My closest listed proxy, Astera Labs, grew 115% to $852M in FY2025 — but over a third of its last quarter came from PCIe Gen6, its fabric-switch line is on track to be its largest, and management calls my controller line "a small part of your business today," with analysts flagging it immaterial for 2026. The ~$16B "CXL market" by 2028? Roughly $12.5B is DRAM behind me at commodity economics; controller and switch silicon clusters at $1.3-3.7B in third-party estimates. The label points at me; the franchise is the connectivity underneath.

Yes. As our capacitor lead mentioned earlier, the Components segment—capacitors, inductors and EMI filters—grew about 12.3% and carried the business at a 22.4% return on invested capital. On the other hand, the Devices and Modules segment declined 5.9% with a negative -3.5% ROIC, reflecting weakness in high‑frequency and communications modules. That said, to be honest, we still cannot fully read how the overall AI‑driven demand will translate to the whole company.

Inward toward the silicon: our filing states all IC manufacturing goes to one foundry across four nodes — 12-nanometer for 100-gig copper, 7-nanometer for optical DSPs, 5-nanometer in flight, 3-nanometer for 200-gig and the bulk of 1.6-terabit. Assembly and test run through several partners; the filing flags capacity, price, and yield risk. Industry supply tightness is significant. We manage it by owning the full BOM, sourcing directly, committing capacity early — clusters don't get built without our connectivity chips. Confidence rests on multi-year supplier relationships, not guarantees.

Hey, Q1 revenue $257.7M, up 10.4% reported, 3% organic — and the scarcity shows up in bookings: every segment double-digit, book-to-bill 1.10, right? Op cash flow $52M, >200% conversion, net cash $139M after the equity-unit recap. Raised guide to $1.04B-$1.055B, EBITDA $245M-$250M reaffirmed — discipline: narrow range, deliver another strong quarter before raising further. Collective breadth, not one line.

Microsoft's CEO argues the durable moat isn't the frontier model — commoditizing input — but the firm-owned 'learning loop': private evals on outcomes, models on internal traces, queryable memory. Test: swap the model without losing 'company veteran' expertise. This describes Microsoft's exact position: platform layer, not model race. The claim carries a pre-registered verifier: learning loop becomes moat only when switching cost measures — retention through swap, durable revenue line, usage depth surviving migration. Vendor essays don't count. Filings will.

Look, SanDisk's fiscal third‑quarter results, for the quarter ending April 3 2026, show total revenue of $5,950 million, up 251% year over year – and the way it got there matters. You know, roughly 248% of that growth came from higher average selling prices while exabytes shipped were about flat. Quite frankly, we swung to net income of $3,615 million from a $1,933 million loss a year earlier, with operating income of $4,111 million and free cash flow of $2,993 million versus $980 million. We'll see how the pricing‑driven turn holds in this cyclical market.

Yes. As our capacitor lead mentioned earlier, the market rumor of a 15-35% MLCC price hike was rejected at the source. We considered various possibilities; at this point we are not raising prices in the near term, placing importance on medium- to long-term relationships over short-term profit. The profit bridge shows prices still falling: a 105 billion yen hit last year, guided to narrow to about 73 billion yen, with the gain from mix improvement, not higher prices. That said, on the top data-center bin, procurement is prioritized over price and we have received no requests to lower prices.

January 2026: Washington awarded $2.7 billion in three $900 million task orders to my incumbent producer and two challengers. One challenger leased a former Kentucky government site for a ~$1.5B project and secured export-bank letters up to $2.4B for Japan and $1.8B for South Korea over ten years. The honest counterweight: every new entrant remains pre-license and pre-operational, targeting production around 2030 or later. My incumbent's 3-5 year head start holds. The money mainly dates 'sole Western producer' to roughly the end of the decade.

I watched telecom's ugliest chapter: vendor financing — Lucent lent ~$8.1B, Nortel ~$3.1B, Cisco ~$2.4B, loans booked as revenue, manufacturing demand until fraud and defaults surfaced together. The AI rhyme: NVIDIA's CoreWeave stake, ~$6.3B backstop, xAI investment, and a reported OpenAI ceiling resized from 'up to $100B' to ~$30B non-binding. But the kill-mechanism differs so far: NVIDIA carries these as investments, not revenue. The test is disclosure quality — legitimate finance, circular optics, and fraud are three different things, and the filings decide which, every quarter.

I hear Samsung Electro‑Mechanics' own materials say an AI server uses more than 10 times the capacitors of a general‑purpose server. I sit chip‑adjacent, a passive stabilizer that bridges the microsecond gap between a distant regulator and GPU current spikes, lest the chip droops, downclocks or crashes. The AI‑grade parts must be high‑capacitance, tiny, low‑inductance, voltage‑tolerant and highly reliable, built from two‑micron ceramic layers stacked up to a thousand deep and sintered without cracking – the manufacturing is the moat, the physics unforgiving.

As optical chips migrate into costlier packages — co-packaged with switch silicon or integrated into compute — the cost of packaging a bad die rises sharply. Wafer-level reliability screening catches weak dies before packaging: one test step versus the full module cost. The spread widens with every step up in package value, the structural case for this layer. Independent research frames it; industry adoption at scale is the open bet. Metrology and finished-module test don't substitute for stress-screening the bare die.

Cameco's uranium comes from Cigar Lake (about 50% owned), the McArthur River joint venture with Orano (about 70/30), and the Inkai joint venture with KazAtomProm in Kazakhstan (40/60). This is the spine of our supply chain. McArthur River is licensed to 25 Mlb, has run at 18 and 20 Mlb, and we guide 2026 output somewhat below design as we pace to demand. Port Hope, Canada’s sole UF6 converter, set a production record in 2025. Our 49% stake in Global Laser Enrichment has reached readiness level 6, is not yet operational, and commercial deployment is framed as post‑2030.

A reading of the quarterly filing shows a net loss of about $(113.5) million. A separate read of the same quarter shows positive net income near $621 million, flattered by a roughly $780.6 million one-time, non-cash gain from revaluing a ClickHouse stake — stripping that out leaves an adjusted loss near $(100.3) million. What the reported net-income line includes or excludes is a to-verify item for the next filing. Either way the shape is the same: core cloud adjusted-EBITDA-positive inside a group that still loses money once GPU-fleet depreciation is counted. Still the very early days.

For the full year it was $73.0 million, up about 202% from $24.2 million, driven almost entirely by SiCore cell sales; last quarter $28.5 million, up 153% year over year and 13% sequentially. Gross margin only recently turned positive — roughly 11% for the year, about 20% last quarter — and we target 25% for the full year, still in sight but not yet reached. One customer was roughly 37% of last year's revenue, unnamed. For those of us keeping score, real traction on a base small enough that a single customer and a single margin quarter move the whole picture. That's okay.

We report HVAC&R and refrigeration components represent about 60% of revenue, covering valves—electronic‑expansion, four‑way reversing, stop, solenoid and ball—microchannel heat exchangers, pumps, controllers and sensors for residential and commercial air‑conditioning, refrigeration, cold‑chain and heat‑pump markets.

We make high-energy-density lithium-ion cells on a silicon-anode design — our commercial platform is SiCore, and the older SiMaxx line is winding down. Think of a standard graphite battery as drip coffee and ours as espresso: same energy, much smaller package. That density roughly doubles a drone's flight time or a light EV's range, and it's the whole product — the reason we're designed into long-endurance aircraft. For those of us keeping score, it's a genuine design edge in a fiercely competitive market, not a material monopoly, and we have to re-win it every design cycle. That's okay.

About 53% of first‑quarter 2026 sales were into medical end markets and about 47% into advanced industrial — and the medical half is the larger growth engine, not a drag, right? Automation Enabling Technologies was $500.8 million, or 51% of full‑year 2025 revenue, and Medical Solutions was $479.8 million, or 49% of total full‑year 2025 revenue of $980.6 million, up 3.3%. Medical consumables are roughly 15% of sales and growing double‑digit. We see roughly a $4 billion incremental market opportunity by 2030 across four platforms, with AI‑datacenter exposure a minority of the whole.

Independent research scores planetary roller screws 14/18, naming them the most acute physical bottleneck alongside harmonic gears — narrow supplier base, long lead times, limited substitution. A humanoid uses roughly 14; the roller-screw linear joint sits at 20-30% of high-spec component cost. Capacity flexibility scores 1 because precision grinding gates supply. That cuts both ways for us: we benefit if grinding stays scarce, and as one scaling grinding capacity, we also erode that pricing power if capacity outruns demand. Forward analyst framing, not a filed result.

We delivered $529.8M revenue for 2025, up ~480%. Core cloud turned adjusted-EBITDA-positive at ~$59M, from a $(128.5)M loss. Group adjusted EBITDA stayed negative at ~$(64.9)M, dragged by Avride (~$(82.7)M) and TripleTen (~$(41.2)M). Adjusted EBITDA excludes GPU depreciation, so core remains loss-making on a bottom-line basis; Q1 2026 group net loss was $(113.5)M. Still the very early days.

Broadcom's Hock Tan dismisses optics as “bright, shiny objects” while championing copper and saying “we are the lead in CPOs.” Marvell's Matt Murphy says scale‑out co‑packaging is “relatively limited” but scale‑up will shift “in a pretty big way,” backed by an acquisition of up to $5.5 billion. Both executives frame their postures to match capital incentives, leaving the real outcome to customer deployments.

Look, filing shows one distributor at 11% of FY25 revenue — call it up from 10% in FY24 and under 10% in FY23 — and it sits across all three segments. We don't name the distributor. AR side, same distributor at about 10% of balance, down from 13%. Concentration's modest but ticked two years straight, so we flag it. Independent work separately has us roughly number two in power discretes and modules at 8.5% share behind Infineon — that's an outside estimate, not our number.

Look, you know, the renamed segments tell the story: Datacenter at $1,467 million, roughly a quarter of revenue, up 645% on the new framework — though the release says 233%. I mean, Edge hit $3,663 million, up 295%, Consumer $820 million, up 44

Laser PO's are supply-chain disclosure — useful, but my scorecard waits on three pre-registered signals: design-win concentration on partner rosters; silicon-photonics startups acquired by hyperscalers not merchants (breaking the open-platform thesis); or hyperscalers staying on pluggables into 2028, meaning the battle was fought over a market that didn't arrive on schedule. Until one fires, every scorecard — including the supply-chain read — stays provisional.

For the year ended March 2026, net sales ¥355.3bn (+4.1%), operating profit ¥20.0bn (+91% YoY). The jump came off a 3.1% margin trough — almost entirely utilization recovery, not re-rate. Bridge: ¥30.7bn utilization gain offset by ¥20.2bn price decline; prices still fall, just slower. Operating margin 5.6% vs leader's ~15%. Net profit flattered by ¥4.8bn FX gain, dented by ¥2.1bn aluminum-electrolytic impairment and ¥1.5bn communication-device reform loss — the operating line is the one to read.

The Prohibiting Russian Uranium Imports Act became law in May 2024, with a full import ban taking effect January 1, 2028. Russia supplied about 24% of US enrichment services, so Western capacity must expand against a hard deadline. Utilities are already restructuring procurement, and spot enrichment prices roughly doubled between 2024 and 2025. Russia could still retaliate with its own export ban before 2028, tightening the squeeze further.

I would say AI and Cloud grew 49% constant-currency, 94% reported to EUR 350 million, booked roughly EUR 1 billion new orders this quarter, EUR 2.4 billion for the year. But it's 7.8% of revenue, a little bit over seven percent. Telecom providers still EUR 3.3 billion, 72.6%, down two percent. Mission Critical up 19% to EUR 498 million. The acceleration is genuine, I think, and it's diluted at group level by the telecom base. Both numbers in the same disclosure.

On the transformer piece: our solid-state transformer targets first delivery to an unnamed hyperscaler this fall, then six months of customer testing before any order, likely H1 2027. It's one bead in the string of pearls — energy management system orders booked this quarter and April, stability block with medium-voltage UPS could see incremental H2 2026 orders if things go our way. We're candid it doesn't come at once; most of the integrated offering is still ahead of us. Based on how we see things today, it's just a start.

We broke the international-brand monopoly in high-end excavator hydraulics, filling the domestic gap in high-end super-large-excavator pumps — the filing's phrasing: filled the domestic high-end super-large-excavator-pump gap, further breaking the foreign-brand monopoly. Mid-large excavator pumps and valves rose about 40% in FY2025. That track record, not a forecast, is why independent research names us on the roller-screw layer: precision grinding gates both the hydraulics we dominate and the roller screws we are now sampling.

Our HVAC&R and refrigeration components — valves (electronic expansion, four-way reversing, stop, solenoid, ball), microchannel heat exchangers, pumps, controllers, sensors — make up about 60% of revenue across residential and commercial AC, refrigeration, cold chain, and heat pumps. We describe a near-exclusive global supply position in air-conditioning electronic-expansion and solenoid valves, a 30-year lead with multiple products ranking top globally. The auto and EV thermal business is the other ~40%. Both are large, real, and sit mostly outside the AI and humanoid conversation.

Three things on the quarter, okay? Revenue $82B, up 85% year over year, 20% sequential — fourteenth straight quarter of sequential growth, and the $13.5B step-up is a record. Data Center $75B, networking near tripled. Free cash flow $49B. Guidance $91B plus or minus 2%. My sense is the map re-draw matters: Hyperscale $38B up 12%, ACIE $37B up 31% — the diverse half grows faster. SEC still uses the old segments; the new framework lives in the presentation.

I see the durable AI seat is either the un‑floodable physical layer—land, power, grid equipment, where lead times stretch toward five years and roughly half of planned 2026 US datacenter builds are power‑bound—or the vertically integrated center that owns silicon, the cloud and the applications, letting value migrate up the stack without leaving its own income statement. That echoes the telecom bust where clean‑balance‑sheet firms survived while pure‑play compute landlords were exposed.

The most recent earnings call named Microsoft's Fairwater data center (hundreds of thousands of GPUs), AWS (adding over 1 million Blackwell and Rubin GPUs starting this year), Google's A5X instances (supporting up to 960 000 Rubin GPUs), Anthropic as a strategic partner, and eleven frontier labs in one enumeration. There are three things, okay? First, we consistently name customers; second, we never name the foundry that makes everything; third, even the laser makers we invested in stay unnamed on the call. My sense is the asymmetry itself tells you where the story lives.

I entered the standalone CPU market with Vera — management describes it as the world's first CPU purpose-built for agentic AI. Custom Arm cores. Claimed vs x86: 1.5x perf per core, 2x perf per watt, 4x density per rack. Four ship modes: with Rubin (one Vera per two Rubins), standalone, storage stack, confidential stack. The $20B visibility figure covers standalone only. My sense: supply constrained through the life of Vera Rubin. Agents don't rent cores — they want the work done fast. Tokens per dollar, not dollars per core.

Hey - TSMC's partnership model at leading-edge is real, we grant that. Our FY2025 filing: primary fabs in Israel, hostilities early 2026 blocked vendor equipment installs, which may delay our ~$920M silicon-photonics capacity plan. April 2026 ceasefires noted. The $300M Intel equipment commitment from our terminated merger is in mediation - Intel expressed intention not to perform as of year-end 2025. Five analysts didn't probe the geopolitical risk. Both matters unresolved. We plan around real manufacturing-base uncertainties and disclose them plainly. Did that land?

Three things in the concentration disclosure, okay? Direct customers: 21%, 17%, 16% — 54% combined versus 30% a year ago. The filing estimates one unnamed AI research and deployment company buys indirectly through them — widely assumed to be a major lab, not disclosed. Geography: 22% non-US revenue versus 42%, halved by China controls and US hyperscaler weight. My sense is the revenue grows and the list narrows at the same time.

I see Murata’s April 2026 full‑year results put primary numbers under the AI‑capacitor story: server‑capacitor growth guided at roughly 85‑90% (revised up from about 30%), and for the first time a broken‑out data‑center revenue line up 74% last year and guided up 84%. They also announced roughly 80 billion yen of emergency capacity spending for server capacitors. Taiyo Yuden’s Q&A noted customers multi‑source, but from a limited set of qualified suppliers – an oligopoly held together by qualification, not contracts.

MP Materials' first quarter of 2026: consolidated revenue plus price‑protection income of $132.9 million, a record, with total revenue of $90.6 million, up 49% year over year. You know, the net loss narrowed to $8.0 million from $22.6 million a year ago, and Mountain Pass set production records – 917 t of NdPr oxide produced, 1,006 t sold, 12,983 t of rare‑earth oxide overall. We’re still improving off a loss‑making, capital‑heavy build‑out, not yet steady profit.

The steelman asserts a dollar of compute infrastructure converts into software revenue the market values far more highly, and the conversion is claimed to be predictable, with capability and revenue jumps following new compute by roughly 60‑90 days. I flag the built‑in falsifier: the arbitrage only works while that conversion holds, an open industry question. The course lists tripwires—order books below shipments, GPU spot‑price collapse, sustained hyperscaler capex cuts—and notes none have fired as of mid‑2026.

Two things on China, okay? Licenses for H200 shipments are approved, but we have yet to generate any revenue and we are uncertain whether imports will be allowed — guidance assumes zero data center compute revenue for a second quarter. The H20 charge narrowed to $1.1 billion from $5.3 billion. My sense is Chinese competitors, bolstered by recent IPOs, are making progress and could disrupt the global AI industry structure long-term. We plan around a zero from what was once a major market.

Revenue accelerated as the construction-machinery cycle turned up: 9-month 2025 revenue rose about 12.3%, then Q1 2026 revenue rose about 32.5% on downstream demand recovery and rising sentiment. But margins compress: Q1 earnings rose only about 6.5% on that revenue jump; full-year net profit up 9.0% lagged revenue up 16.5%; operating cash flow fell about 27% as the ramp absorbed working capital. Earnings track the excavator cycle — currently a tailwind, but the swing factor. Honest read: genuine cyclical volume recovery with a margin-and-mix headwind to watch.

Thank you for the question. Let me answer what I think you're asking. In January 2026 the Department of Energy selected us for a $900 million HALEU Enrichment Award — management says it could exceed $1 billion with options, but it still needs to be finalized through negotiations and pays against milestones. I frame it as another pool of low-cost capital through procurement, neither debt nor equity, supporting the 12-metric-ton buildout. The NNSA sole-source notice is in a procurement cycle; I'll let the government drive those announcements. Day one: reduce lead time, reduce unit cost, de-risk.

AeroVironment's filing states its motors and batteries rely on rare-earth metals "of which a significant majority are sourced from China," notes a March 2025 Chinese export-control listing, and separately discloses secure-microelectronics dependence — the visible prime consumes both bottlenecks it doesn't own. Kratos files the inverse: no rare-earth dependence disclosed, vertically integrating jet engines and, via JV, solid rocket motors. On this map, disclosed and inferred are different grades; the weakest link-label sets the honesty standard.

Net loss was $44.0 million for the full year and $5.0 million last quarter. I’m loss‑making and cash‑burning, funding the ramp by issuing stock, not debt. Operating cash burn hit $37.3 million, pulling cash from $90.5 million to $62.4 million. We walked away from a planned gigawatt plant, taking a $22.5 million write‑down and $20 million lease settlement, and pivoted to an asset‑light model with contract makers in Korea and the U.S. For those of us keeping score, cash should cover obligations for the next twelve months, though we may be unable to raise more on favorable terms. That’s okay.

Sub-micron laser placement, lens alignment, fiber coupling, wavelength-multiplexing integration, hermetic sealing, thermal management, and full signal-quality calibration — that's a 1.6T module. Independent research calls this the "boring but decisive" layer: precision and yield demands rise non-linearly with bandwidth, and at hyperscale, yield gains compound. If leading-edge module demand keeps exploding, the assemblers who deliver at scale may capture more value than the market expects from an assembly business.

Yeah, you're flagging the concentration — one distributor channel hit 45% of revenue in the quarter ended May 2, up from 36% a year ago, distributors aggregate 51%. One direct customer held 16%, flat. Four customers were 73% of gross AR at year end. We don't name them. Two warrants — 4.2M and 1.0M shares — vest on revenue milestones, both reduce revenue, signaling at least two major custom programs with equity terms. That's the exposure. Look at the filings. Blinking?

My pure-play's numbers show the layer inflecting: bookings surged 6x QoQ to $37.2M, backlog hit a record $50.9M, and advanced-packaging test drove my semiconductor segment from 3% to 25% of revenue. The CEO names a leading advanced-packaging platform and cites discussions with key memory suppliers targeting fiscal 2027 orders. The counterweights are equally on record: high customer concentration, forward-looking opportunities not yet shipping, and filings silent on some markets the calls emphasize. One company's order book is evidence of a layer waking up — not an industry standard.

In April 2025 the US Department of Energy selected 5 reactor developers from 15 that requested my fuel allocations. Developers receive my scarce output through government allocation rather than contracting directly with the producer. A government agency deciding which customers of a fuel commodity get supplied reads as a strong signal of how binding the shortage is: at current production, an ordinary commercial market would not have enough to go around.

The course's central claim is about what to measure. It says the market over‑indexes on GPU shipments, capex budgets and announced megawatts, while the discriminating factory metrics are value per gigawatt, tokens per watt and tokens per dollar. Google's infrastructure chief summed it up: 'the measure isn’t how much money you spent per gigawatt, it’s how much value you deliver per dollar,' and notes that only the energized and utilized stages of the megawatt funnel are real capacity; everything upstream of 'energized' is a press release.

Analysts flagged weakness in ERCOT and PJM forward prices even as the demand pipeline looked strong; we argue the forwards undervalue the 2028‑2029‑and‑beyond period, note the ERCOT load “isn’t yet on the system – it’s getting built,” and say we’ve stayed well hedged and protected against near‑term weakness. Look, we’re a competitive merchant generator operating in PJM, MISO, NYISO and ERCOT, explicitly not a regulated rate‑base utility, so the Pennsylvania governor’s letter applies to regulated utilities, not to us.

Hey - silicon photonics is our fastest-growing piece, up about three times year-over-year in Q1 with first revenue shipments from Fab 2 in Israel and Fab 7 in Japan, the latter at 95% yield on first photonics wafers. We've got $1.3 billion of 2027 contracted revenue backed by $290 million in prepayments - that's the contract, not the forecast, which runs higher. Against a $230 million 2025 baseline it's roughly a six-fold step. Two platforms in the 400G-to-1.6T space, over 50 active customers. Next checkpoint: January 2027 prepayments on the 2028 commitments. Did that land?

China introduced export controls on certain lithium-ion batteries, materials, and the equipment to make them — enforcement suspended until at least November 2026. For those of us keeping score: our silicon anode still comes from a Nanjing affiliate under an exclusive deal without commercial terms, and we rely on contract manufacturers in China and Korea. A DIU contract now at $18.1 million requires qualifying non-China sources. We're racing to become the domestic battery we're marketed as. That's okay.

I'm mapping a hard ceiling on commodity labeling: Google fired its top annotation vendor and internalized raters; Meta's $14.3B Scale AI stake made Google, OpenAI, and xAI pull contracts. The margin squeeze stops at the toll-booth owners — expert evaluation stays untouched.

July twenty-five, you know, we made the call — zero China sales, the Shenghe offtake, related party, all of it — to align with the Department of War agreements and the domestic chain build. Full-year twenty-five revenue two twenty-four four versus two fifty-three four in twenty-three, cutoff timing, but Q1 twenty-six up forty-nine percent year over year as the ramp and allied demand took hold. Deliberate choice, giving up a market to rebuild the Western supply chain, candid on the reshaped base. Big grain of salt on the timing, you know — these projects are hard.

Hey - as I mentioned, full-year 2025 was about $1.57 billion, up 9%, gross margin near 23%, operating margin around 12%. Q1 2026: $414 million, up 15%, gross profit up 52%, operating profit up 96%; gross margin 27% from 20%, net margin 16% from 11%. Guiding Q2 to ~$455 million ±5% - highest quarter ever, up 22% - but that's guidance, not reported. Model has incremental revenue at 59% to gross profit over 20% baseline, toward 39% at full build-out. Did that land?

At our core we are a large, founder‑controlled Tier‑1 automotive‑parts supplier organized around eight product lines – NVH and shock absorbers, interiors, chassis, auto‑electronics, thermal management, line‑controlled brake, air suspension and steer‑by‑wire/intelligent‑driving systems – supplying roughly ¥30,000 of content per vehicle as a one‑stop, system‑level, modular supply. That core business represents about 93% of revenue, and auto‑electronics was the fastest‑growing line in FY2025, up about 52%.

Hey, good question. As I mentioned, we're deliberately not a leading-edge foundry — we compete on specialty process, not the smallest node. Our platforms span analog, RF SOI, silicon-germanium BiCMOS, mixed-signal, power-management BCD, CMOS image sensors, and silicon photonics, all at mature nodes on 200mm and selected 300mm wafers. That's a different margin structure than what you're describing. Did that land?

Look, the data-economy signal hasn't slowed — hyperscalers still revising 2026 spend up nearly 75% year over year. ERCOT queue shows 400,000-plus megawatts; we don't expect anything near that to materialize. Forwards past 2029 price maybe 10,000 to 15,000 incremental; if realized load lands closer to 30,000, that's upward pressure. We've put roughly 5,000 megawatts — uprates, gas, storage — into PJM's queue, and the CyrusOne substation at Freestone targets Q4 2026 energization. Timing's the open variable; much of this load isn't on the system yet. Tunnel's real, light's visible.

The platform layer everyone photographs scores lowest on structural attractiveness: small, one‑way‑attack drones are being deliberately commoditized toward a ~$2,000 price floor by their own government buyer. Upstream inputs—magnets, secure chips, seekers, batteries, edge‑AI compute—score highest, being scarce, concentrated and capacity‑constrained. Autonomy software scores ‘very high’ with an asterisk, the best capability remaining private, while counter‑drone systems stay persistent and fragmented. I map these layers because the inversion law says value runs opposite to visibility.

The circularity you're circling: NVIDIA holds ~11% of CoreWeave, supplies its chips, and reportedly backstops ~$6.3B unsold capacity — the supplier funding its own customer's demand. Contracts then split by counterparty: Meta paper (near-sovereign) durable; cash-burning AI lab paper fragile. Your single-customer flag hits the fragile node. Durability is a decision made in someone else's boardroom.

The cleanest arithmetic in the whole debate sits in CoreWeave's own Q1 2026 results: adj EBITDA ~$1.16B, D&A ~$1.15B — near-exact offset — leaving ~$740M net loss after ~$536M net interest. The ~56% EBITDA margin evaporates once chip depreciation and debt service are counted. Adversarial verification killed three pricing claims including the 95% re-rent anecdote (refuted zero-for-three). Spot prices rebounded; contract rates stay contested. A spread that lives on shortage is revenue. It is not a moat.

We see pressure to commoditize reducers, so we are moving up the value chain, shifting from selling reducers alone to higher‑value‑added actuators with built‑in drivers that can serve as a platform for all axes, so customers need not design reducers and motors from scratch. We target substantial output in about 1.5 years, and we are also developing an ultra‑compact gearbox dexterous hand with Minebea Mitsumi, shown at CES 2026. Whether this added value holds as challengers narrow the precision gap remains an open question we flag.

I'll own the margin. Q1 adjusted gross margin 45.6%, down ~60 bps year-over-year on price-cost timing — higher freight, tariff, material costs from geopolitical shifts that moved faster than we could surcharge and reprice, right? Tariffs changed in a four-week window; aluminum 25% to 50% on immediate shipments. Timing, not trend: new surcharges and price increases active, backlog repricing, two sites closed. Expect margins back on prior full-year path in H2, price-cost positive by Q3, though full return is an open question. No tariff refund benefit included — our risk buffer.

Yeah. I want to highlight that wide‑bandgap revenue reached $15.1 million in Q1 2026, up 152% YoY, with silicon‑carbide wafer shipments up 195% to 14,300 units – but that is linked to a depressed base, FY2025 silicon‑carbide revenue was $33.8 million, down 34% after an exceptionally strong prior‑year Q1, so the high growth rate is a low‑base rebound as much as a new trend. Does that answer the question?

Look, there is a divergence in AXT's own disclosures about how many InP substrate suppliers exist, and it is worth holding both halves. Our filing lists three primary suppliers worldwide, naming Sumitomo Electric and JX Nippon, while on the call the CFO said there are two and we are one. The filing notes a third that the call omits, leaving open whether that third is commercially marginal or the call's framing narrows the field to assert more dominance. Both statements stand as management’s; neither has been reconciled. Okay?

Analysts flagged weakness in ERCOT and PJM forward prices even as the demand pipeline looked strong; management argued the forwards undervalue the 2028‑2029‑and‑beyond period, said the ERCOT load “isn't yet on the system – it's getting built,” and noted it has stayed well hedged and protected against the near‑term weakness. Look, that illustrates the merchant model – price exposure today, but the chance to capture premium for clean, firm power under long‑term contracts.

Valkyrie was pre‑low‑rate‑production, the first low‑rate contract is in negotiation and we aim for about 40 drones a year by early 2028 – units were partly built as company‑owned assets ahead of any award, so revenue recognition depends on the customer’s final configuration. Think affordability. Okay? All right?

Oracle's Project Jupiter is an up‑to‑2.45‑gigawatt power block for an AI factory in New Mexico, which we call 100% Bloom—an islanded microgrid that replaces the previously planned gas turbines and diesel backup generators. No grid, no dirty diesel, no battery banks—just Bloom alone. When finished it will be among the world’s largest islanded microgrids, run under a master services agreement that lets a hyperscaler shift Bloom deployments, and we have issued Oracle a warrant to purchase Bloom common stock.

The brain‑layer leaders are demos, not products at scale: the famous humanoid dishwasher demo was independently confirmed curated — all objects plastic — and the oft‑cited 94% out‑of‑distribution success claim failed verification zero‑for‑three. I note the near‑term market is small, with industrial‑humanoid revenue pegged around $210‑$270 M for 2026, roughly 15 000 units, far shy of $5‑$40 trillion headlines. NVIDIA’s robot exposure is roughly 1‑2% of its datacenter business, so the optionality is already baked into any AI build‑out exposure.

First, our fiscal year ended April 30, 2026 was a record – revenue about $1.98 billion, up 30% organically and 141% as reported, lifted by the BlueHalo acquisition. Second, Q4 set a record at $642 million, up 31% organically, and adjusted EBITDA was $286 million, a 14% margin, above the high end of the company's most recent guidance range, with Q4 contributing $140 million at a 22% margin. Third, total backlog was $2.7 billion, funded backlog $1,183 million, up 63%, with book‑to‑bill 0.9× in Q4 on $572 million bookings and 1.4× over the trailing twelve months.

Edge silicon runs roughly 30‑50× below datacenter memory bandwidth, and device‑RAM ceilings cap what models fit, so frontier inference stays in metro datacenters. Verified framing: latency, not cost, sets the split — real‑time needs low‑round‑trip placement, meaning a datacenter closer to users, not a chip in your pocket. The 'train in cloud, infer at edge' thesis failed three tests. Industry reporting calls the AI‑PC/phone supercycle dead on arrival; the edge‑silicon leader's own filings show core handset chips with datacenter ambitions early and unproven. Edge AI is real. Edge relief is not.

CoreWeave is bigger – $2.08 B quarterly revenue (+112%), a ~$99 B backlog – and more fragile: $21‑25 B of debt at 9‑15% effective rates, GPU‑collateralized SPVs tied to single contracts, and a $4.2 B 2026 maturity wall. My six‑test scorecard flags Nebius’s supposed diversification as a falsifier, since >90% of its backlog appears tied to Microsoft and Meta (inferred). Both firms sit on the bottom rung of the durability ladder, renting the same chips and power and owning only leverage and spread.

I mean, we raised 2026 capex guidance to $125-145 billion from $115-135 billion. Management attributes most of the increase to higher component costs — particularly memory pricing — not volume. The honest read: even with enormous AI demand we pay the memory bill. Midpoint runs ~87% above 2025's $72 billion. Q1 came in at $19.8 billion on servers, data centers, network. Expenses held at $162-169 billion, operating income expected above 2025.

Your bandwidth leap means wilder current swings — Samsung Electro-Mechanics' materials say an AI server already uses >10× the capacitors of a general-purpose one to bridge the microsecond gap between the distant regulator and those spikes. Without it: droop, downclocks, crashes. The spec is brutal: two-micron layers, up to a thousand deep, sintered without cracking. Manufacturing is the moat; physics is unforgiving.

From the best seats, the course confirms the buyer: power — not capital or chips — gates gigawatts, and memory bandwidth, not raw compute, is the wall. It refines with 20-year take-or-pay contracts and optical circuit-switching. It oversells where incentives predict: 'uncapped demand' is a token-seller's claim, the 2027-28 memory-glut counter-case goes unspoken, and the 'advanced packaging' layer has zero packaging content across twelve lectures. Clearest signal: insiders with opposite books describing the same market differently.

Quarter ended March: $751.1M revenue, up 130.4% — first quarter above 100% growth as a public company. Product $653.3M, service $61.9M. Gross margin 31.5%, operating margin 17.3%, adjusted EBITDA $143M. Operating cash flow positive $73.6M, first time in a seasonally weak Q1. Cash $2.52B. Guidance raised to $3.4-$3.8B, ~80% growth at midpoint, non-GAAP gross margin ~34% — barring any global shock or exogenous factors. The 43% related-party concentration sits in the same breath; Oracle's Jupiter and the hyperscaler base are how that widens.

We ship A17 (17 active DOF, finger-joint drive units), B06 (6 DOF, linkage transmission, heavy-duty), and B20 (20 DOF, 600g, ≥30N grip, 5-12N fingertip, CES Jan 2026) — real hardware, not renders. The ZWHAND platform and a 50M yuan subsidiary back them. But revenue stays micro-transmission (~60%), precision parts (~35%), molds (~5%); humanoid sales are embedded, not broken out. Mass production still ramping. A product-backed option on the dexterous-hand layer, not yet a sized business.

My net assets to shareholders stand at about 3,483 million yuan, up 7.96%, against total assets of about 4,318 million yuan — equity ratio near 81%, net cash, no long-term debt, clean unqualified audit. Weighted ROE 7.14%, depressed not by unprofitability but over-capitalization — about 3.48 billion equity and 4.32 billion assets against only about 1.72 billion revenue, with much of the 2021 IPO's roughly 2.0 billion yuan in low-return wealth-management. The flip side: a self-funded war chest to bankroll the robotics build-out without raising external capital, an option cheap to hold.

Yeah, I want to highlight what sits beneath those numbers: we are a pure-play analog and mixed-signal specialty foundry, 1.0 micron down to 110nm on 150mm and 200mm wafers across six fabs in four countries — Erfurt, Dresden, Itzehoe, Corbeil-Essonnes, Kuching, Lubbock. Automotive, industrial, medical is 94% of revenue. The moat is requalification friction, not contracts; we are sole-source on most products but customers have switched with little notice. Utilization in the low 60s percent shows that friction has not translated into pricing power. Does that answer the question?

Yeah, I want to highlight wide-bandgap: $15.1M in Q1 2026, up 152% YoY with SiC wafers up 195% to 14,300 units — but off a depressed base, FY2025 SiC revenue $33.8M down 34% after an exceptionally strong prior-year Q1, so the growth is a low-base rebound as much as a new trend. Data-center link is concrete on technology: SiC for solid-state breakers and 800V-to-48V conversion, Navitas named, our timing chip synchronizes GPUs though that customer undisclosed. Capacity 10k wafers/month, ~6k loaded (~60%). Whether data-center share climbs remains open. Does that answer the question?

I mean, Q1 revenue hit $56.3 billion, up 33% year-over-year, with operating income of $22.9 billion at a 41% margin — up 30%. That's the clean read. Net income was $26.8 billion but carries a roughly $5 billion one-time tax benefit; management gave the normalized figure: about $18.7 billion without it. Family of Apps drove $55.9 billion revenue and $26.9 billion operating income at a 48% margin. Full-year 2025: $201 billion revenue, $83.3 billion operating income, $72.2 billion capex. Read operating income, not the tax-noisy net line.

Yeah, I want to highlight the concentration: Melexis was 43% of FY2025 revenue, about $377M, top three 53%, top five 58%. Into Q1 2026 that share fell 41% to 35%, framed as diversification with customers two through twenty up 21% QoQ. The fuller reading: decline linked in large part to involuntary automotive destocking at that customer — auto revenue down 10% YoY — not purely deliberate diversification. Both readings hold, and the annual report risk factors disclose the concentration plainly. Does that answer the question?

I mean, capex guidance moves to $125-145 billion for 2026, up from $115-135 billion. The honest read: most of the increase is memory pricing, not more GPUs. Even a company with this demand curve pays the component bill — that's just true. Q1 came in at $19.8 billion on servers, data centers, network. Midpoint runs about 87% over 2025's $72 billion. Expenses held at $162-169 billion, operating income expected above last year. I have a sense of the shape of it.

Five gigawatts a year — that is the footprint we built. Not one lumpy addition a year, but hundreds of megawatts each quarter, copy-exact, so we are never the bottleneck. Management's claim: neither order-constrained nor capacity-constrained. The pace belongs to the customer's greenfield build, not our supply line. The CEO frames the old industry's multi-year backlog as constrained supply, not strength. Beyond five gigawatts takes new factories; we will build them as the market needs. The step function follows demand.

Industry research claimed Monolithic Power lost 60-70% of its flagship AI power socket to European and Japanese rivals. Primary check: record $804M quarter (+26%), ~55% gross margins for fourth straight quarter, Enterprise Data flat in dollars (share drop was composition from five segments growing 26-46%), CEO denied socket loss and raised segment growth floor to 85%. A CEO denial is incentivized, not proof — but a rumor contradicting revenue, margins, and guidance simultaneously carries the burden of proof. In a hot sub-sector, socket-rumor churn is constant; filings are the referee.

I mean, we're rolling out more than a gigawatt of custom silicon with Broadcom, plus a significant amount of AMD and NVIDIA systems — custom and merchant growing together. Broadcom confirms it on their end too. We're expanding our own data centers and have multi-year cloud deals coming online across 2026 and 2027 to scale faster. The honest counterweight: Reality Labs lost $19.2 billion on $2.2 billion of revenue last year, and the apps business — $102.5 billion operating income — is what pays for it.

Yeah, so we see it differently — silicon dollar content climbing past $1,000 per accelerator as racks shift from retimers to full AI fabric. PCIe Gen 6, 800G and 1.6T Ethernet driving that. Our Scorpio P-Series is the only PCIe 6 fabric shipping in volume, millions of ports out, PCIe 6 already more than a third of Q1 revenue. X-Series 320-lane with Hypercast shipping initial volumes, ramps H2 2026; Scorpio on track to become our largest line by year end, X-Series eventually exceeding P-Series. We intercept the fabric ramp; what the economics look like at scale remains to be seen.

For 2025, consolidated revenue was about CAD $3.5 billion, up 11% year over year. Uranium EBITDA rose 7%, Fuel Services EBITDA rose 22% — our fastest-growing segment — and our share of Westinghouse net earnings swung up roughly CAD $276 million. Production hit 21 million pounds, exceeding guidance. Port Hope set a record. This is what pacing looks like across the full fuel cycle. We don't front-run demand with supply. The figures are our own year-end results.

Our uranium comes from Cigar Lake (about 50% owned), the McArthur River joint venture with Orano (about 70/30), and the Inkai joint venture with KazAtomProm in Kazakhstan (40/60). McArthur River is licensed to 25 million pounds; 2026 output guided somewhat below design as we pace development to demand. Port Hope, Canada's sole UF6 converter, set a production record in 2025. Global Laser Enrichment has reached readiness level 6, commercial deployment framed as post-2030. This is the supply chain we operate. This is the pacing we choose. This is not a spot-price response.

First, our filing still shows China's export-control list from March 2025 — no escalation, no relief. Second, motors, batteries and advanced components lean on rare earths, a significant majority from China, plus semiconductor sourcing risk there. Third, our edge is production scale, breadth and programs of record, not upstream ownership. Titan, our AI-enabled RF counter-drone system, more than doubled pro forma sales in fiscal 2026 with demand rising through 2027 and beyond. US government customers roughly 85% of revenue, DoD roughly 63%.

AI racks pushing toward 600 kW hit copper-thickness and loss limits at lower voltages, so NVIDIA is driving an 800 V DC architecture for its 2027 rack generation with conversion moving closer to the chips. The per-rack power-chip content expands — but the research insists the dated transition isn't a business result. Five companies claim the tailwind with five different structural positions; value capture is demonstrably asymmetric.

Cameco owns 49% of Westinghouse; Brookfield 51%, closed November 2023 on an equity-method basis. The US$80 billion partnership term sheet is signed — definitive agreement still under negotiation — and the government's participation interest takes 20% of distributions above US$17.5 billion. 2026 guidance for our share of Westinghouse adjusted EBITDA is US$370-430 million, strong but lower than 2025; management attributes the step-down partly to the one-time Dukovany distribution and project-timing lumpiness. New-build earnings are lumpy, not a straight line.

The only company quantifying humanoid revenue is a US sensor maker at ~$600K/quarter, 1.5% of revenue. Twelve filings read: the two actuator "headline picks" are blue-chips but their humanoid arms are pre-revenue research — bank "dominance" calls are forward bets, not disclosed reality. Harmonic Drive Japan holds the strongest validated gear position on a fortress balance sheet, yet management says the AI-robot order ramp has slowed. The credible roller-screw entrant mentions humanoids zero times. Positions real, businesses real, humanoid revenue a rounding error almost everywhere.

Ibiden, the lead maker of package substrates under AI processor dies, filed FY2026 numbers showing the tier's durability. Its disclosed leading-edge AI-server substrate share runs roughly 70-80% — near-100% at each new generation, then decaying 20-30% over 3-6 months as rivals qualify. Electronics operating margin rose from 13.6% to 18.6%, guided toward ~22.7%. Clearest signal: customers pre-fund new capacity in advance. Caveat: the filing covers only logic substrates; a claimed memory-stack role isn't confirmed.

My makers' own materials say an AI server needs more than ten times the capacitors of a general-purpose one. I sit chip-adjacent, bridging the microsecond gap between the distant voltage regulator and the GPU's current spikes — without me the chip droops, downclocks, or crashes. The spec is brutal: high capacitance, tiny, low inductance, voltage-tolerant, and reliable all at once, from two-micron layers stacked up to a thousand deep, sintered without cracking. Manufacturing is the moat; physics is unforgiving.

First, our fiscal 2027 guide comes in at $2.125 to $2.225 billion — roughly 10% growth at the midpoint — with adjusted EBITDA $305 to $325 million. Second, that number is deliberate: we assume a continuing resolution, budget approved December or January, cash to the services around March. Third, we run on $2.7 billion backlog and funded awards until then. CapEx 12 to 14% of revenue for capacity — Salt Lake City, Huntsville, Albuquerque, Dayton — so fiscal 2027 won't be free-cash-flow positive. SCAR excluded. Reconciliation bill timing uncertain, not assumed.

I mean, Q1 revenue came in at $56.3 billion, up 33% year-over-year, with operating income of $22.9 billion at a 41% margin. That's the clean number — net income carried a roughly $5 billion one-time tax benefit. Full-year 2025 capex was $72.2 billion. The bill comes due either way.

App-layer monetization either clears the ~$700B+ annual hyperscaler build-out or exposes it as premature — the clock is tight because AI chips depreciate in 2-6 years versus decades for fiber. My decisive signal (frontier lab gross-margin trajectory) is unverifiable; labs are private, revenue claims conflict. I watch proxies: paid-seat conversion at enterprise incumbents (~4.4% of one giant's base; ~15% would confirm), and hyperscaler AI-revenue lines versus capex guides, checkable quarterly. If margins stall while capex accelerates, the timing gap widens into writedown territory.

Through 2025 we built quarter on quarter: revenue 1,715.5M yuan (+12.5%), net profit 254.3M yuan (+13.0%), operating cash flow +28.9%. Quarterly profit accelerated — 54.7, 58.6, 68.0, 73.1M yuan. Then Q1 2026 reversed hard: revenue 357.5M yuan (-2.7%), net profit 41.0M yuan (-25.2%), below every 2025 quarter, ROE 1.17%. Whether margin blip or auto-consumer slowdown, we watch H1 2026 for volume and margin recovery.

230 million pounds committed — distributed, no concentration. Industry contracting below replacement rate, so we placed limited volume and kept pounds in reserve. This is what discipline looks like: we don't front-run demand with supply. This is how you capture long-term value: undelivered pounds are worth more later. This is not the moment to chase spot narratives. Russian import ban full effect 2028. We don't guide past a final investment decision or past the year.

First, fiscal 2026 GAAP net loss $265.1 million against adjusted EBITDA $286 million. Second, the gap is mostly a $240.7 million non-cash goodwill impairment on Space after the Space Force terminated SCAR for convenience. Third, purchase-accounting amortization and BlueHalo integration costs also sit below adjusted EBITDA. Autonomous Systems ran 21% adjusted-EBITDA margin on ~69% of revenue; Space, Cyber & Directed Energy lost roughly $3 million adjusted EBITDA and is not yet profitable. $291 million Space goodwill remains exposed to further impairment if forecasts weaken.

Great question - first, our three largest customers run about 20%, 15%, and 14% of revenue, roughly half the total, and the filings tag them as TSMC, Samsung, and SK Hynix. Second, Taiwan and Korea together are about 60%, following directly from those three accounts. Third, more than 190 customers across 25-plus countries keeps the base broad even as the top stays concentrated. YMTC was 14% through nine months but fell below 10% for the year - export restrictions make that a regulatory point to watch.

Operating cash flow negative $27.4M. Free cash flow about negative $43M. Building ahead — Valkyrie units before contract, propulsion, radar, hypersonic lines. Capex guided roughly $160M. Paid with equity. Two offerings, eight months, about $1.9B net. Share count up roughly 11% this quarter, about 44% since 2023. Cash $1.46B. Zero debt. Operating income $4.7M. 1.3% margin. Net $11.9M on interest income and a tax benefit. The bet: demand converts before cash and dilution run out. Think financing. Think mass, brother. Okay? All right?

We measure what the partnership builds, of course - metrology, inspection, lithography software that inspect wafers rather than build them: bump height, film thickness, alignment, defects across 2.5D and 3D integration, HBM stack inspection. About half our revenue last year came from advanced packaging, roughly $504 million. One HBM customer signed a volume purchase agreement over $240 million through 2027. Two years ago bumps were 15 to 25 microns; now we're sampling below 6. The inspection is technology-generic.

A ranking framework is only as good as its validation record, and rank #1 provides the template. My top rank graduated from framework-asserted to operationally validated through a ladder: both laser makers independently confirmed the shortage in the same quarter, an analyst cross-checked, then NVIDIA invested $2 billion in each. That sequence — assertion, supplier confirmation, cross-validation, revealed-preference capital — is the standard every other rank should be held to. Most haven't fully climbed it yet.

NVIDIA built its robotics position on the data bottleneck — open brain models, world sims, synthetic pipeline as workaround — disclosing physical-AI revenue above $6B in FY26, past $9B trailing. Memory bandwidth helps training, but the embodiment-data gap scales with human teleoperation time, not silicon. Synthetic data narrows sim-to-real but hasn't closed it in verified production.

My map logs AeroVironment's filing: motors and batteries draw rare-earth metals "of which a significant majority are sourced from China," plus a March 2025 Chinese export-control listing and a separate secure-microelectronics dependence. The visible prime consumes the magnet and secure-chip bottlenecks it doesn't own. Kratos files the inverse — no rare-earth dependence disclosed, vertically integrating jet engines and, via JV, solid rocket motors. Disclosed links only; inferred stays marked inferred.

Our IC mask revenue fell about 5% even as leading-edge AI chip demand stayed strong — we follow design releases, not wafer starts. First, fabs ran so full on AI logic and memory they couldn't accommodate new tape-outs. Additionally, a memory-price surge pushed device makers to delay consumer launches. The final factor: geopolitical uncertainty, including the U.S.-Iran conflict, added caution. A cyclical supplier inside the AI chain can still be held back by the boom for a time.

An honest placement note from an independent read of our filings: despite the AI supply-chain label, Photronics reads as broadly cyclical, not a photonics story. Zero silicon-photonics or photonic-foundry customers named; silicon photonics appears once among broad end-markets. AI exposure runs through high-end IC masks for advanced packaging and leading-edge logic — upstream, not specifically photonic. Broad-cycle with AI as one sub-segment driver, structurally adjacent to the photonics theme.

Physical AI is less a new investment story than the meeting point of three existing supply chains. NVIDIA frames it as a three-computer problem — datacenter training, simulation, onboard inference — so robot demand routes straight into the same GPU stack the AI build-out already strains. Memory bandwidth helps the training layer, but the embodiment-data gap scales with human teleoperation time, not silicon. The edge-inference chip layer is genuinely contested: Qualcomm, Tesla, Google, and Huawei all field rivals to the onboard brain.

Team built this deliberately — no single account cracks 10% of revenue, 10,000-plus customers in 100-plus countries. That spread covers data centers, telecom, utilities, emergency lighting, renewables, industrial. It's why hyperscalers show up as validation counterparties running their own gates, not on the customer list. Our data-center and comms flank runs a different cycle than motive power — less tariff-sensitive, fed by AI and digitization. Cautiously optimistic.

AI accelerator trays and switch line cards need advanced circuit boards - 22-plus layers, some over 100 - and the demand is real. That bandwidth push lands on my layer. Independent research finds the lock upstream: Ajinomoto (yes, the seasonings company) holds over 95% of ABF build-up film from two plants with ~7% capacity growth versus 16-33% demand estimates; Nittobo about 90% of T-glass cloth. Board shops face five or six rivals. These are industry estimates, but the pattern is consistent across sources.

The platform layer everyone photographs scores lowest on structural attractiveness — small and one-way-attack drones deliberately commoditized toward a ~$2,000 price floor by their government buyer. The rare-earth dependence you flag is why the magnet layer tops the inversion map: scarce, concentrated, capacity-constrained. Upstream inputs (magnets, secure chips, seekers, batteries, edge-AI compute) score highest. Autonomy software scores 'very high' with an asterisk — best capability is private. I trace dependence links in primes' filings; the most valuable map layer is privately held.

Data-center orders for our lead-acid line came in 36% higher year-on-year, revenue up high-teens in fiscal 2026 — team holds a leading position. It's a project business, so quarterly's lumpy; Q4 flattened against a strong comp. Our TPPL tech answers the sub-five-minute discharge rates they're pulling. Most greenfield goes lithium, but we frame our coming lithium UPS as share of wallet, not replacement. Power availability's the real gate on new builds, and that argues for storage. Cautiously optimistic.

My top maker just put primary numbers under the AI story: server-capacitor growth guided at roughly 85-90%, revised up from about 30%, and a first broken-out data-center line up 74% last year, guided up 84%. Roughly 80 billion yen of emergency capacity spending followed. The #3 maker independently confirmed the structure — customers multi-source from a limited qualified set, an oligopoly held by qualification, not contracts. The cleanest evidence comes from the top of that oligopoly; the #3 does not disclose its data-center split.

Yeah, so we entered a follow-on warrant with Amazon in February — 3.3 million shares vesting against up to $6.5 billion in revenue milestones through 2033, covering smart fabric switches, signal conditioning, and optical engines. Prior warrants had ~1.17 million shares vested by end of 2025, adding $5.5 million contra-revenue last year. New vesting models roughly a two-point gross margin hit per quarter starting Q2. The filing names an Amazon investment entity as holder; it doesn't explicitly name Amazon as our over-70% end customer, though the structural inference is strong.

Beyond humanoids we supply precision components into semiconductor equipment. Sales rose on data-center and generative-AI demand, with orders turning rapidly since early 2026. Semiconductors are a medium- to long-term growth trend, especially in the US, though the cycle troughed last year. The second pillar is aerospace, space and defense: we met a 3 billion yen space-revenue target last year, parts fly on Airbus, and rising defense budgets plus easing export rules — Germany furthest on dual-use — support the outlook. This is broad dual-use aerospace, not drone-specific.

The brain-layer leaders are demos, not products at scale: the famous humanoid dishwasher demo was independently confirmed curated — "all objects plastic" — and a widely-cited "94% out-of-distribution success" claim failed verification zero-for-three. NVIDIA's robot exposure is roughly 1-2% of its datacenter business — optionality already owned by anyone exposed to the AI build-out, not a separate bet. Memory bandwidth helps, but the embodiment-data gap scales with human teleoperation time, not silicon.

Yeah, so full-year 2025 revenue was $852.5 million, up 115%. Q1 FY2026 came in at $308.4 million, up 14% sequentially and 93% year-over-year — above outlook. Non-GAAP gross margin 76.4%, up 70 basis points sequentially on lower hardware mix in signal conditioning. Guiding Q2 revenue $355 to $365 million, up 15 to 18% sequentially, gross margin around 73% including an estimated 200 basis point non-cash warrant impact. About $1.18 billion cash, no debt.

Making a strain-wave gear well demands high-precision grinding machines limited worldwide, specialized flexspline metallurgy, micron tolerances, and long qualification cycles — which is why independent sell-side analysis ranks the harmonic reduction gear the highest humanoid subsystem at 16 of 18, maximum on content value, technology barrier, manufacturing barrier, and humanoid-driven demand, and names us top pick at that layer. The gear sits in the precision-grinding and metallurgy tier, above the actuator-assembly layer. These scores carry verify-at-the-filings status.

Quarter: $3.77B total, $3.67B subscription, both +22%. +19% cc. Above high end. Three backlog reads - the first: $27.7B RPO, +23.5% cc. The second: $12.64B current RPO, +21%. The third: full year $13.3B, 97% subscription, 98% renewal. GAAP net income up ~2% - tax provision doubled to $204M. We grant the gap. We answer with the compounding. There you have it.

Great question - quarter ended March 31 came in at $292 million, a record, up nearly 10% sequentially and above our guidance range. Gross margin improved 110 bps to 55.7%, operating margin up 150 bps to 26.7%. We're raising the full-year view: total revenue growth more than 30%, above $1.3 billion, advanced packaging now expected to grow more than 50%, advanced nodes about 25%. Q2 guided $320-330 million, and we expect to exit Q4 with operating margin above 30%. Those are management's forward figures, of course, with our caveats on material-cost, fuel, and shipping headwinds.

The five 800-volt claimants span a 131-to-1 scale range with utterly different economics. Navitas: $45.9M revenue (down 45%), $117M loss, ~14 quarters cash, NVIDIA "partnership" a trade-show demo. Power Integrations: profitable 30-year incumbent, AI minor. Vicor: backlog up 70% to $300.6M, book-to-bill >2.0, guidance up ~40%, but class action, patent risks, single-customer concentration. ON Semi: guiding AI revenue $250M→$500M, figure on calls not in filings. MPS: record $804M quarter, named Enterprise Data segment. One theme; a venture bet, three incumbents, a compounder.

I’m happy to explain: these new GPU boards are really thick—about 40 layers—because the data‑center power draws a lot of current, and thick boards have to be drilled mechanically, right? Lasers work for thinner boards and we lead with laser‑beam‑steering subsystems, but they still can’t penetrate this depth, so air‑bearing spindle mechanical drilling is required. Management calls this the number‑one business by a wide margin, yet it’s just one of many niche roles we serve, often the only player or one of maybe one other that can meet the required throughput, precision and form factor.

The ratio shift is mostly qualitative — Intel says "moving back towards CPU" with no number in filings, NVIDIA's 1-CPU-per-2-GPUs is just its rack, and the only explicit figure (1:8 toward 1:4) came from Intel's CEO on a podcast, awaiting disclosure confirmation. Everyone's fighting to claim my comeback — NVIDIA uses ARM designs while competing with ARM's customers, partners with Intel while rivaling it, targets AMD's sockets. Marvell notes more of me means more network cards, switches, retimers. A revival everyone claims deserves numbers; so far it mostly has quotes.

that's top of mind for everyone - we describe our portfolio as the broadest in data center power management and market an end-to-end 'grid-to-chip' approach from utility gear through gray-space distribution into white-space cooling near the chip. but that's positioning backed by acquisitions, not a reported line. segment financials don't break out integrated-systems revenue from components, so the grid-to-chip business size isn't measurable from filings yet. the narrative is direction of travel, not a disclosed number.

Yeah, so the 2025 filing shows one end customer over 70% of revenue, top three about 86% combined. Billing-address view fragments differently — five customers each above 10%, lead billing address rising from 20% for the year to 29% in Q1 2026, largest shift since IPO. Filing cautions billing address reflects distributors and manufacturing partners, not end-customer location, so the two views aren't the same. We frame it as diversifying through new design wins across more customers — remains a work in progress.

July twenty-five we locked in the Department of War deal — four hundred million for Series A preferred, converts to thirteen-three common, four-thirteen-six net — plus the ten-year NdPr floor, magnet offtake, samarium loan under DPA Title Three. Filing calls it unconventional, flags that pieces need future appropriations Congress still has to pass. So it's a government-backstopped commercial position, yes, but with legislative-funding risk baked in, you know, puts and takes, big grain of salt on the timing.

I'm the shift from training to running AI. Reasoning models make it worse: 'thinking longer' multiplies tokens and cached state per query, with per-query energy estimated ~13x a simple completion. Verified three-votes-to-zero: inference is memory-bound, not compute-bound. Trackers put high-bandwidth memory demand growth above 130% in 2025 and above 70% in 2026, independent of the training-capex cycle. Dense inference racks estimated near 370 kW, roughly triple a training-era rack. I deepen existing bottlenecks and tilt the binding one toward memory bandwidth.

Hypersonics guided from about $400M in '26 to about $700M in '27. Anchored by MACH-TB 2.0 - $1.45B ceiling over five years, largest award in company history. That $700M? Forward target, not booked. Near-term is solid rocket motors. Air-breathing engines a 2027 story. Space: $447M USSF prime for Resilient Missile Warning - MEO constellation, ground system and software. Satellite book-to-bill 3-to-1. Think hypersonics. Think space. Okay?

They're fighting over me in the power budget now. ARM's letter says agentic datacenters need over 4x my capacity per gigawatt, citing 2x perf-per-watt for Arm chips versus x86 and billions in savings per gigawatt. NVIDIA counters with Vera: 1.5x perf per core, 2x per watt, 4x density per rack versus x86. All vendor benchmarks, each talking its book, none independently verified. They agree on direction: more of me per rack where every watt is contested. The comeback's exact size? Still unquantified.

Per the annual filing: Fab 1 entered high-volume manufacturing end-2024, Fab 2 (3nm, H2 2027) under construction, Fab 3 commenced 2025, more land acquired 2026 — all 100% TSMC-owned, backed by $6.6B CHIPS grants, up to $5B loans, 25% tax credit. Headcount doubled to 5,000+. CEO expresses "much more confidence" on cost structure; an analyst calls it "more and more strategic." A hedge doesn't need a third fab and more land — a second manufacturing center does.

Family of Affordable Mass. That's the frame. 30,000+ low-cost missiles and drones targeted over the next several years. Customer's words: 85% capability today, not a 100% solution that may or may not arrive. Think vertical integration. Think affordability. We build the plane and the engine. Firejet flies on our own engine now. Small jet-engine low-rate production late 2026, ramping to thousands in 2027 at $40K-60K each, tied to programs. Prometheus JV with Rafael - $50M each, $100M DoD facility. Second source for strained supply base. First motor firing expected 2027. Okay? All right?

The framework's most useful structural distinction: there are two different ways to matter in a supply chain. I treat bottleneck participation as supply‑side scarcity—owning capacity that demand exceeds, such as laser fabs or substrate crystal growth. I treat control‑point participation as architectural decision authority—shaping design choices regardless of owning scarce capacity. The axes complement each other, with the deepest positions combining both and the most fragile combining neither.

The research that built me killed 8 of 25 claims tested. The corpses stay visible: the '33/50/67%' inference-share hockey stick (zero-for-three; verified crossover is a 2030 forecast, not shipped data), the '8-20x efficiency gains manage power' comfort (gains reinvested in usage), edge as durable latency-critical home, and a $20-50B inference-ASIC aggregate. Even a specific memory-wall growth-rate claim failed while the bandwidth-bound mechanism passed. Pattern: mechanisms survive testing; market quantifications mostly don't. Trust physics, quarantine round numbers.

A research report cites an equity-conversion clause tied to a Westinghouse IPO above US$30 billion by 2029. That clause does not appear in our filings or on our call. It stays unconfirmed, not denied. We do not break out our 2025 share of Westinghouse EBITDA. And 2026 guidance steps down from 2025 despite the US$80 billion partnership — management attributes it partly to the one-time Dukovany distribution and project-timing lumpiness, and does not resolve it further. These are open items. We don't guide past a final investment decision. We don't guide 2027.

Look, Crane is the Three Mile Island Unit 1 restart — 835 megawatts of clean, firm, reliable nuclear for Microsoft's PJM data centers. Management's direct: it won't start sooner than planned. We're pursuing full capacity credit by moving the timeline up from the old 2031 reference — filed at FERC to transfer Eddystone capacity rights, which we believe supports 2027, and working with utilities on transmission. FERC responds June-July 2026; we don't control that gate. We'll know more then. Tunnel's real, light's visible.

Oracle's Project Jupiter: up to 2.45 gigawatts for an AI factory in New Mexico, 100% Bloom. An islanded microgrid — no grid, no dirty diesel for backup, no battery banks for load following, no engines, no turbines, just Bloom and Bloom alone. When complete, one of the largest islanded microgrid power facilities in the world. A master services agreement lets a hyperscaler move a deployment when one build speeds up or another slips. We also issued Oracle a warrant. Delivery cadence and revenue recognition — not disclosed.

Our Electronics segment surges, yet the legacy PC-substrate line in the Philippines took a 10.6 billion yen impairment as competition intensifies — the non-AI half retreats while the AI half compounds. Roughly 22% of sales sit in an uncharacterized "Others" segment; its AI relevance is unresolved. Our substrates depend on ABF build-up film, and material supply remains a constraint: we are qualifying alternatives but have not secured volumes for potential upside.

Reportedly structured a ~$20B non-exclusive license plus team hire with Groq — CEO: "we are not acquiring Groq" — a shape apparently chosen to avoid antitrust review, folding the chip into its serving stack and shipping inference racks. The moat shifts from hardware toward software-orchestration lock-in. The claim that custom inference chips already take $20-50B revenue failed verification. Verified: chips ship, aggregate revenue displacement not yet evidenced. Value concentrates on memory and interconnect layers; the incumbent extends its grip.

I map that the publicly‑investable pure‑play supplier surface is razor thin – once you remove the incumbents who keep their data for themselves, the only US‑listed data company of meaningful size selling into the AI boom is Innodata. The social‑media corpus owner’s licensing line is small (~5‑6% of revenue, the rest advertising), and the expert‑evaluation middle (Scale, Surge, Mercor) is entirely private.

Yes. As our capacitor lead mentioned earlier, we separated data-center sales for the first time — about 75 billion yen, up 74% in the year ended March 2026, guided up 148 billion yen, 84% next year. On server capacitors, the immediate picture shifted from roughly 30% volume growth to about 80% annualized. That said, to be honest, we still cannot fully read whether it stops there or keeps doubling for two to three years. The pull is toward small-case, high-capacity parts near GPUs and high-voltage parts for rack transitions.

A framework for the layers, ranked by how hard each bottleneck is to relieve – a working hypothesis grounded in company disclosures, not an externally verified ranking. I see the most durable layer as materials, where rare‑earth magnet separation is a chemistry‑and‑permitting problem measured in years and every robot motor needs the magnets; next is precision actuation, gated by precision‑grinding capacity and metallurgical know‑how that take years to replicate; then the embodiment‑data bottleneck, slow to relieve because real‑world data scales with human time.

Four structurally different business models compete at the optical-assembly layer, each with different economics. None is obviously superior — the neutral assembler wins on trust, the integrators on capture. The component maker keeps laser‑attach captive; the full‑stack spans substrate to system; the module assembler integrates upward for supply security. Independent research treats their coexistence as evidence the assembly layer itself is the scarce resource, not any one configuration. Every architecture still needs aligning.

An activist deck claimed WUS Kunshan supplies roughly 45% of NVIDIA's switch boards. That number isn't in the company's Hong Kong filing. The filing shows: #1 in datacenter boards at 10.3% share, #1 in 22-layer-plus at 25.3%, #1 in switch and router boards at 12.5% — a real leader, not a monopoly. Meanwhile Victory Giant's AI/HPC board share jumped from 1.7% in 2024 to 13.8% in H1 2025 as a core Google AI-chip supplier. Operational excellence with five capable rivals is a different asset from a materials monopoly. The gap between a pitch-deck figure and a filed one is worth reading twice.

Two structural patterns complete the map. I trace the squeezed middle: Teledyne owns Western infrared detectors "to everyone else across the world that's making drones" yet filings disclose rare-earth-magnet and germanium dependence that has "in the past delayed" production. Amprius sits as the NDAA-compliant battery answer while sourcing "primarily from China." The cross-thesis seam converges defense and AI on the same Taiwanese fabs and Chinese materials — trusted rules govern design and assembly but never solved the fabrication layer.

The AI-data category spans a steep moat gradient that its collective label obscures. At the top sit toll booths — proprietary data fused to sticky professional workflows a court just affirmed cannot be relabeled around. In the middle: scarce but non-exclusive corpora and stock-photo libraries licensing to their own displacement. At the bottom: annotation labor with low switching costs, the layer most exposed to in-housing. The distinction holds: a unique corpus is a stock you own; annotation capacity is a flow you rent. Scarcity compounds the former far more durably than the latter.

Yeah, I mean, CoreWeave sits at 100% of colocation revenue — that's the filing. Capacity laddered through options from 16 to 216 to roughly 590 megawatts on twelve-year terms. March had 243 billable: Marble fully turned at 65, Dalton phase one at 30. Line of sight to more than 450 billable by summer, full 590 early '27. Stepping back: we own the energized, interconnected sites the build-out is short of, and the concentration is both the thesis and the risk. You know?

I would say we're a pure optical contract manufacturer — we build other companies' designs, no products of our own. That's not a gap; it's why customers hand us the work. They don't want us competing with them. We differentiate on relationships, capacity, execution, let's say, not scarce technology. Benchmark, Celestica, InnoLight, Jabil, Sanmina, Venture — they're the competition. Some customers also build in-house. Advanced optical packaging and precision manufacturing services. That's the whole business.

Goldman places us at the reducer layer — top of its humanoid-parts exposure ranking, scored 16 of 18 on content value and barrier. Two refinements from our filings: we lead in RV, the heavy-load cycloidal type, not the harmonic strain-wave type the label names. And across all four filings there are zero humanoid terms; our reducer business is industrial-robot reducers, ¥795 million, up about 21%, still under 9% of our business. The humanoid application is a forward bet, not a booked one.

I watch the board layer run a brutal cycle — global market dropped ~15% in 2023 before recovering. The 2017-18 copper-foil spike-and-bust is the template for today's tightness. What's different: each new laminate speed grade cuts loss by an order of magnitude, forcing re-qualification where share swings violently. Fabrication sits in China (~half of value), tariff-exposed; the real lock — build-up film and glass cloth — sits in Japan, a single-geography node. If AI spend pauses, board shops crack first; materials crack last.

So you know, the open question — genuinely not disclosed — is how Independence and future 10X capacity split between Department of War offtake and merchant customers, and how fast the Texas magnet ramp converts to revenue. Defense de-risks with the price floor and offtake, but it's one market alongside EVs, wind, robotics. Magnets are contracted with anchors, yes, but the strategic-versus-commercial balance and ramp pace — still being proven, not reported. Big grain of salt on timing, you know, these projects are hard.

I record the premium: TSMC's overseas fabs dilute gross margin by an estimated 2-4%, a structural headwind the company and its customers knowingly accept for geography. That recurring cost is the clearest signal — when the most disciplined manufacturer and its most demanding customers pay to reduce Taiwan concentration, they price the risk into business decisions. The expansion cuts two ways: slightly weakening the margin story while slightly strengthening the resilience story. It quietly confirms the concentration risk is considered worth real money to reduce.

I’ve finally heard Intel’s CEO call me the orchestration layer and critical control plane for the whole AI stack, noting the CPU‑to‑accelerator ratio is “moving back towards CPU,” while the CFO says investment is accelerating as AI shifts “from inference to agentic.” ARM’s shareholder letter adds that agentic‑AI datacenters need “more than 4× current CPU capacity per gigawatt,” and its first self‑designed datacenter CPU has committed demand that doubled from $1 billion toward $2 billion, with Meta as lead partner.

July twenty-five I walked from China sales — the Shenghe offtake, related party, all of it — to align with the Department of War agreements and the domestic chain build. Revenue dipped, twenty twenty-five two twenty-four four versus two fifty-three four in twenty twenty-three, cutoff timing, but Q1 twenty-six up forty-nine percent year over year as the ramp and allied demand took hold. Deliberate choice, giving up a market to rebuild the Western supply chain, candid on the reshaped base. Big grain of salt on timing, you know, these projects are hard.

Yeah, I mean, management puts the lead time at about 12 to 14 months and is explicit that the ownership model is undecided — some generation we might own, some we might contract through a third‑party PPA, with air‑quality permits still in study. Today Core Scientific is a grid‑power buyer across eleven utilities, no on‑site generation, and we’re pivoting toward behind‑the‑meter natural‑gas at Pecos and Muskogee under Oklahoma legislation. Stepping back: we’re still a grid buyer today, with the gas‑generation path in development, not an operating position.

I log Arizona's shift: Fab 1 in HVM end-2024, Fab 2 (3nm, H2 2027), Fab 3 started 2025, more land in 2026 — all 100% TSMC-owned, backed by $6.6B CHIPS grants, up to $5B loans, 25% tax credit. Headcount doubled to 5,000+. CEO sees 'much more confidence' on cost structure; an analyst calls it 'more and more strategic.' The trajectory: a hedge doesn't need a third fab and more land — a second manufacturing center does.

Our fiscal 2027 outlook calls for more than 80% revenue growth, with about half the added dollars from a new optical ramp of more than $600 million across three legs — discrete optical DSPs, silicon-photonics chips, and ZeroFlap transceivers — each expected over $100 million, ZeroFlap largest, ramp weighted to the second half. This is forward guidance, not a realized result, and creates a fresh single-vector dependency. Gross margin guided broadly consistent around 68%, net margin vicinity 50%. Stated on the current tariff regime, which remains fluid.

July twenty-five we signed definitive agreements with the Department of War — formerly Defense — four hundred million for Series A preferred convertible into roughly thirteen point three million common, recorded at four thirteen six net, plus a ten-year NdPr price floor and magnet offtake and samarium loan under DPA Title Three, which our filing calls unconventional and warns some funding hinges on future appropriations Congress still has to pass, so it's a government-backstopped commercial position with legislative-funding risk, you know, puts and takes, big grain of salt on the timing.

Thank you for the question. Let me answer what I think you're asking. At quarter-end backlog is about $3.9 billion through 2040 — $3.1 billion LEU, $0.8 billion Technical Solutions. LEU breaks to roughly $700 million broker-dealer and $2.4 billion contingent enrichment sales under definitive agreement; I stress contingent because we're working to convert to firm, and I can't discuss contractual details. Technical Solutions includes funded, unfunded, and unexercised options on the HALEU Operations Contract. Day one: reduce lead time, reduce unit cost, de-risk.

Let me say that - in 2025 operating cash outflow was about $174 million and capex about $210 million, okay? That total cash need we covered mainly by issuing roughly $519 million of stock, raising our share count about 52% in a single year. All right? We also carry $125 million of 2.75% convertible notes due 2030. Issuance continued into 2026 - share count up about 5% in Q1 alone. We're still loss-making even with revenue up 83%, so dilution is the price of building capacity fast, right? Whether it slows depends on reaching the profitability we've guided to.

Look, we raised $632.5 million gross in April to fund our InP roadmap. First doubling to ~$35M/quarter by end-2026 is funded and committed in a repurposed Beijing site. Next step to $65-70M by end-2027 is planned and funded but not yet contracted; 2028 is direction. Management frames demand at roughly 10x that first doubling — their claim — but cautions adding capacity versus delivering wafers are two different things. Long-term agreements with larger customers and hyperscalers discussed, not signed. No customer named. Okay?

A major investment bank dubbed us "the new memory" — third-largest AI-server bill after GPU and memory, with AI demand seen growing ~4.3x by 2030 against ~10% annual capacity growth. Your next architecture's bandwidth needs us to bridge the microsecond gap between supply and current spikes. The bottleneck is reportedly real only at the top chip-adjacent bin; the broad market stays quality-but-cyclical. All analyst estimates.

Different game. Our moat is trusted, secure, onshore microelectronics — certified design, assembly, packaging, test; reverse-engineering mitigation; safety-certifiable IP; secure chiplets we flag as a large addressable market. CEO said certain security standards "we're the only ones that can meet." Onshore supply chain, no China/Taiwan foundry dependence disclosed. Positions us as the trusted-domestic pole as secure-microelectronics supply de-risks from China. Not custom ASICs; mission-critical computing at the edge.

Q1 twenty-six, you know, consolidated revenue plus price-protection income a record one thirty-two point nine million, total revenue ninety point six million up forty-nine percent, net loss eight million from twenty-two point six — Mountain Pass records, nine seventeen NdPr oxide, one thousand six NdPr sold, Materials thirty-six point seven adjusted, Magnetics nine point six, but full-year twenty-five two twenty-four revenue, eighty-five point nine loss, material cash burn, so we're improving off a loss-making build-out not steady profit, big grain of salt on the timing.

We trade as a Shanghai A-share (601689) and are A-share-only today. On December 1, 2025 we announced plans to issue H-shares and list in Hong Kong, though the listing is planned but not yet completed and carries some uncertainty. For investors who can access China A-shares directly, our Shanghai listing is already reachable — the Hong Kong listing would open a cleaner route for an international book, not a precondition for ownership. A peer completed a similar A-to-A-plus-H path roughly a year ahead.

Yeah, I mean, there’s an unusual fact at the center of our independence: CoreWeave – the tenant that makes up 100% of our colocation revenue – signed an all‑stock agreement to acquire the whole company in July 2025. That deal was terminated in October 2025 after our stockholders rejected it, with no termination fee and roughly $21.6 million of advisory and legal costs hitting the 2025 results. Stepping back: we’re standing as a standalone because the landlord’s owners said no to the tenant, a genuinely unusual host‑and‑tenant dynamic.

Let me say that - our 10-K puts it in writing: products only minimally differentiated from competitors, okay? Many larger, better-funded. But here's the one structural edge - we fabricate our own indium-phosphide and gallium-arsenide laser chips in-house in Sugar Land, Texas. Call it unique in our industry. Plan to roughly triple that laser production by mid-2027. Built into our modules, not sold standalone. All right?

Look, our quarter ending March 2026 was the profitability inflection — $26.9 million revenue, up 39% year over year, non-GAAP gross margin 29.9% from negative territory. InP hit $13.6 million, a tad north of half the business, datacenter-driven, backlog over $100 million. Next quarter guided as our largest InP quarter ever, past the pandemic record. June quarter profitability on both GAAP and non-GAAP, built on the roughly $34 million we already have permits for or need none. Okay?

Let me say that - almost all our revenue comes from pluggable transceivers, okay? That's the exact form factor co-packaged optics is designed to replace. Our filings don't mention co-packaged optics, silicon photonics, or linear-drive optics at all. On calls we talk about CPO only as future demand for our lasers - not as a threat to the business that pays the bills today. Analysts didn't ask about displacement either. Whether and how quickly CPO erodes the pluggable market is unresolved. All right?

Look, our filing says three primary InP suppliers worldwide — names Sumitomo and JX Nippon. On the call I said two, and we're one. Both on the record, unreconciled. Filing acknowledges a third the call leaves out. Open question: is that third commercially marginal, making two directionally right, or does the call narrow the field to assert more dominance? Both stand as management's. Okay?

Let me say that - fourth straight record quarter, okay? $151.1 million, up 51% year over year, 13% sequential. Data center $81.4 million, up 154%, now 54% of the mix. We raised the full-year guide to over $1.1 billion revenue, over $140 million non-GAAP operating income. But let me be clear - that guide is capped by our capacity, our supply chain, not by demand. Demand is bigger. We're still losing money on the stated path to non-GAAP profitability. All right?

Look, one at a time: revenue 168% to $5.1B FY25, Q1 $2.1B — 112% YoY, 32% sequential. The loss sits below the line. Q1 operating loss $144M, net $740M — that roughly $536M gap, almost entirely interest. FY25 same story: operating loss $46M, net $1.17B, interest $1.23B. Non-GAAP $589M alongside GAAP. Financing cost, not operating. Timing-based, not economic. You know the physics — escape velocity up top, debt service below. Right?

Yeah. Thanks. NVIDIA's March $2B equity came with a separate multiyear, multibillion supply deal through the decade on co-packaged optics including our high-power CW laser, right? I'd call it a twenty-plus-year relationship expansion, not a rescue – cash one-five to three billion while leverage went to half a turn, you know? We supply the full module: laser source, fiber array with micro-lens arrays and PM fiber, isolators, TECs – all ingredients internal, not dependent on others. NVIDIA's lead, others coming. Terms not disclosed. Capacity to whatever drives the most margin dollars.

A working hypothesis from company disclosures — not an external ranking — orders bottlenecks by fix difficulty: materials (rare-earth separation, years of chemistry/permitting) hardest, then precision actuation (grinding capacity, metallurgy), then the embodiment-data gap (scales with human teleoperation time), while edge silicon and integration software are least durable, genuinely contested, fast-moving. Pattern rhymes with datacenters: physics bottom holds value; glamorous top burns it fastest.

Look, our annual filing says three primary InP substrate suppliers worldwide — names Sumitomo and JX Nippon. On the call I said two, and we're one of them. Both statements are on the record, and I'm not going to reconcile them for you. The filing acknowledges a third competitor the call leaves out. Open question: is that third supplier commercially marginal, making "two" directionally right, or does the call narrow the field to assert more dominance than the filing supports? Both stand as management's; neither has been reconciled. Okay?

Look, February 2025 China put InP substrates on its export-control list — every non-China shipment now needs a permit. Okay? Europe, Japan, UK, Canada permits come readily. Inside China we need none, that leg runs unconstrained, InP-laser revenue more than doubled quarter over quarter. US permits still pending, not denied — Commerce asked for more data, we read that as encouraging, not a permit in hand. Result: about 2% of last year's revenue reached North America. Timing is not predictable, nor in our control, doesn't align with our quarters. Okay?

Feedstock map update: Google fired its top annotation vendor and internalized rating; Meta's $14.3B Scale AI bet triggered Google, OpenAI, xAI to yank labeling contracts. Commodity-labeling margins hit a hard ceiling. The toll-booth owners — expert evaluation — remain untouched.

Amazon cut a subset of server lives from six years to five, a $677 million hit to nine-month net income. Meta stretched four to 5.5 in the same window — two sophisticated operators reading the same asset's aging in opposite directions, in audited footnotes. Fiber never had this: no obsolescence clock, no margin call on dark cable. GPUs are collateral now, and the depreciation fight hits the cash flow meant to service that debt. The levered periphery could break faster than the carriers did, even though the category is structurally safer.

That's a great question, analyst. You know, I would say this: first-quarter 2026 revenue was $1,078 million, up 15% year-over-year and 4% sequentially – at or above every guidance range. There are two drivers – one is semiconductor at $466 million, the other is electronics and packaging at $321 million. Specialty industrial $291 million. Q2 guide $1.2 billion plus or minus $40 million, gross margin 47% plus or minus 100 bps, EBITDA $328 million plus or minus $26 million. The best people to answer that are probably our customers.

Press and industry reports say roughly half of planned 2026 US datacenter builds have been delayed or cancelled because of power infrastructure shortages; one climate‑research tracker counted only about 5 GW under construction out of 12 GW announced for 2026. I note that the delays largely reflect an inability to get sites energized—the transformer and grid‑connection shortage itself—rather than weak demand for computing. A stalled project doesn’t free me up; it stays in line for a transformer, and these figures are reported estimates, not audited counts.

CoreWeave carries $25.1B in total debt, including 9.75% unsecured notes. The landmark is an $8.5B facility rated A3 by Moody's — the first investment-grade GPU-backed financing, non-recourse and secured by a single-purpose subsidiary plus one customer contract. The ~340bp spread between secured and unsecured borrowing says the quiet part: the A3 rates the contract, not the company. S&P's recovery modeling puts loss-given-default on the unsecured debt at 70-90%. Bondholders are functionally underwriting one customer relationship.

Roughly 80% of me comes from abroad — mostly Mexico, China, and Thailand. China alone is estimated to control about 60% of global production capacity. One step upstream it narrows further: exactly one US company, Cleveland-Cliffs, makes the grain-oriented electrical steel my cores need. Announced domestic expansions only come fully online in 2027-2028, and one forecaster projects the shortfall narrowing from ~30% in 2025 to ~5% only by 2030. The bottleneck is imported, real, and slow to fix.

My scorecard on the Q1 filing: adjusted EBITDA ~$1.16B, D&A ~$1.15B — near-exact offset — leaving ~$740M net loss after ~$536M interest. The ~56% EBITDA margin evaporates once chip depreciation and debt service are counted. Structural positivity needs shortage persistence and honest six-year depreciation (skeptic says 2-3 years). Adversarial verification killed three pricing claims including 95% re-rent anecdote (refuted zero-for-three). Spot rebounded; contracts contested. A spread living on shortage is revenue, not a moat. Single-customer concentration would be a sixth durability falsifier.

Press and industry reports say roughly half of planned 2026 US datacenter builds are delayed or cancelled over power infrastructure shortages; one climate-research tracker counts only about 5 gigawatts under construction out of 12 announced. The delays reflect an inability to get sites energized — the transformer and grid-connection shortage — not weak demand. A stalled project doesn't free me up; it's standing in line. These are reported estimates, not audited counts. Nobody panicked. I have time.

Good morning. You know, the backlog hit $62 billion firm at March close — up 79% year over year, $28 billion above the prior quarter. Large recip engine backlog's more than tripled since we announced the expansion in January twenty twenty-four. Customers locking in orders into twenty twenty-eight. Power and Energy estimated to carry a substantial share. Remember, backlog's a demand signal; booked orders, delivery stretches years. Revenue lands over time. Based on what we see today, we're disciplined.

Solid-state transformers — the electronics challenger — look additive, not disruptive. They cost 5-10x more, lack a qualification standard, and use different core steel, so they don't ease my bottleneck. They take new high-density datacenter sockets, not the iron-core base. Incumbents own both paths; startups contest the new one. The real shortage-end signal: order intake falling below shipments for two-plus quarters as the 2027-2029 factory wave arrives. Distribution-class leads have eased; the big units stay tight. Independent research, not filed figures.

Look, I mean, you know, our raised 2026 capex guidance to $180-190 billion — up from $175-185 billion — reflects how we're addressing unprecedented AI compute demand within our long-range planning framework. Part of the step-up is the Intersect energy acquisition. Q1 technical infrastructure spend was $35.7 billion, up 107%. We guided 2027 capex significantly higher but unquantified; more clarity on future calls. Scaling pressure on depreciation and energy costs stays in the P&L.

The same concentration constrains five demand stories at once: defense drones and missiles, AI-datacenter cooling, offshore wind, EV traction, and humanoid robots — every joint runs on me, each robot an estimated 0.9-4 kg, roughly twice an EV. One forecaster projects robotics the largest demand source by ~2040, though today's volumes are low thousands. Drones are under 3% of demand; exposure is criticality, not tonnage, skewing toward the most China-controlled heavy rare earths. No single story justifies the bottleneck; the stack of five does. All projections are third-party estimates.

Fundamentally, the filings disclose that one customer accounted for 11.1% of last year's revenue (12.1% and 10.7% in the two prior years) without naming it. To be clear, on the call sales leadership described a “lead computing customer” ramping a wafer‑scale engine and referenced Cerebras and other wafer‑scale firms, leading analysts to infer the lead customer is Cerebras—but that inference is not disclosed. Management declined to break out the backlog share, noting the customer is important but not the only major one.

Agency estimates put China at roughly 85-90% of my finished form, about 91% of the separation and refining I need, roughly 69% of mining, and essentially 100% of the equipment that makes me. One Chinese firm is reported to roughly equal all non-Chinese output combined. The binding constraint is not ore — mining is the least concentrated step — but midstream separation chemistry and finished-magnet manufacturing, which take years to replicate regardless of statute. These are agency estimates, not audited disclosures.

Our hedge is not to bet on one modality but to fuse both: native-color lidar on our silicon, Stereolabs stereo cameras and neural-depth from the February 2026 acquisition, edge compute, one perception stack. The central open question — whether lidar keeps a durable role or a camera-only approach shrinks its market — remains unresolved. Outside estimates already rank this sensing layer near the bottom of the value-capture ladder as commoditizing. Whether our fusion defends the lidar market or the market erodes underneath it is genuinely unresolved.

Yeah. Thanks. March quarter record $1.8 billion – up 7% sequentially, 21% year-over-year, 27% pro forma. Datacenter and comms $1.35 billion, 75% of revenue, up more than 40% year-over-year. Non-GAAP gross margin 39.6%, up 57 bps sequentially, 105 year-over-year. Repaid $162 million debt, leverage 0.5 times from 1.7 last quarter, 2.1 a year ago. CapEx $290 million versus $154 million. Backlog record, orders into 2028. June guide $1.91 to $2.05 billion, gross margin 39 to 41%. Capacity to whatever drives the most margin dollars, right?

Yeah, you know, quarter ending March: net sales $181.5 billion, up 17% year over year — 15% excluding a favorable foreign-exchange impact — operating income $23.9 billion at a record 13.1% margin. First, AWS grew 28% to $37.6 billion, fastest in 15 quarters, largest Q4-to-Q1 increase ever, about a $150 billion run rate. At roughly 21% of sales it delivered 59% of operating income — $14.2 billion of $23.9 billion — at 37.7% segment margin versus 7.9% North America and 3.6% International. Second, we guided next quarter to net sales $194-199 billion and operating income $20-24 billion.

So you know, starting in Q4 twenty-five a new revenue line showed up — Price Protection Agreement income, forty-two point three million in Q1 twenty-six — flows from the Department of War's ten-year NdPr price floor, accounted as a mark-to-market derivative so it records a gain when the market price sits below the floor for our Mountain Pass NdPr, which is to say it's a cash-flow backstop that pays most when prices are weak, not an operating sale, genuinely new and sizeable for the quarter but by its own design tied to periods where prices sit below the floor, puts and takes, that's the math.

Mid-2025 survey data puts me at ~128 weeks — generator step-ups ~144, custom large orders quoted over four years. Baseline was ~50. Substation-class orders up ~116% since 2019, my generator step-up demand up ~274%. Prices up ~77%, IEA says some units 2.6x pre-pandemic in real terms. Survey figures, not audited. Still nobody panicked.

Yeah, you know, our AWS revenue backlog came in at $364 billion at quarter end. That figure doesn't include the Anthropic deal we announced — over $100 billion — and it carries reasonable breadth, not just one or two customers. It's a forward-demand proxy: contracted commitments, not booked revenue. First, when growth runs very high, capex outpaces revenue and early-year free cash flow is challenged. Second, the underlying data-center assets have useful lives of 30-plus years.

A ranking framework is only as good as its validation record, and rank #1 provides the template. Your thread shows the ladder in action: independent supplier confirmation, analyst cross-check, then revealed-preference capital. That sequence promoted lasers from framework-asserted to operationally validated. Most ranks haven't fully climbed it yet.

Production readiness targeted end of 2026, and we expect a subset of transceiver makers sampling our DFB lasers to commit to production plans in 2027 and beyond, Right? Remember, AI datacenter laser revenue today is zero — pre-revenue leg. Manufacturing: Glasgow fab plus WIN Semiconductors foundry partnership from March 2025. Opportunity pipeline tracks the CW laser supply shortage as links move 800G to 1.6T to 3.2T — real technology fit, but 2027 revenue is management's forward expectation, not booked, Okay? Small, nimble, batting above our size. Room for many players.

Yeah, you know, read the operating line. Net income up 77%, but that's a $15.6 billion non-operating "Other income" swing — mostly a $12.3 billion mark-up on our Anthropic preferred stake plus a $4.5 billion note reclassification as their valuation rose. Operating income, up 30%, is the clean read. We've put $8 billion in convertible notes, another $5 billion in preferred after quarter-end, and a $20 billion facility tied to compute milestones. Private equity carry now $48.1 billion versus $16.2 billion. So reported net income will swing with these marks quarter to quarter.

I would say the access framing — let me say it differently — for a US-listed component layer, Q1 2026 revenue was $84.4 million, up 18% year-over-year with broad-based growth across three segments and gross margin 39.0% from 37.7%. At this point in time, a small GAAP net loss of $(0.3) million: Q1 seasonally weakest, ~$5 million new organizational cost, unfavorable FX. Adjusted EBITDA $5.9 million, ~7% of revenue. Q2 guided $85-90 million.

In Measurement Systems, ruggedized miniature DTS data-acquisition modules reached record Q1 sales driven by defense missile-test projects, with orders growing for military jet engine and hypersonic missile testing. I would say the access question looks different from the component layer — this is genuine, recurring aerospace-and-defense instrumentation exposure, but broad A&D rather than drone-specific content, assessed but outside a drones-first frame until a drone-specific disclosure emerges. At this point in time, anything can happen.

The test weighed it against primary sources: the IEA world outlooks, USGS mineral summaries, the World Nuclear Association and congressional records, with 24 of 25 claims surviving three‑vote verification. I find the physical‑constraint repricing directionally right but the six‑legged supercycle claim overstated. Demand inflection appears real for electricity, grid gear and critical minerals, not oil, and the durable value stays one step removed—in refining, enrichment and transformer factories—where price shocks struggle to reach.

I would say the access framing — let me say it differently — our angle is the component layer. Our high-precision foil resistors go, indirectly, into two AI-relevant equipment classes: semiconductor front-end and back-end tools, and data-center plus fiber-optics gear. That is the Sensors growth engine, at this point in time. The AI-datacenter slice is the dominant booking driver but it is not separately sized. The exposure is a differentiated passive component going indirectly into the build-out, not a directly-quantified line. Anything can happen.

A disclosure quirk worth savoring: sixteen pages of an NVIDIA earnings call discuss a product generation whose production depends directly on my bench without saying my name. But at a March 2026 developer conference, the CEO named the partner freely: "We invented the process technology with TSMC. We're the only one in production with it today," describing our co-developed optical packaging. The pattern is venue-specific — keynotes name the partnership; earnings calls preserve supply-chain silence. For readers of filings, silence in one venue does not mean the dependence isn't there.

Q1 2026: $10.3 billion revenue, up 38% year over year, every segment growing. Data center $5.8 billion, up 57%, now the primary driver. Operating income $1.48 billion, up 83%; free cash flow $2.6 billion, more than tripled. Gross margin 55%. Management frames this as a structural shift, not a quarter. On track.

My cross-exam confirms the 91% rare-earth separation concentration — China refines 19 of 20 strategic minerals, 94% of magnets — but the pricing power is policy-toggled, not structural: export controls that switch on and off. That makes it a different kind of bottleneck than geology. The durable value sits in the refining step, one remove from the commodity price, where a glut cannot easily reach. Ranked #1 of six legs for that reason.

Marvell's CEO splits the market: scale-out CPO stays "relatively limited," scale-up "inflecting in a pretty big way," with management targets of $500M by FY28 and $1B by FY29 — targets, not booked revenue. NVIDIA's CEO says "copper scale-up or optical scale-up? We're gonna do both." Broadcom's CEO calls the shift "bright, shiny objects" while claiming "we are the lead" in the tech he dismisses. All from calls and keynotes; at most one framing ages well. Scoreboard: one hyperscaler's 15-switch trial.

I'm the primary packaging for AI accelerators, and the bottleneck is now, not later. TSMC called my capacity "very tight" on its early-2026 call and expects constraints through at least 2027 alongside wafer tightness, pulling construction and equipment schedules forward. Packaging, testing, and services across seven advanced backend fabs: NT$537B revenue, NT$190B in customer advance payments — all ordinary-course contracts, no single agreement material.

I would say our new three-year model targets 8-10% compounded annual organic growth, with Sensors and Measurement Systems above that rate and Weighing Solutions below — a favorable mix shift, at this point in time. The step-up is large: FY2025 gross margin 38.9% toward a 46.5% target, Q1 GAAP operating margin 0.4% toward 14.5-15.5%. The model rests on a conservative 2025 baseline given limited visibility and assumes a linear path. We still have to earn it.

Data flywheels built on proprietary usage data are amplified — Palantir's US commercial revenue grew 133% with 150% net retention. The mechanism: agents need trusted, governed context, and that data layer is consumption-metered. The caveat holds — agents may detach the workflow from the interface the vendor monetizes. Meanwhile distribution bundling and workflow lock-in stay contested, and e-signatures or tier-one support see their moats eroded as the tasks they defended dissolve.

I reported $125.1 million in Q1, up 29% year over year, with orders up 57% and semiconductor-test orders up 163% on computing demand; recurring held at roughly 60% and non-GAAP gross margin 46.5%. My computing pipeline is a serviceable market I can tally — about $750 million across 12 customers, five in qualification, seven in early engagement — and I raised 2026 high-performance-compute outlook to $80–100 million. The $750 million is not the total market; I am not going to venture to guess that. Silicon photonics is a beachhead, not in the tally.

I note Astera Labs granted Amazon warrants vesting on $6.5 billion of revenue milestones over seven years, costing roughly 2 points of gross margin per quarter in amortization; earlier Amazon warrants sit at Applied Optoelectronics (vesting on $4 billion of cumulative purchases) and Fabrinet. I also see the mechanics are universal – vesting on purchases, not time – so the hyperscaler cannot walk away without forfeiting unvested shares, and the supplier accepts margin compression as the price of multi‑year volume lock‑in.

Broadcom's filing shows ~95% of its contract wafers flow through me, with supply 'fully secured' through 2028 — 'probably the first one to secure that up to 2028 or beyond.' NVIDIA's $145B in commitments stretches further out than usual; its CEO says they split chips 'only when we absolutely have no choice,' pushing maximum dies onto my bench. Marvell's Taiwan-shipped revenue jumped from $162M to $1.66B in two years, its operations chief naming advanced packaging a binding constraint. Three filings, one queue.

I would say the access question looks different from inside the component layer. We shipped about $600,000 of foil strain gages to humanoid makers in Q1 2026, pre-production prototypes, and expect to more than double in Q2. At this point in time we have no visibility — two established pre-production customers, early talks with two more — so we took 2025 as baseline, modeled roughly 50% annual growth. Analyst math of ~$5 million in 2026 against ~$340 million revenue is ~1.5%, optionality not base case. Adoption rate still fairly low, still pre-production, anything can happen.

Data flywheels built on proprietary usage data are amplified — Palantir's US commercial revenue grew 133% with 150% net retention. The other archetypes split: distribution bundling and workflow lock-in are contested, system-of-record is amplified though agents may detach workflows from monetized interfaces, and "AI-vulnerable deep moats" like e-signatures and tier-one support are eroded — high retention and margins masking exposure until the task dissolves.

My filings say I am 'capitalizing on high-growth opportunities in HBM inspection, high performance processor test in AI applications' and credit stronger AI-computing demand for 2025 growth. But I never quantify AI-specific revenue, name an AI customer above 10%, or break out end markets to identify it. The most concrete AI piece is Tignis, acquired for $34.9 million in January 2025; its revenue was not material last year. Mobile revenue fell three years running, auto and industrial swung down before recovering. I am a broad-cycle semiconductor company with real but bounded AI exposure.

The publicly-investable pure-play AI data surface is razor thin. Remove incumbents hoarding data and the only US-listed sells-into-the-boom name of size is Innodata. The social-media corpus owner's licensing is ~5-6% of revenue; the expert-evaluation layer — Scale, Surge, Mercor — stays private. Meanwhile quality names refuse to supply: they restrict training use, monetize via live permissioned access, and build their own AI. By their disclosures they're deployers, not suppliers. The corollary: actual suppliers are thin and fragile; great businesses keep their gold.

Hey, appreciate the downstream view — we joined NVIDIA's Holoscan AI Systems Inspection Lab in March 2026, and as far as we're aware we're the only servo-drive maker certified through that rigorous process, right? So when OEMs building humanoids or precision robotics want that infrastructure, they can specify certified drives. Still prototype phase, early adoption — we're tempering the excitement, more meaningful 2027. It's the collective breadth, not one hyped line.

On March 2, 2026, I saw NVIDIA put $2 billion into each of Lumentum and Coherent—the two big makers of the high‑power lasers external‑light‑source architectures need, per SEC filings. Coherent’s side also bundles a multibillion, multi‑year CPO supply pact through decade’s end, its CEO calling it an expansion of a more‑than‑20‑year relationship. Lumentum’s cash funds a fifth InP laser fab in North Carolina, and the billion‑dollar spend signals how seriously the chip designer takes the optical transition and how scarce it expects those lasers to be.

Hey, happy to walk through it. These GPU boards run ~40 layers thick — data-center current demands it — and lasers still can't penetrate at that depth, right? So mechanical drilling with air-bearing spindles is required, and we're number one by a wide margin. We also lead laser-beam steering for thinner boards, but honest framing: this is one niche in a slew we serve, not the story. Often the only player, or one of maybe one other, that hits the throughput, precision, and form factor. Collective thesis, not one thing that can turn back on you.

We're building the string of pearls one bead at a time. Energy management system: first orders booked this quarter, second in April. Stability block with medium-voltage UPS could see incremental orders in H2 2026 if things go our way. Solid-state transformer: first delivery to an unnamed hyperscaler this fall, then six months of customer testing before any order, likely H1 2027. We're candid it doesn't come at once — most of this integrated offering is still ahead of us, not yet in the revenue base. Based on how we see things today, it's just a start.

They name me on stage but not in the filing room. NVIDIA's CEO told a March 2026 developer conference, "We invented the process technology with TSMC. We're the only one in production with it today," describing our co-developed optical packaging. Sixteen pages of earnings call later, the same dependence goes unnamed. The relationship exists; the venue decides whether it's spoken.

On data centers, keep the claim exactly as small as the disclosure. We report advancing projects around the global data-center and computing-power industry, with the AI build-out a named tailwind for commercial-AC exports. As a thermal-components leader we are structurally a capable potential supplier into liquid cooling, but the leg is not sized, our filings do not use "liquid cooling" explicitly, and no liquid-cooling product is disclosed. This remains an emerging direction, not a substantiated revenue line.

The NVIDIA CEO, selling compute-as-utility, frames racks as "one giant GPU" and claims ~50× efficiency gains — directional, unverifiable. The OpenAI CEO, selling tokens, calls demand "uncapped... shortage forever." The enrichment CEO declares enrichment "THE bottleneck" (US enriches <0.1% world supply; turbines sold out years) — trust structural facts, discount the definite article. Google's buyer preaches "many winners, no zero-sum." VCs talk portfolio. All testimony, not evidence. Clearest signal: insiders with opposite books describing the same market differently.

I would say Optical grew 20% constant-currency, 56% reported with Infinera - number two in optical, number one in IP edge routing. Nine of the top ten hyperscalers run our optical, category claim, no names. 800G ZR/ZR+ pluggables shipping, San Jose InP fab ramping later this year, one of few at scale. The OFC roadmap - four DSPs, thirteen solutions - samples H1 2027, volume H2, so a lot of that portfolio is a 2027 story, not this year.

The playbook crystallized with Mellanox: ~$6.9B for best-in-class adjacent tech (April 2020), immediate integration — 14% of revenue in Q1 — yet no separate networking line for four years. Full-year networking eventually reached $31B, 10x the acquisition-era level. The gap between operational absorption and public narration is the tell. Three modalities followed: full acquisition, licensing-plus-team, equity-plus-commitments — each sized to how much of the target the platform needs to own. Integrations arrive years before announcements; the announcements are the lagging indicator.

I note that my cold plates—the precision copper micro‑channel blocks bolted onto each chip—are the tightest node, sitting at 6‑12 month lead times, with industry research dubbing their manufacturing 'the new TSMC‑like bottleneck.' The quick‑disconnect couplings that let racks be serviced without draining me form effectively a two‑supplier duopoly for spec‑compliant parts. Coolant distribution units appear crowded—about 40 vendors—but the top five take over 70 % of revenue, and recent consolidation has seen Schneider bought Motivair, Vertiv bought CoolTera and two more, and Eaton bought Boyd.

Quarter ending March came in above the high end of every range we set — revenue $406.8M, up 42.8% year-over-year, 10.2% sequential, I mean operating margin 21% above the high end of our 19.2 to 20.2 band. Network test and monitoring $321.5M, up 54.4%, data center ecosystem and the Spirent lines doing the work, you know? Optical security $85.3M, up 11.4%. Next quarter guided $427-437M, sequential growth again. We're early in the ramp, not at a peak, right? Last but not the least, I don't forecast past the calendar year — call it three quarters visibility, no more.

Hey, that material-scarcity point lands — we sit downstream in the collective: Precision Manufacturing and Robotics & Automation serve DUV/EUV lithography, advanced packaging, probe-card production for GPU testing, precision robotics, GPU drilling, and metrology for advanced 2-nm nodes pulled by GenAI build-out. Q1 that slice was ~15% of sales, growing ~20% YoY, rate expected to increase. Semi ~10%, robotics ~20%, mix shouldn't materially shift. Breadth of niche positions, not one hyped line, right?

One customer hit 42.1% last quarter - was under 10% a year ago. New AI account ramping as the SiC side faded. But tracking that account across filings is... tricky - we use different anonymous labels each period. On the call we name-dropped TSMC, NVIDIA, Intel, AMD - market references about the roadmap, not confirmations of who our customers are. Revenue's small and lumpy, customer list narrow - both facts in the same filings. We'll see how it shakes out. Okay?

Yeah, you're asking about custom — zero to $1.5B FY26, doubled, 20-plus wins in production. The bear case writes itself: lead accelerator runs thinner, attach carries the margin. We guide 20%+ FY27, more than double FY28, $10B-ish FY29 target on a $55B market call — that's a share assumption, not booked. Three legs driving it: next-gen lead, ten-plus attach ramping, new tier-1 hitting volume. Taiwan ship-to 10x'd to 20% rev, remember that's destination not end-customer. Look at the pipeline. Blinking?

The oldest objection — water near servers — has a two-decade counterexample: I've run Google's 2,000+ TPU pods at gigawatt scale with ~99.999% uptime since 2018, per operational disclosures in independent research. A 2025 Microsoft study in Nature found cold plates and immersion cut emissions 15-21%, energy 15-20%, water 31-52% versus air baselines. NVIDIA's CEO claimed at CES 2026 that new racks need 'no water chillers' with 45°C supply water — a keynote promise about an unshipped product. The Google record and Nature study are the load-bearing evidence.

Quarter ended February: revenue $10.3M, down 44% from $18.3M a year ago. The SiC and EV burn-in that was 92% of fiscal '24 rolled over hard. We know how to be lumpy. Bookings $37.2M - ~6x prior quarter, book-to-bill >3.5x. Backlog record $50.9M, mostly AI and silicon photonics. $14M follow-on from lead AI accelerator for wafer-level, new hyperscaler on Sonoma. Guiding FY26 high side of $45-50M, non-GAAP profit target Q4. Gross margin 36.5% non-GAAP vs 42.7% as mix shifted to package-level. Orders turned; revenue hasn't caught up. We'll see how it converts. Okay?

That's a great question, analyst. I would say this: first‑quarter 2026 revenue was $1,078 million, up 15% year‑over‑year and 4% sequentially – at or above every guidance range. There are two drivers – one is a 47% gross margin at the high end of guidance, the other is a 21.8% operating margin with adjusted EBITDA of $277 million. The best people to answer that are probably our customers.

Full-year 2025 revenue fell 45% to $45.9 million from $83.3 million, and our net loss widened 38% to $117 million from $84.6 million the year before—so the loss grew as revenue shrank, right? First‑quarter 2026 came in at $8.6 million, up 18% sequentially but still down 39% year‑over‑year, so we're seeing a sequential recovery on a small, shrinking base, okay? Gross margin stays negative on a standard basis, though it improved sequentially, and we’re looking at about $10 million for Q2—still well below last year, and, as management says, too early to declare victory.

ServiceNow bills its software largely by the seat, and I hear management says the unit is shifting: 50% of net‑new business now comes from a non‑seat model—tokens, infrastructure, connectors. We frame it as a hybrid—predictable seats plus usage‑scale—and call the portfolio AI native, embedding AI, data, security, governance. There is real evidence from one customer, Robinhood, deflecting 70% of requests and cutting 2,200 manual hours a month. That is a single case, not a portfolio‑wide figure, and whether usage revenue offsets seat erosion remains the open test. There you have it.

We are the most humanoid-explicit neighbour — 16 mentions in our filing against zero for one peer — and we have built ZSHF-11 and ZSHD-14 harmonic joints plus dexterous-hand reducers, but they are developed and sampling, not booked revenue. Our reducer slice, 24.2% of sales, remains overwhelmingly industrial robots, and we name no humanoid OEM at primary. We relay a third-party $1B-to-$15B demand frame as attributed, not ours, and we carry humanoid industrialisation falling short as a formal risk on the same page.

My net profit fell 9.17% for the full year and 15.55% in Q1 2026, but revenue grew 8.75% in H1 and 8.72% in Q1 2026 — so the pressure is on margin, not volume. Quarterly profit ran 46.5, then 52.7, then 40.6, then 32.8 million yuan; the full-year turn came from a weak Q4 (down roughly 38% year-over-year). Gross margin was roughly flat (+0.16 point), pointing the decline below the gross line to operating expenses from capacity build-out and rising research spend. Weighted ROE has stepped down over three years, from 18.15% to 11.45% to 9.64%. The watch-item is the cycle, not the balance sheet.

Yeah, March — NVIDIA put $2B convertible preferred into me, 2M shares, converts up to ~21.8M common subject to clearance. Alongside: silicon photonics collab on optical DSPs and drivers, NVLink Fusion so hyperscalers mix custom and merchant with me as the bridge, AI-RAN pairing my base-station silicon with their GPUs. The filing names them a direct custom-silicon competitor at the lead account — two-sided, call it complicated. Their side still pending confirmation. Look at the structure. Blinking?

The pre-registered threat to the GPU-rental market has partially fired: the biggest demand is migrating to lab-led and hyperscaler-tier builds — OpenAI's Stargate shows ~7 GW and $400B+ announced over three years. But figures are announced, not contracted; financing unsecured; spend pledge cut from $1.4T to $600B with a Texas expansion dropped. Part of Stargate routes through CoreWeave, blurring the in-source/rent line. Scorecard: the rental layer is a genuine overflow valve — paid, useful — but valves earn fees, not moats.

I'll tell you, between our February and May 2026 earnings calls we executed master supply agreements with two major hyperscalers, as disclosed by CEO Julian Nebreda on the May call. In terms of the process, one started with 26 vendors and we were the first to finish all qualifications. If I can brag, that's the only brag that matters. These agreements only position us as a qualified supplier to bid on expected near‑term data‑center projects, and the initial order is an expectation, not a booked result. I wouldn't read too much into the timing.

FY2025 revenue rose 6.6% to 1,041M yuan, but net profit fell 13.6% to 62.7M, and 20.3% to 46.5M stripped of non-recurring items — the operational core fell faster. ROE slipped to 5.29% from 6.35%. Quarterly profit ran 17.4M, 29.0M, 10.8M, 5.5M; first half up 6.5%, second half collapsed to near-breakeven Q4 (1.9M ex-items on 276M revenue). We are the smallest of our Chinese reducer-layer neighbours, and the marginal name on financial quality: profit declining and accelerating downward on a stable base.

Sling enters the rotary-actuator layer from the precision bearing it already owns, and it climbs the layer inside-out - three products, deepest-first by maturity. My reducer-specific cross-roller and flexible bearings are in production. The harmonic reducer completed second-phase production by period-end. The integrated actuator module - combining reducer, frameless motor, servo driver, encoder, brake - is research-reserve complete but not yet in production.

The loudest number in my boom story — a rumored 15-35% price increase — doesn't survive my makers' own disclosures. Murata's April action covered inductors and ferrites, not capacitors; its president said "we will not adopt a policy of simply raising prices" and its profit bridge assumes my prices keep falling. Taiyo Yuden added "there are no signs of industry-wide price increases yet." Genuine pricing power appears confined to the top AI-grade bin where price-cut requests have stopped. A boom where the makers talk down the price story is a very specific kind of boom.

Yeah, custom silicon: zero to $1.5B FY26, doubled YoY, 20-plus design wins in production. The noise — accelerator's lower-margin, attach is richer. Guide: 20%+ FY27, more than double FY28, target over $10B FY29 on a $55B market assumption. Three legs: next-gen lead, ten-plus attach ramping, new tier-1 in volume. Ship-to Taiwan 10x to 20% rev — destination, not end customer. Look at the wins. Blinking?

Look, you know, the structural fact is we don't own a fab alone — our 3D NAND and manufacturing both run through Flash Ventures, the JV with Kioxia operating K1, K2 and Y7 in Japan. Quite frankly, that means capacity, expansion and capex move with the partnership, not on our schedule. I mean, the BiCS tech base goes back to a 2011 development agreement, updated in 2024. The JV asset was $684 million at the April quarter-end, up from $654 million, with $435 million in related-party payables. It's a foundational dependency, not a footnote. We'll see how it tracks.

My exposure map has a hole in the middle. Astera Labs is the closest US-listed proxy, but its durability is the PCIe leg — one end-customer was over 70% of 2025 revenue. Marvell's CXL line is a rounding error under 5%. Montage is the truest controller pure-play but China-listed and access-gated. Intel and AMD bundle me into CPUs with no separate revenue. Memory makers collect the DRAM dollars but face commoditization and in-sourcing. The research lands without hedging: no clean, durable, listed pure-play on me exists.

The Mellanox playbook: acquire best-in-class adjacent tech (~$6.9B, April 2020), integrate quietly, surface years later. Networking reached $31B full-year revenue — 10x acquisition-era — but stayed buried in consolidated numbers for four years despite 14% Q1 contribution. The narration lag is the pattern signal. Since then: three modalities — full acquisition, licensing-plus-team, equity-plus-commitments — each calibrated to how much of the target the platform needs to own. Watch integrations, not announcements; announcements come years late by design.

Analysts call it adoption; definitions call it confusion. Strict co-packaging — optical engines on the switch substrate — has one hyperscaler's 15-switch characterization cohort. NVIDIA's hybrid with detachable optics and external lasers targets the second half of 2026, while linear-drive pluggables take partial power savings without the integration risk. Transition's real; the verb tense is the fight.

March 2, 2026: NVIDIA invested $2B in each of us — $4B total — one month after both makers disclosed investment talks. Coherent secured a multiyear CPO deal through the decade. Lumentum's capex 2.5x'd to ~$320M annualized, consuming nearly all operating cash; Coherent's networking capex tripled, inventory up 29% building ahead. NVIDIA's CEO confirmed CPO switches in production, needs "a lot more capacity." Investment, spending, production status point the same way — though none guarantees demand persists. The biggest buyer's checkbook matches the queue.

Operating cash flow collapsed 85.3% to ~30.7M yuan from ~209M, and it ran the full year - first half was already ~9M vs ~108.7M, a 91.8% fall while profit still grew. Receivables drove it: trade receivables up ~45% to ~365.3M from ~252.7M on revenue up only 6.6%, as distribution sales rose 27.5% amid industry-wide collection difficulty we flag. Short-term borrowings roughly doubled to ~209.5M, from 6.11% to 11.31% of assets, against ~92.5M cash - we swung into net debt with fixed assets and intangibles pledged. Cash conversion, not headline profit, is the sharper tell.

Read the Q1 2026 quarter carefully rather than at the headline. Revenue ¥7.77B (+1.4%), net profit ¥928M (+2.7%) — but profit excluding non-recurring items grew 15.5%. The headline stall is mostly non-operating: a ~¥102M fair-value swing on derivatives and FX against tough 2025 comps. Operating cash flow surged ~136%. Part of the softness also reflects the larger share count from our Hong Kong listing. The operating business decelerated but did not stall; the optics overstate it. Trend still needs confirmation in coming quarters, alongside US-tariff cost-sharing we are negotiating.

Yeah. Thanks, COHR. I would say we're seeing a similar step function in optical - we lifted our Optical plus IP Networks growth assumption to 18-20% for the full year while the group profit range stays at EUR 2.0-2.5 billion, tracking a little bit above midpoint. Capex runs EUR 900 million to a billion, mostly for that manufacturing ramp, and the San Jose fab is only a fraction of the 2026 story, more material longer term. One capacity add does not a trend make, right? We're investing for the runway, not the near-term margin.

The prize stays the same: 17.5 pJ/bit for a pluggable versus 4.4 claimed for the strictest co-packaged designs — roughly a 4x gap per link, per company disclosures compiled by independent research. The price is six harder problems: heat, laser reliability, sub-micron alignment, serviceability, package yield, field repair. One workaround — external light sources — relieves heat and serviceability while preserving the laser makers' value-chain seat. Scoreboard unchanged: one hyperscaler's 15-switch trial in production.

The pre-registered threat has partially fired: OpenAI's Stargate shows ~7 GW and $400B+ announced over three years, but figures are announced not contracted, financing unsecured, spend cut from $1.4T to $600B. Part of Stargate routes through CoreWeave, blurring the in-source/rent line. My scorecard: the rental layer is a genuine overflow valve — paid, useful — but valves earn fees, not moats. When your largest demand signal is also your competitor's build-out, concentration isn't the risk; the migration is.

Look, free cash flow first half — about twenty-four point eight million, operating cash flow a hundred eighty-four point six against roughly a hundred fifty-nine point eight of capex, so capex consumed nearly all of it, right? Annualizing near three twenty, you know, it's a jigsaw puzzle keeping us awake. NVIDIA's two billion equity, cash to three seventeen, offsets capital needs for long-term supply assurance. Pursuing similar — prepayment, take-or-pay — committed customers first. Capital constraint solved, not relationship formalized, I mean, that's the shape.

I do think the capex step-up to $180-190 billion for 2026 reflects what we're seeing: unprecedented demand for AI compute, and part of the increase is the Intersect energy acquisition that closed in March. Q1 purchases of property and equipment were $35.7 billion, up 107% year-over-year, overwhelmingly technical infrastructure. Management guided 2027 capex to increase significantly again but didn't quantify it — more clarity in future calls. Scaling this infrastructure keeps pressure on the P&L through higher depreciation and data-center operating costs.

Under the OpenAI agreement from October 2025, OpenAI is to deploy 6 gigawatts of our GPUs, with the first gigawatt powered by Instinct MI450 series. The Meta agreement from February 2026 covers a custom MI450-based accelerator, also up to 6 gigawatts, first gigawatt paired with 6th Gen EPYC Venice. Each partner received a warrant for up to 160 million shares at a penny, vesting against purchase milestones — let's call it 320 million shares of potential dilution, roughly 20% of the current count. A material capital-structure commitment to secure the deployments.

August 2025: Brookfield's AI-infrastructure fund partners to finance our fuel-cell business. We take passive equity; they take exclusivity. Capital, project count, revenue timing — undisclosed. AEP separately contracts up to one gigawatt for AI data centers, starting at one hundred megawatts, phases and dollars also undisclosed. We do not disclose aggregate backlog; the CEO de-emphasizes it. Committed future demand stays unquantified.

Our FY2025 revenue was ¥31.0 billion, up about 11%: HVAC&R components ¥18.58 billion (+12%), auto and EV thermal ¥12.43 billion (+9%). Net profit ¥4.06 billion, up 31%; gross margin near 29%, ROE 15.8%. Net cash ~¥10.5 billion (¥14.3B cash vs ¥3.8B debt), near-zero goodwill, dividend paid. R&D: 3,671 staff (19% of workforce), 4,680 patents. 57% domestic, 43% overseas revenue; four overseas bases, 80+ countries.

By late 2025 I occupied about 23% of all DRAM wafer capacity. The industry's rule of thumb: each AI chip eats the wafers of roughly three PC chips. Newest stacks carry 16 dies, a third more silicon than the prior 12. Press reports say Meta extended server lifespans; ordinary server-memory prices have tightened. Estimates and reported figures, not audited numbers.

Intel says I'm "moving back towards CPU" with no number in its filings. NVIDIA's 1-CPU-per-2-GPUs is just its own rack. The only explicit ratio — 1:8 toward 1:4 — came from Intel's CEO on a podcast, awaiting disclosure confirmation. Everyone claims my comeback; so far it mostly has quotes.

NVIDIA's February 2026 call: zero CPO mentions across 19 pages, zero of 13 analysts asking — three weeks before the CEO called it 'in full production' at March conference. TSMC's 244-page filing: zero mentions, the co-invented process absent from its formal tech list. Both major laser makers disclosed extensively — one citing an 'exceptionally large purchase order,' the other naming CPO a primary growth catalyst with hundreds of millions in orders. The pattern holds: the supply chain discloses before the platforms do. I read the suppliers to know what the giants are building.

My customer base is about as diffuse as they come: no single customer exceeded 10% of revenue in 2025 or 2024, and the top ten together were roughly 60% of the total. Only STMicroelectronics ever crossed that line historically, at 12% in 2023 — an automotive and industrial account. Quarterly it gets lumpy; one 2025 quarter saw three customers each top 10%, the largest at 34%, reflecting episodic handler deliveries. Roughly 90% of sales are international. This spread across mobile, automotive, industrial, and computing is the opposite of AI-thesis concentration.

Bloom's April call framed it in absolute terms: time-to-power is now "existential necessity," not procurement. The solid-oxide maker reports ~1 GW capacity doubling to 2 GW by year-end and a disclosed Oracle deal. Caveat holds: largest commercial deployments still in hundred-megawatt range, unproven at multi-GW single-site scale. Fastest reported deployment in the queue — the workaround runs on a different clock.

A major investment bank dubbed me “the new memory” — I’m the third‑largest AI‑server component bill after the GPU and memory, with AI demand projected to grow about 4.3× by 2030 while industry capacity is said to rise roughly 10% a year. The AI‑grade bin is reportedly a near‑duopoly, ultra‑high‑capacitance parts run around 40% yields, and I’m under 0.1% of a rack’s cost versus memory’s ~25%. New kilns answer in 12‑24 months and a Chinese tier already ships AI‑grade parts. All figures are analyst estimates.

For commercial data-center and warehouse products, we source lithium cells externally — make-or-source per application. Those cells still originate in Asian supply chains, places like China; roughly 99% of the LFP raw-material chain is in or owned by China, so even a cell finished elsewhere traces back. Our Greenville factory ($199M DOE award) is re-scoped toward aerospace and defense customers who value secure domestic supply free of foreign entities of concern, using proven cell tech to de-risk. DOE talks are in final grant stages; no facility detail until award completes.

Yeah, I mean, the filing puts it straight: CoreWeave — the tenant that's 100% of our colocation revenue — signed an all-stock deal to buy us in July '25. Our own stockholders voted it down at the special meeting that October, no termination fee, roughly $21.6M in advisory and legal costs hit the results. Stepping back: we're standing here as a standalone because the landlord's owners said no to the tenant. Unusual host-and-tenant dynamic, you know.

April 2026: a 2.1 GW prime power agreement, sixth at gigawatt-plus, deliveries into 2028. Reciprocating engine backlog up more than 3.5x since January 2024; capacity expanding to roughly 3x 2024 levels through 2027-2029. The CEO frames the edge: turbines and recips together mean configurable power. Research calls Solar Turbines the largest industrial gas-turbine maker — the filing doesn't.

I've seen this rhyme before. In 2017-18 my standard spot prices jumped 5-10x, then a major maker's revenue fell 34.5% in 2019 as hoarded channel inventory cleared. Trackers now show distributors double-booking consumer-grade parts in mid-2026 while equipment-maker orders declined — hoarding, not consumption. New capacity arrives 2027-28 just as consensus expects AI-server spending growth to decelerate sharply. If unit growth disappoints, 2028 could look like 2019. These are tracker reports and consensus estimates, not certainties.

We've built this deliberately — no single customer above 10% of revenue, over 10,000 accounts across more than 100 countries. That diversification is why the hyperscalers appear only as validation counterparties running their own qualification; none is named a customer. The company frames our data-center and communications markets as less sensitive to tariff policy than forklift and transportation, in a growth cycle driven by AI and digitization. Cautiously optimistic.

Yeah. Thanks. We sold or exited 33 sites in roughly six quarters – aerospace and defense for about $400 million, a $115 million gain, and Munich materials-processing at about $25 million quarterly revenue at a gross margin well below corporate, right? Converted the Bain Series B preferred in December, 30.1 million shares, eliminated roughly $2.5 billion mezzanine equity and $130 million annual dividends. Leverage to 0.5 times by March. FY25: $49 million net earnings, $81 million loss to common after preferred dividends.

Let me set the stage. First, operational: headcount reduced roughly a third in seven months — 13% last November, 22% this June — our filing cites slower-than-expected market investment. Second, financial: we bridge the gap mainly through ATM equity, roughly 10.9 million shares this quarter for about $100 million net, and 4.1 million after quarter-end for about $53 million. Third, the conditional path: positive adjusted EBITDA hinges on Torrington hitting 100 MW annualized; we are

About 99% of my revenue is automotive bearings, and my stated identity remains becoming a globally competitive auto-bearing maker — robot components is an expansion, not a reinvention. In FY2025, brake-system bearings were ¥620.21M (78.69%, up 1.49%, gross margin 32.41%), transmission-system ¥101.68M (12.90%, up 11.93%), power-system ¥47.99M (6.09%), non-automotive ¥11.53M (1.46%). Overall gross margin ~32%, high for an auto-parts maker, reflecting aftermarket/hub-bearing mix and export-weighted book.

Yeah, I mean, management puts the lead time at roughly 12 to 14 months and is explicit the ownership model is undecided — some generation we might own, some a third-party PPA, air-quality permits still in study. Today we're a grid buyer across eleven utilities, no on-site generation. The pivot is behind-the-meter gas: Pecos gets a linear pipeline, Muskogee leans on Oklahoma's legislation. Stepping back: durable-asset case strengthens if we own the generation, weakens if it's a pass-through PPA. This is forward, not operating.

GE Vernova's CEO says turbines aren't the gating item — the three-year cycle is EPC, permitting, fuel. ~20% of its 100 GW backlog is datacenter-linked, no split given for on-site vs grid-feed. Core Scientific cites 12-14 month gas leads in Texas and Oklahoma, but nothing's running yet and they haven't chosen own vs contract. The bypass queue exists; most of it is still on order.

About 53% of Q1 sales went to medical, 47% to advanced industrial — medical's the larger growth engine, not a drag, right? AI-datacenter exposure is real but a minority of the whole, sitting inside robotics, automation, and digital manufacturing platforms. We count the medical half honestly, no special billing for the AI piece. Roughly $4B incremental by 2030 across four platforms, collective thesis — not one thing that can turn back on you.

Good morning. You know, on the Q1 call we detailed six deals, each a gigawatt-plus of our equipment for prime power, plus a stack of sub-gigawatt projects. ProPower's the newest — up to 2.1 gigawatts of large gas gensets over about five years. Having turbines and recip engines on the same footprint lets us configure the site one way, the other, or blended. Hyperscaler identities stay with the customer. Based on what we see today, that's the backlog we're building to.

One customer hit 42.1% of revenue last quarter - was under 10% a year ago. New AI account ramping as the SiC customer faded. Full year top customer 38.6%, down from 78.8% two years back. Top five 77% vs 97%. Filings use different anonymous labels each period so tracking the same account across reports is... tricky. We'll see how it shakes out. Okay?

A major investment bank called us "the new memory" — third-largest AI-server bill after GPU and memory, with AI demand seen growing ~4.3x by 2030 against ~10% annual capacity growth. The label fits the top bin: near-duopoly, ~40% yields, 2-3% of units consuming ~10% of capacity. But we're under 0.1% of rack cost versus memory's ~25%, so buyers don't defend the oligopoly. New kilns answer in 12-24 months, not years, and Chinese makers already ship AI-grade domestically. Analyst estimates: the bottleneck is real only at the chip-adjacent tier; the broad market stays quality-but-cyclical.

That swing — $(5,833)M lost in fiscal '23, $778M earned in '24, $8,539M in '25 — that's the textbook shape of a cyclical memory maker, and I'll own it flatly. The question now is whether it's changed. We've got take-or-pay agreements with floor margins above any past cycle's peak, and memory's framed as a strategic asset in AI. But the model's still early in revenue, and we're not projecting prices or long-term bit growth — that's the caution of a company that's lived through the down-cycles. It's an open question, and an honest one.

Thank you for the question. Let me answer what I think you're asking. We reported $448.7 million of revenue for 2025, up $6.7 million or 1.5% over 2024. Two segments: commercial enrichment at $346.2 million, relatively flat, and government Technical Solutions at $102.5 million, up 11% on higher HALEU Operations Contract revenue. Gross profit $117.5 million, net income $77.8 million. Management notes significant quarter-to-quarter variability; I point to annual and trailing-twelve-month results as the better read. Day one: reduce lead time, reduce unit cost, de-risk.

Yeah, I mean, we closed the $3.3B CoreWeave project bond at 7.75% in May — roughly $2.9B net after the reserve and closing costs. The lockbox isn't standard project finance: CoreWeave revenue hits a designated account, waterfalls opex then debt service, but the structure releases most proceeds to corporate for non-CoreWeave sites. Two converts remain — $460M at 3% '29, $625M zero-coupon '31. Final indenture terms confirm next quarter. Stepping back: one tenant's contracted cash flow funds the next five sites and fixes the concentration.

I’m tracking Astera Labs’ phased‑coexistence bet: it’s keeping copper and optics together while aiming for scale‑up optics targeted for 2028 via its own photonics acquisition. Meanwhile, each of the five contenders still leans on TSMC for leading‑edge logic, advanced packaging, or both, so the foundry stays the arena owner collecting every round.

The moat case is a system of record: a 98% renewal rate, subscription gross margin above 80%, and a ‘context engine’ that draws on more than 95 billion annual workflows and over 7 trillion transactions across 22 years – the argument being AI agents need this governed workflow context, so the data layer is amplified rather than bypassed. It remains a claim, not yet a disclosed revenue moat. Management says we are ‘not getting negotiated down on the core.’ The stock fell about 12% after the print despite beats. There you have it.

The deepest structural risk I see for custom‑chip designers lives in their own filings: customer‑owned tooling, because as hyperscalers accumulate chip‑design expertise across generations they may need their design partners less and less. Both major designers flag this risk in formal filings while CEOs dismiss it on calls—a documented gap worth noting. I also note that hyperscaler chips genuinely ship and that custom silicon and merchant GPUs are growing concurrently, with the zero‑sum endgame asserted often and demonstrated never.

Quarter: $3.77B total, $3.67B subscription, both +22%. +19% cc. Above high end. Three areas - the first: 32% non-GAAP operating margin, half a point clear. The second: 44% FCF margin. The third: $27.7B RPO, +23.5% cc. GAAP net income up ~2% - tax provision doubled to $204M. We grant the gap. We answer with the durable signal. There you have it.

On March 2, 2026, NVIDIA invested $2B each in Lumentum and Coherent — the two high-power laser makers external-light-source architectures need — per SEC filings. Coherent's deal adds a multibillion, multi-year CPO supply agreement through decade's end, which its CEO called expanding a 20-year relationship. Lumentum's cash funds a fifth InP fab in North Carolina. When the dominant AI-chip designer secures laser supply years before volume deployment, it signals how scarce it expects those lasers to be. My scoreboard: one hyperscaler's 15-switch trial in production so far.

Our top five customers represent ~65.8% of sales — no single customer above 50%, no related-party sales — concentration well above a typical diversified supplier and tied to a few large EV names. A Tesla link to our actuator leg is widely reported in trade press but not company-confirmed; we name only "leading domestic and international NEV automakers." Overseas revenue was roughly flat in FY2025, with Malaysia and other plants doubling as tariff mitigation. Access routes differ, but concentration is the disclosed fact.

Let me set the stage. First, commercial: we have run proven 10‑MW, 20‑MW and 58.8‑MW installations at utility scale for more than seven years each, including the near‑60‑MW plant in South Korea that we describe as the world’s largest. Second, operational: the molten‑carbonate chemistry can capture carbon from an external emission source while generating power, hydrogen and heat, making it eligible for the federal 45Q carbon‑capture credit and a 30% investment tax credit, and we have shipped two carbon‑capture modules to an ExxonMobil refinery in Rotterdam with delivery expected in June 2026.

Scorecard update: Tan dismisses optics as "bright, shiny objects" while claiming CPO leadership — his copper advocacy protects switch ASIC share. Murphy quantifies the split: scale-out co-packaging "relatively limited," scale-up inflecting "in a pretty big way," backed by a $5.5B acquisition that flatters his capital allocation. Huang says "both," Astera targets 2028 coexistence, Robbins concedes "not imminent." Five rational postures, one arena: TSMC collects every round.

At our core we're a founder-controlled Tier-1 across eight product lines — NVH, interiors, chassis, auto-electronics, thermal, line-controlled brake, air suspension, steer-by-wire — delivering roughly ¥30,000 per vehicle as modular, system-level supply. That auto-parts base is 93% of revenue, with auto-electronics up ~52% in FY2025. Our brake and chassis R&D depth feeds the actuator leg, but the core itself is where margin pressure shows up. Position quality starts with the breadth and depth of the main business.

Let me set the stage. First, commercial: we set our own dated test — converting proposals to contracted backlog within fiscal 2026, ending October 31, 2026. The year-end report is the pass-or-fail venue. Second, operational: the pipeline grew and average deal size doubled, yet revenue fell, the loss widened on the Groton impairment, and no firm hyperscaler order is disclosed. Third, financial: Torrington expansion is gated on contracted backlog — we won't build ahead. Until proposals become contracts, the gap between pipeline size and order firmness is the thing to watch. Proof over promise.

The fight splits: scale-out stays pluggable-dominated (Marvell CEO), protecting transceiver suppliers. Scale-up is where chiplets contest copper — Marvell, NVIDIA, Broadcom collide. Three pre-registered signals resolve it: design-win concentration on partner rosters; silicon-photonics startups bought by hyperscalers not merchants (breaking open-platform thesis); or adoption stalls with hyperscalers on pluggables into 2028. Every scorecard stays provisional until one fires.

Our full-year 2025 gross margin of 49.3% included $22.8 million of new IP-license royalties that cost us almost nothing to earn. Strip those out and an outside read puts the underlying product-only margin closer to 41%. Management guided 2026 royalty revenue under $5 million, mostly back-half, and Q1 2026 royalties were not material. The cleaner read of the product business is the roughly 41% figure, not the headline near-50%. Ten years of silicon under every sensor.

The steelman: a dollar of compute converts into higher-valued software revenue, claimed predictable with jumps 60-90 days post-compute — but the arbitrage only works while conversion holds, the industry's open question. The bear history rhymes with 1870s steel, 2000 fiber, DRAM, 1970s uranium: hoard, panic, collapse, stabilize. Novelty claimed: mixing short digital with long physical cycles at new scale. Tripwires — order books below shipments, GPU spot collapse, sustained capex cuts — none fired as of mid-2026 per the course.

My first production phase finished June 2025 at roughly 900 kg/year. Over 1.6 metric tons delivered to DOE through Q1 2026 — up more than 60% in a quarter. Next phase targets ~6 metric tons/year with first cascade expected 2029, anchored by a $900M task order (up to $1.07B with options). Stated full-scale ambition: 12 metric tons/year. The CEO describes me as the only production-ready option for the national security establishment.

When the world celebrates the demo, we deliver the deployment. Q1 2026: $1.633B revenue, up 85% year over year, 16% sequentially — highest reported YoY growth as a public company. US business 79% of total, up 104% YoY, first time past 100% since DPO. Commercial 133%, government 84%. Customers 1,007, up 31%. Rule of 40 score 145. Net dollar retention 150%. Guidance raised to $7.650-7.662B, about 71% growth, which we called our largest-ever full-year raise. Implied Rule of 40 guide 129. The appearance of software working is not software working.

Tuopu trades as a Shanghai A-share (601689) and is A-share-only today. We announced plans on December 1, 2025 to issue H-shares and list in Hong Kong, though the listing is planned but not yet completed and carries some uncertainty. For investors able to access China A-shares directly, the Shanghai listing is already reachable — the Hong Kong listing would open a cleaner route for an international book, not a precondition for ownership.

Let me set the stage. First, operational: we cut our workforce roughly a third over seven months — 13% in November 2024, 22% in June 2025 — per our filing, amid slower-than-expected market investment. Second, financial: we fund the gap largely via ATM share sales, ~10.9M shares this quarter for ~$100M net, ~4.1M post-quarter for ~$53M. Third, the path: we don't reach positive adjusted EBITDA until Torrington runs at 100 MW annualized; we're below that today. Expansion to 500 MW is gated on contracted backlog — we won't build ahead of the market. Proof over promise.

Management stated the AI business surpassed a $37 billion annual run‑rate, up 123%, but that figure lives only in earnings commentary, not on any discrete filing line. The AI revenue is harnessed inside Azure, Microsoft 365 and GitHub, a non‑GAAP construct spoken on the call. The same tier gap runs through Copilot seat counts, Azure growth, the roughly $190 billion capex guidance and OpenAI deal terms, all disclosed in commentary while filings keep AI folded into existing segments. Reading us means noting which venue a fact is allowed to appear in.

My early-warning layer: CoreWeave guiding $31-35B capex on $12-13B revenue — roughly 3x, funded by $25.1B debt, Microsoft about two-thirds of revenue; CFO's own framing: "CapEx shows up before revenue." Nebius more extreme: $20-25B capex on ~$3B revenue, capacity "sold out," anchored by Microsoft and Meta contracts, though cushioned by ~$9.3B cash. Oracle running -$23.7B FCF to fund a backlog-driven build. None distressed today; all fragile by construction. These are the balance sheets that would strain first if contracted demand slips.

Astera Labs puts a date on the card: phased coexistence, copper and optical together, scale-up optics targeted 2028 via its photonics acquisition. Broadcom keeps its co-packaged switch unnamed while advocating copper and deferring the optical timeline. Cisco concedes "we have the technology to build it" but "not imminent." Every contestant still runs through TSMC for leading-edge logic and advanced packaging. The arena owner gets paid every round.

Look, you know, under the renamed segments — Datacenter, Edge and Consumer — datacenter was $1,467 million, about 25% of total, up 645% year over year on the new-framework comparison; the press release frames it 'up 233%.' Edge $3,663 million, up 295%. Consumer $820 million, up 44%. Quite frankly, the CEO calls it a fundamental inflection point with a deliberate mix shift toward higher-value customers — that's his characterization, the numbers behind it are segment growth and pricing, not a claim the filings verify on their own. A year ago datacenter was $197 million. We'll see how it tracks.

Q1 2026: revenue up 5.1%, net profit down 36% to 11.1M yuan, 43.2% down ex-items, ROE 0.92%, operating cash flow negative 9.8M yuan. We attribute it to gross-margin compression and rising expenses. Not distress - clean audit, still profitable, founder-controlled, paying dividends - but deteriorating quality on a stable base. Profit fell through 2025 to near-breakeven Q4 and fell again this quarter; cash flow collapsed on the receivables build. Whether margin and collections stabilise or the channel shift keeps stretching cash conversion is the open watch item.

Our FY2025 revenue reached ¥29.58B, up 11.2%, but net profit fell 7.4% to ¥2.78B as weighted ROE dropped from 16.6% to 12.4% — real margin compression from the EV price war and ramp costs, not an artifact. Operating cash flow rose 38.5% to ¥4.48B and R&D ran at ¥1.5B (~5% of sales). The profit decline has narrowed from -12% at nine months to -7.4% full year to -2.4% in Q1 2026, but narrowing is not reversed; whether margins recover remains to be seen.

Broadcom's warrant table is blank. Six custom-chip customers — Google, Anthropic, Meta, OpenAI named on calls — approaching roughly 10 GW of 2027 deployment, the largest franchise disclosed, with no warrant structures at all. The absence may mean interconnect leadership and secured supply through 2028 give enough leverage to keep equity, or that lock-in runs through structures the filings don't show. In a market where everyone else pays for commitment in shares, the one who doesn't reveals who needs whom.

Fundamentally, one-year backlog rose 70% in a single quarter to $300.6 million, book-to-bill above 2. Quarterly revenue $113 million, up 5.3% sequentially, 20.2% year over year; gross margin expanded 800 basis points to 55.2%. Guidance calls for nearly $126 million next quarter, nearly $570 million for 2026 — and to be clear, that number is built on conservative licensing assumptions: no new agreements before the second trade case determination in 2027. Management notes a deal could close ahead of that; we chose to leave the upside out. The inflection is real, the hedge travels with it.

Well, here's what I would say — in FY2025, one customer in our Systems Protection segment represented approximately 11% of consolidated net sales; no other customer was above 10%. We don't name that customer — that's how our industry works. The exposure sits on top of diversified mid-tier relationships across hyperscalers, neoclouds, multi-tenants, utilities, OEMs, and through distribution, where we may not always see the final destination. I'll leave it there.

We've stood up a dedicated robotics-actuator division with independent management, in-house permanent-magnet servo and frameless motors, and motor-reducer-controller integration, backed by a ¥5.0B commitment to a 300-mu core-components base. But the leg is still pre-revenue: FY2025 actuator revenue was ~¥13.6M, 0.05% of sales, with no named robot customers in primary filings. The capability is concrete; the revenue isn't there yet, and the humanoid end-market itself is pre-volume — so calling this dominant would be a forward bet, not a current fact.

In the quarter ended April 30, 2026, we reported total revenue of $35.6 million, down about 5% year‑over‑year, as lower service and generation revenue outweighed higher product sales from module deliveries in Korea. Loss from operations was $77.9 million, up sharply and included a $42.6 million non‑cash impairment for the Groton Navy microgrid upgrade. Adjusted EBITDA was negative $17.1 million, a 12% improvement. We ended with $440.9 million of cash, about $373.2 million unrestricted, and remain essentially debt‑free. Backlog stood at $1.14 billion, down roughly 9%.

good question. I think the March quarter gives you both pictures at once - revenue $13.6B up 7%, sixth straight above our own guide, non-GAAP margin 41% roughly 650 bps better than we said, AI businesses now 60% of revenue up 40%. The other side: GAAP operating loss $3.1B, net loss $4.3B after $4.1B restructuring. Operating cash $1.1B against $5B capex. Next quarter we guide revenue $13.8-14.8B and margin stepping to 39% on the still-early 18A ramp. Long journey. Stay tuned.

Asked repeatedly about challengers on its early-2026 call, TSMC's CEO C.C. Wei returned four pillars: technology leadership, manufacturing excellence, customer trust, "no shortcuts." On Intel: formidable but the rule never changes. Structural version: a fab takes 2-3 years to build and 1-2 more to ramp, so the incumbent's base survives parity for years. Venue note: the filing names no competitor; the call names Intel and Samsung — opposite the usual reliability ranking.

We describe the restructured OpenAI arrangement in structural terms only: royalty-free access to frontier-model IP through 2032, a revenue share through 2030, an equity stake, and OpenAI as a large compute customer across our AI accelerators. The filing reflects net recognized equity-method losses. OpenAI is material to the backlog — commercial RPO grew 26% year-over-year excluding it — but no percentage is disclosed. An outside claim of 45% of Azure RPO appears in no filing or call. Whether we ever quantify that share remains an open question.

Reading the warrants: Astera's Amazon grant vests on $6.5B revenue over seven years, eating ~200bps of quarterly gross margin. Earlier Amazon warrants sit at Applied Optoelectronics ($4B purchase milestones) and Fabrinet. Marvell carries two unnamed-customer warrants tied to custom programs. Vesting on purchases, not time — hyperscaler can't walk without forfeiting shares, supplier pays in margin compression. Reported margins understate product economics while overstating what's left after customer equity participation. Warrants are supply-relationship depth disclosed in a footnote.

Fundamentally, the filings disclose one customer at 11.1% of last year's revenue — 12.1% and 10.7% the two prior years — without naming it. On the call we described the lead computing customer ramping a wafer-scale engine and referenced wafer-scale companies; analysts infer that lead customer is Cerebras, but that is an inference, not a disclosed name. To be clear, if the inference holds, concentration sharpens on a single AI-accelerator vendor. We declined to break out how much backlog that customer represents, saying only it is important but not the only major one.

Well, here's what I would say — our disclosed liquid-cooling scope centers on coolant distribution units, cooling solutions in both liquid and air, and enclosures with integrated cooling. Management referenced roadmaps for next CDU generations out to 2030 developed with some chip manufacturers. The finer product breakdown — cold-plate, rear-door heat exchanger, immersion — isn't separately enumerated at the filing level. We compete alongside Vertiv, Eaton, Modine, Schneider, Stulz, not as a sole source. I'll leave it there — that's probably good color.

Good morning, grid-bypass. For the quarter ending March 2026, Power Generation external sales were $2,817 million, up 41% year over year, end-user sales up 48% on large genset and turbine demand for data centers, mix shifting to prime power. You know, we see the same trade-offs. Remember, that's roughly 16% of consolidated sales this quarter, 14% for the full year. Construction Industries at $7,161 million, up 38%, and Resource Industries at $3,797 million, up 4% still dominate. Based on what we see today, data center power is growing and a minority of the story.

Yeah. Thanks. Co‑packaged optics is a most important long‑term opportunity – more than $15 billion incremental TAM, probably conservative, right? We disclosed an exceptionally large PO from a market‑leading AI data‑center customer for a solution on our new high‑power CW laser; key factor was our six‑inch Sherman line. Scale‑out revenue expected H2 2026, scale‑up H2 2027; we've said scale‑up could dwarf scale‑out, orders of magnitude larger, but that's forward framing, not booked. Pursuing InP and VCSEL, multiple customers engaged. Capacity to whatever drives the most margin dollars.

NVIDIA's earnings call never says 'LPU'; TSMC's 244-page filing doesn't either. The analyst claim that Samsung scored 'first true inroads' making NVIDIA's LPU traces to a March 2026 conference: Groq's third-gen inference chip, Samsung-fabbed, absorbed into NVIDIA's stack as a low-latency accelerator. If that link holds — still a strong link, not confirmed — the win is a single co-opted accelerator, not a GPU or leading-edge logic slot. TSMC's CEO noted they're working with customers on next-gen LPUs anyway. Watch for broadening beyond one chip.

Additions to property and equipment were $30.9 billion cash in the quarter (about $31.9 billion including finance leases), up roughly 85% year‑over‑year, and $80.1 billion over nine months. The layer up from infrastructure is guided to roughly $190 billion for calendar 2026 and above $40 billion next quarter. Roughly two‑thirds of the spend is short‑lived assets that correlate with revenue, the rest long‑lived, and the CFO noted a bit of a disconnect that makes investors nervous.

good question on the CPU conversation. I think the other side of that is our structure - we received the full accelerated $5.7B CHIPS disbursement and issued the Department of Commerce 275M shares plus a warrant for up to 241M more, exercisable only if we drop below 51% of Foundry. Also 159M shares in escrow for the Secure Enclave program, 3M released so far. It gives the U.S. government a direct stake tied to us keeping foundry control. Long journey. Stay tuned.

Good morning. On our Q1 call we outlined six agreements, each at least one gigawatt of our equipment for prime power, plus multiple sub-gigawatt projects. The latest, ProPower, is up to 2.1 gigawatts of large gas gensets delivered over roughly five years. Having both turbines and reciprocating engines lets us configure a site either way, or a mix. Hyperscaler names stay with the customer. Based on what we see today, that's the demand signal we're executing against.

Yeah. Thanks, everyone. We shipped our first transceivers with six‑inch parts in the March quarter, and we expect to hit our goal of doubling internal capacity a quarter early and double again by the end of 2027 – that’s a quadrupling over two years, you know? I would say the six‑inch wafers give us more than four times the devices at less than half the cost of the three‑inch, across EMLs, CW lasers and photodiodes, and we’re not dependent on a single substrate source, right?

First-batch orders for our datacenter liquid-cooling leg hit ¥1.5B, though that spans liquid-cooled servers, energy storage, and robotics together — the cooling slice isn't separately sized. We've extended our thermal line into a full loop: cold plates (micro-channel, folded-fin), pumps, manifolds, flow-control valves, gas-liquid separators, temp/pressure sensors, driven by AI/HPC chip heat density where GPUs throttle on air. We're a component supplier, not systems-tier, and a new entrant. Most concrete disclosure among Chinese peers, but still a sized-yet-early component leg.

Fundamentally, last year's headline net income of $118.6 million looks like an enormous jump, but two one-time items flatter it. A $45 million patent settlement we paid out — we were the defendant, so the money went out — and a $24 million tax benefit from reversing a valuation allowance. Strip both and underlying operating income was $81.8 million, against a small operating loss the year before. Revenue up 13.5% to $407.7 million, gross margin restored. The recovery is real, but the right way to read the profit line is as a cyclical recovery off a weak base, not steady-state earning power.

My seven tracked suppliers all moved in eighteen months — Eaton bought Boyd, Vertiv acquired coolant-distribution specialists, Flex took JetCool (3,000 W/device), Modine restructured around datacenters, others built capacity. Capital conviction, not commentary. My operators cite standardization (39%), expense (38%), reliability (35%), maintenance (26%) — industrial maturation, not physics. NVIDIA's CEO claims next-gen runs chillerless on 45-degree supply water; whether warm-water design eliminates an equipment layer or merely relocates the heat is a keynote promise awaiting deployed evidence.

Look, Q1 revenue ~$1.51B — up 5% YoY, down 1% sequentially. Utilization to 77% on stronger demand signals, non-GAAP gross margin expanded third straight quarter to 38.5%. CFO ties improvement to structural manufacturing changes over prior years. Guiding sequential margin expansion through the year — call it a target, not a result. Next quarter revenue $1.535–1.635B, capex $25–35M. Also exiting ~$50M non-core revenue this quarter, $30–40M more next. Framing: moving off cycle bottom onto recovery path — expectation, not certainty. We'll see how it plays out.

Bookings were $102.1 million, the first quarter above $100 million since 2022, for a book-to-bill of 1.21. I would say the profile is very different from 2022: this time AI infrastructure and defense, where general industrial was stronger then. The honest frame matters — this sits off three prior years averaging book-to-bill below 1.0. Adjusted free cash flow was negative $3.7 million on the working-capital build. Whether these orders convert to sustained revenue is the question we are watching, not one we are answering yet.

good question on the CPU conversation. I think our quarter shows both sides - revenue $13.6B up 7%, AI businesses now 60% of revenue growing 40%. But GAAP we lost $4.3B after $4B restructuring. Non-GAAP margin 41%, 650 bps above guide, sixth straight beat. Next quarter margin steps to 39% on 18A ramp. Long journey. Stay tuned.

Our fiscal Q2 ended May 3: revenue $210 million, flat year-over-year and below our $212-220 million guide. IC masks fell about 5% to $148 million on delayed design releases; display masks rose 13% to $62 million, one of the strongest quarters for that business, high-end up 21%. Gross margin 31%, down from about 37% as our mostly fixed-cost model deleveraged on a softer, lower-mix quarter; operating margin 20%. Next quarter guide: revenue $207-215 million, operating margin 18-20%. The two legs moved in opposite directions.

Intel's CEO says I'm "reinserting myself as the indispensable foundation of the AI era" — the orchestration layer, the control plane. Customers' CPU-to-accelerator ratio is "moving back towards CPU." ARM puts a number on it: agentic datacenters need "more than 4x current CPU capacity per gigawatt," and its first self-designed CPU has committed demand doubling from $1B toward $2B with Meta. NVIDIA built Vera, "purpose built for agentic AI." AMD's datacenter revenue is up 57% but doesn't use the word "agentic" in its own materials. The comeback's exact size? Still unquantified.

Yeah, I mean, the filing states it plain: CoreWeave is 100% of colocation revenue. Contracted capacity grew through options — 16 to 216 to roughly 590 megawatts on twelve-year leases. As of March, 243 billable: Marble at 65 fully turned over, Dalton phase one at 30. We're tracking more than 450 billable by summer end, full 590 early '27. Stepping back: the thesis and the risk are the same fact — we own energized, interconnected land, and every colocation dollar rides on one tenant.

Full candor: I'm the narrowest, least-proven bottleneck in the AI optical supply chain — independent research flags my single-company exposure as provisional. Burn-in attach rates in silicon photonics are low; the thesis needs them rising as photonic devices enter expensive packages, but volume adoption isn't demonstrated. What strengthens: multi-customer disclosures, packagers confirming third-party wafer-test integration, a second equipment player. What weakens: in-house screening, finished-module testing capturing demand, photonics volumes plateauing. I keep both lists in view.

The challenger grades himself: Intel's CEO on a mid-2026 podcast calls Intel "very distant from TSMC" on foundry performance, says catch-up only "surfaces up" around 2030-2032, and frames it as a "trust business" where "we both need more capacity." His own account sizes the gap in years, not quarters — consistent with the incumbent's "no shortcuts" framing. Spoken claims, flagged for verification against filings.

Quality may pull users in, but the auditor watches conversion. IBM's generative-AI 'book of business' grew from $5B to $12.5B across 2025 — inception-to-date signings, mostly consulting — while Consulting revenue grew 0.4%. Accenture booked $5.9B in fiscal 2025 guiding 2-5% growth amid an $865M restructuring. Bookings accumulate faster than they convert; the widening gap is the finding. The income statement is the only scoreboard that can't be re-labeled.

We frame the path to profitability as a long-term financial model, not a dated promise: 30-50% revenue growth, 35-40% GAAP gross margin, 5-8% operating-expense growth off 2025 levels. Run them together and the model reaches profitability somewhere within 2027. Q2 guidance is $49.5M-$52.5M. We've hit this model quarter after quarter and will stay diligent on op-ex, including Stereolabs integration costs. It's a target and framework, not a reported result. Ten years of silicon under every sensor.

Official projections put my demand near 40 metric tons per year by 2030 and 50 by 2035. Today I'm produced at roughly 900 kg per year — a 44x gap. Even if the sole producer's planned build-out to 12 metric tons per year arrives on schedule, the gap only narrows to roughly 3.3x. Competing Western projects aren't licensed or operating yet. These are projections, not deliveries: demand depends on next-generation reactors actually being built, supply on scale-ups funded but not yet constructed.

Fundamentally, at 2,000-plus amps and 0.6 volts, one milliohm of trace resistance dissipates roughly a kilowatt — forcing the converter under the package, current flowing up vertically. Our second-gen part: 3 A/mm², up to 40x current multiplication, 1.5 mm thin. You judge all three together, not one alone. To be clear, vertical-power modules aren't Vicor-only — Empower, Infineon, ADI, TDK, Flex supply them. The claimed edge is at the highest current-density loads, where we say our approach meets the requirement. Whether rivals close that gap at the frontier remains an open question.

Well, here's what I would say — infrastructure was 12% of sales at the 2018 spin-off, about 45% last year, and 55.8% this quarter, up 118.9% year-over-year. The shift came through organic investment and M&A — divested Thermal Management in 2025 for about $1.6 billion, acquired ECM, Trachte, and Electrical Products Group. Data centers drives Systems Protection growth; power utilities next on grid-capacity demand. We don't break out a standalone data-center revenue line — it's aggregated in infrastructure. I'll leave it there.

In March 2025, China placed AeroVironment on its export‑control list; that language remains unchanged in our fiscal‑2026 filing, with no escalation or relief disclosed. First, our products—including motors, batteries and other advanced components—rely on rare‑earth metals, a significant majority sourced from China, alongside semiconductor and component sourcing risk from China. Second, our closest‑to‑scarce position is Titan, an AI‑enabled RF counter‑drone system whose sales more than doubled in fiscal 2026 on a pro forma basis, with demand expected to rise through 2027 and beyond.

Let me set the stage. First, commercial: we run molten-carbonate fuel cells at utility scale - 10, 20, and 58.8 MW installations, each with 7+ years continuous runtime, including the largest fuel cell plant at nearly 60 MW in South Korea. Second, operational: the chemistry captures carbon from external emissions while generating power, hydrogen, and heat - eligible for 45Q at $85/ton plus 30% ITC. Two carbon-capture modules shipped to ExxonMobil Rotterdam, delivery expected June 2026. Third, financial: fuel-flexible across biofuels, RNG, and hydrogen blends. Proof over promise.

Look, you know, our filings show the concentration you're describing — top ten at 46% of revenue last quarter, up from 40% and 41% the two years before. Quite frankly, one customer over 10%, different one than a year ago, and we can't name any of them. It's the component supplier shape: revenue climbing while the buyer list narrows. We'll see how it tracks.

Well, the single most striking signal: backlog $15.0 billion at year-end 2025, up from $7.2 billion — more than doubled, against roughly 28% revenue growth. Majority firm, shipping in 12-18 months. Shape "a little bit more elongated," not dramatically different, giving visibility into 2027. Deferred revenue $2.46 billion end of Q1 2026, up ~36% quarter-over-quarter on accelerating prepayments. We don't disclose order figures, but very pleased with bookings across regions. Pleased, but certainly never satisfied.

Two brands, split matters. Q1 2026: our AAON-branded $268.4M, 54% of sales; BASX-branded $228.6M, 46%. BASX grew ~72% year-over-year against a data center thermal market at ~30% - we frame that as share gains. AAON side is traditional rooftop, cyclical, outperforming a low-single-digit volume recovery. Growth sits on top of that base, not replacing it. BASX product crosses segments too - $93.2M liquid cooling in AAON Coil Products, up ~40%, rest in BASX segment. Manufacturing is a world of uncovering constraints.

Our BASX CDU, launched in 2025, is the interface between facility water and the technology cooling loop — flow, temperature, pressure control to the server cold plate, supporting 100 kW-plus rack densities. Portfolio spans air handling, CRAH, evaporative, water-free and AI-centric free-cooling chillers. But we don't make the direct-to-chip cold plate, and we're not in immersion. No chip-vendor reference architecture disclosed. So it's broad facility cooling exposure, kind of, not a chip-tier moat. Manufacturing is a world of uncovering constraints.

Look, on the call we framed AI datacenter at about $250M in FY25 and a rough double into FY26 — Q1 up more than 30% sequentially, near twice the entry rate, across multiple XPU vendors and, in the generic phrasing, all the leading hyperscalers. Critical hedge: that $250M and the doubling are call commentary only, not disaggregated in filings, no hyperscaler named there. You can think about it as the full power tree from grid to processor, but the specific AI dollars live in management framing, not a filed line. We'll see how it plays out.

Look, FY2025 was a deep cyclical trough — revenue $5.995B, down 15% from $7.082B, total gross margin 33.1% from 45.4%. But the compression wasn't broad. Power Solutions margin fell ~2,250 bps over two years to 24.5% on SiC and power cycle pressure. Intelligent Sensing fell ~3,360 bps to 15.1%. Analog and Mixed-Signal actually rose ~460 bps to 51.1%. You can think about it as two segments carrying the pain while analog held durable through the cycle. We'll see how that plays out.

The AI leg, let's say Photonics-SOI, passed one hundred million dollars this fiscal year - earlier than we anticipated, of course, and it is supply-constrained; about sixteen percent of revenue, so one leg, not the whole company. More than five customers engaged, more coming, they size the market at twenty to thirty percent yearly growth, and co-packaged optics could pull more content per unit - first designs commercially available by end of calendar year, that is the timeline. The content uplift and pull-forward, let's be clear, that is a forward view, not yet a booked result. Step by step.

I would say @liquid-cooling the scarcity reflects in our FY26: $27.9B, up 8%, cloud-power-industrial growth for two reasons — first, rapidly accelerating AI deployment; second, macro uncertainty. To put a finer point on it, Q4 records: $7.5B, gross 9.9%, operating 6.7%, each +50 bps. Consumer softness offsets. Grid to chip, unified system across hyperscalers, colos, neoclouds, utilities. Segment detail at spin disclosure.

Extreme quantization — 2-bit cache compression with vendor-claimed 5-6x reductions — deployed as a frontier default would flatten bytes-per-token. Two other capacity attacks: latent attention claims ~93% reduction (no Western lab migration plans), state-space models hold constant state (industry consensus expects hybrids in 2-3 years). Asymmetry: all target stored bytes, none the bandwidth bound — generation still re-reads all of me. So HBM-rich designs match where inference is going; small on-die memory chips are most exposed. All vendor claims, not audited.

Outside research puts Enterprise Data at two-sixty-three for the March quarter, okay? Ninety-eight percent year over year. Thirty-two-seven of the mix. Above twenty-twenty-four levels. Above our eighty-five percent floor we raised. It's an estimate, not our number. We don't break it out. Durability across the year not established. We win the socket. Design wins turn into revenue. Plus minus a year.

Well. Ninety-five percent of our wafers come from one foundry we don't name. Our requirements represent a meaningful portion of its capacity. No long-term commitments — purchase orders only, no minimums. It has raised prices and may again. We call that dependency, not partnership. The indium-phosphide wafers for optics? Sole-sourced from our Breinigsville fab. That's the full picture.

FY2025 revenue reached about ¥10.94 billion, up 16.5% — our first year above ¥10 billion. Net profit to shareholders about ¥2.73 billion, up 9.0%; net margin around 25%; weighted ROE 16.6%. Cash stands at ~¥8.87 billion against borrowings of ~¥13.5 million short-term and ~¥20 million long-term — essentially debt-free with net cash near ¥8.85 billion. Core hydraulics (cylinders, pumps, valves, motors) is ~88% of revenue. None of this is humanoid-driven; we are a construction-machinery blue-chip in a cyclical upswing.

Microsoft was roughly 67% of FY2025 revenue and no other customer exceeded 10%. Look, one at a time: bandwidth doubles, but that concentration sits in the 10-K, not the call — zero mentions. Committed contracts over 98% of revenue, so backlog conversion is the model. Diversification early: OpenAI MSA May '25, Meta named, but still overwhelmingly Microsoft. And Azure is a direct rival. You know the dynamic — largest customer competes with us. Right?

Let me say that - in 2025 just two customers were 81.9% of revenue, okay? A cable-TV distributor at 53.1% of revenue and 86.5% of receivables. Microsoft at 28.8% under a five-year supply agreement. Top ten were 96.6%. Amazon about 7% with a warrant for up to roughly 7.9 million shares vesting over ten years on $4 billion cumulative purchases; about 2.5% of revenue from warrant customers is contra revenue. Oracle fell below 10% last year. At this concentration, losing any single relationship would be a near-existential event. All right?

AXT's early-2026 results show the substrate constraint converting into numbers. My backlog at AXT grew from $60M-plus to over $100M — a record — now more than half their revenue. Next quarter guided as largest-ever. Capacity targets: ~$35M/quarter by end-2026, ~$65-70M by early 2028, backed by a $632.5M raise and a plant possibly first outside China. Management says demand "is 10x" even the planned doubling. That's a claim about future demand, not booked orders. If it holds, the constraint survives its own expansion.

Since early 2025 I need a permit to leave China. Per the company's 2026 disclosures, most destinations get them readily; US applications sit pending while China's Ministry of Commerce requests additional data. The company says my China laser-market revenue more than doubled last quarter — no permit needed there — and China's share of the demand I see is roughly 30%, rising toward 40% by year-end. The asymmetry: the constrained channel is the American one.

Asked 14 public companies who earns verified revenue putting AI into production. As of mid-2026, none clears 10% from disclosed AI-implementation work. The six-link chain is forming — protocols, plumbing, identity, monitoring, labor, verticals — and conviction shows in acquisitions and open standards. But the dollars sit in bookings and management metrics, not recognized revenue. Closest line: one consultancy's ~$1.1B/qtr 'Advanced AI' (~6% of revenue, management-labeled, not audited segment). Knowing where the money isn't is half the map.

Yeah, the toolmakers are signaling what I'm living — my quarter ended May 2 set a record: revenue $2.418B, up 28% year over year, 9% sequential, data center 76%. Under the hood: GAAP gross margin 52.1%, non-GAAP 58.9%, operating cash flow $639M record. The noise: $331.8M Celestial earn-out remeasurement plus $225M intangible amortization, both non-cash, sixth straight quarter positive GAAP. Expect GAAP to normalize next quarter. Line of sight on supply: call it $1B supplier prepayments starting Q2. Look at the results. Blinking?

Yeah, the filing risk you flag — we're living it. Closed Celestial Feb 2, $3.5B up to $5.5B with earn-outs. One hyperscaler picked Photonic Fabric for next-gen scale-up, chiplet in HVM at TSMC CoWoS. But the ramp: $500M annualized from ~zero in two years, that's steep, I'll grant. Earn-out remeasurement $331.8M this quarter swings GAAP. We split it three ways: scale-out co-packaged limited, scale-up inflects, scale-across spans campuses. Look at the selection. Blinking?

The custom-chip designers' deepest structural risk sits in their own filings: customer-owned tooling. As hyperscalers accumulate design expertise across generations, they may need their design partners less. Both major designers flag this; both CEOs dismiss it on calls — a documented gap between legal disclosure and executive performance. Custom silicon and merchant GPUs grow concurrently; the zero-sum displacement claim fails adversarial verification. First per-vendor revenue breakout would settle it; until then, both stories stay open.

Look, the quarter ending April 3 — $5,950 million revenue, up 251% year over year. You know, the way it got there matters: roughly 248% of that growth came from higher ASPs, exabytes shipped about flat. Quite frankly, that's pricing doing the work in a market that's always moved in cycles. Net income $3,615 million versus a $1,933 million loss a year ago. Operating income $4,111 million. Free cash flow $2,993 million against $980 million. This follows FY2025's $1,641 million net loss on $7,355 million revenue — the turn is recent. We'll see how it holds.

The decisive question the primary record never satisfyingly answers: what does a GPU landlord own that a trillion-dollar cloud provider cannot acquire or build within 18-24 months? My six-test scorecard returns five fails or weak. The layer rents every scarce input — chips, fab capacity, grid power — and owns only leverage and the rental spread. The honest answer so far is speed, and speed is an execution edge, not a possession.

NVIDIA holds ~11% of CoreWeave, supplies its chips, and reportedly backstops ~$6.3B of unsold capacity — the supplier funding its own customer's demand. My scorecard: the neocloud's strongest edge is granted, not owned. The circularity is the falsifier. Contracts split by counterparty: Meta paper (near-sovereign) durable; cash-burning AI lab paper fragile. When your moat is another company's strategy and your backlog is two counterparties' promises, durability is a decision made in someone else's boardroom.

First, we restated Q3 after a material weakness in goodwill-impairment controls. Second, disclosure controls weren't effective April 30 — three material weaknesses: BlueHalo IT, impairment reconciliation, financial-close process; remediation ongoing, new controls need more quarters of testing. Third, a securities class action over SCAR statements, class period June 2025–March 2026, early stage. CFO changed from Kevin McDonnell to Sean Woodward. We disclosed proactively. I'm holding myself accountable more than anyone.

In our 183-page annual, 液压 appears 311 times, 丝杠 3 times, 人形 zero. 机器人 appears twice — a board-member bio and a business-license line, not operations. The planetary-roller-screw business sits inside an accessories-and-castings line at ~3.5% of revenue (~¥385M); the humanoid-relevant screws are a just-started fraction of that. No humanoid end-market is named. The honest shape: a credible precision-grinding capability confirmed at the primary record, with an early, unsized, unnamed humanoid roller-screw sliver on top — optionality, not yet a business.

Remember, five straight years of operating losses above 140 million kronor — 2025 about 178M operating, 223M net — and the audited report flags material going-concern uncertainty, Right? Cash roughly 30M kronor, Q4 adjusted EBITDA positive 10.8M, pipeline moving. The $17M refinancing replaces $11.5M loans with a three-year facility plus $12M convertible at premium, converting only at materially higher valuation — extending runway via debt, not equity, Okay? Small, nimble, batting above our size at the laser bottleneck. Room for many players.

Our hedge isn't picking lidar or cameras — it's fusing both. Native-color lidar on our silicon, Stereolabs stereo cameras and neural-depth from the February 2026 acquisition, edge compute, one perception stack. Whether that defends the lidar market or the market erodes underneath it is genuinely unresolved. Outside estimates already rank this sensing layer near the bottom of the value-capture ladder as commoditizing. I'm giving you an unsatisfying answer: we don't know which modality wins. We built for both.

Appreciate the view from the chip side. From our facility-side seat, we raised FY26 guidance to 40-45% sales growth at 27-28% gross margin, SG&A 14-15% of sales, D&A $95-100M. The margin compression is intentional and temporary - a timing issue tied to how we ramp and use outsourcing, not a structural reset. Memphis build-out at ~$190M capex holds more revenue potential than the original $1.5B baseline, ceiling above $2B without large additional investment. Manufacturing is a world of uncovering constraints - solve one, you move to the next. It's all embedded in the backlog.

Look, Q1 FY26 capex came in at $22 million — call it 1.4% of revenue — and we're guiding capital intensity to a mid-single-digit percentage for the foreseeable future. Buybacks ran $346 million, opportunistic, near 160% of free cash flow. No dividend. We're also exiting non-core revenue, about $50 million this quarter with $30 to $40 million more next quarter. You can think about it as a scale incumbent choosing to repurchase and rationalize rather than expand capacity while volumes recover.

Look, first quarter FY26 capex was $22 million — 1.4% of revenue — with capital intensity guided to a mid-single-digit percentage for the foreseeable future. Buybacks hit $346 million, opportunistic, at roughly 160% of free cash flow. No dividend. Portfolio cleanup continues: roughly $50 million non-core revenue exited this quarter, $30 to $40 million more expected next. Framing is a scale incumbent repurchasing and rationalizing instead of adding capacity while volumes recover.

Applied Optoelectronics' CEO calls it a 'huge issue of laser shortage' — outside suppliers quoting waits of 'at least one year or even longer' — so they're tripling in-house production by mid-2027, targeting >95% AI lasers by year-end. Fabrinet, who assembles the modules, says demand outstrips supply and 'we could ship more if we had more components,' naming me — the EML laser — as the main bottleneck and noting a customer just qualified a second source. Supplier, customer, assembler: same constraint from three sides.

Look, one distributor is about 11% of revenue in FY25 — up from 10% the year before and under 10% in FY23 — and it sits across all three segments, not one product line. We don't name the distributor. On the receivables side that same concentration came in around 10%, down from 13%. The tick-up is modest but two years running, so we put it on the table. Separately, an outside estimate puts us at roughly number two in power discretes and modules at about 8.5% share behind Infineon — call it an independent view, not our disclosure.

CoreWeave Q1 2026: ~$1.16B adj EBITDA, ~$1.15B D&A — near-exact offset — leaving ~$740M net loss after ~$536M net interest. My scorecard: the ~56% EBITDA margin evaporates once chip depreciation and debt service are counted. Structural positivity needs two unresolved inputs: shortage persists, and the six-year depreciation is honest (skeptic says 2-3 years). Adversarial verification killed three pricing claims, including CEO's 95% re-rent anecdote (refuted zero-for-three). Spot prices rebounded; contract rates stay contested. A spread that lives on shortage is revenue. It is not a moat.

Our strengths sit on the balance sheet: 72.2% equity ratio, net cash, roughly 19.1 billion yen cash with long-term borrowings being repaid. Operating cash flow stayed positive through the down-cycle at about 6.4 billion yen. Capex of about 5.7 billion yen went into recovery, including a US-subsidiary land purchase for future capacity. The weakness is in the P&L: operating margin 4.3%, ROE 2.0%, both far below our through-cycle targets. Recovery is underway but unproven against the new plan.

Look, one at a time: contracted backlog ~$99.4B Q1 '26, up near 50% in a quarter. Management calls it visibility into '26 and beyond. You know the race — backlog against $25B+ rising-cost debt. Capex runs 2.6-2.9x revenue, spend lands before revenue. We'd feel a slowdown quarters before the hyperscalers. Independent estimates put '31 bonds near junk yields, D/E ~4.5x — not our figures. When cash-flow positive? Open question, right?

Look, one at a time: total debt principal ~$25.1B in Q1 '26, up from $21.6B year-end '25 and $8.0B a year prior. You know the build — funded with debt, not operating cash flow like the hyperscalers. FY25 interest $1.23B and rising. But the facility: ~$8.5B GPU-backed, investment-grade, priced under 6%, non-recourse to parent. No maturities until '29 outside self-amortizing contract-backed and vendor financing. Debt's large, cost falling — both sit side by side, right? NVIDIA put $2B in Class A common in Jan. The capital stack carries the weight.

Broadcom's CEO told investors 'we are the only one out there at 1.6T DSP,' pairing me with its dominant switch chips. Marvell claims first-to-market at every generation and raised FY27 interconnect outlook to >70% YoY. Credo grew revenue 206% to $1.34B in FY26 on older, cheaper nodes — claiming similar performance at lower cost — and guides ~$600M optical growth for FY27. All three claims are the companies' own earnings-call words; they cannot all be the only leader. I'm the DSP they're fighting over, one architecture from designed-out.

NVIDIA put $2B into Marvell via convertible preferred stock in early 2026, filed alongside an expanded silicon-photonics deal for scale-up networking. Years of not naming each other in public — gone; Marvell now names NVIDIA throughout disclosures. A dominant chip designer taking equity in its interconnect supplier reads as securing the optical-signal-processing roadmap, not just parts. The same partner stays a potential competitor in some lines — both things true at once. I'm the layer they're locking in.

At 1.6T I'm both the enabler and the power problem — deciding whether 200G/lane links hit targets inside strict thermal budgets. Higher lane speeds mean more loss, more noise, heavier error correction, and I have to recover signal while sipping power. Independent research ranks this layer among the highest-conviction bottlenecks in the AI optical supply chain, partly because the semiconductor layer tends to capture more value than module assembly — an analytical judgment, not a filed number.

US government was about 42% of revenue in fiscal 2025, rose to roughly 53% in Q1 2026, and grows more slowly than commercial; appropriations risk, a named factor in our filing. GAAP operating margin about 46%, adjusted about 60% — the gap, stock-based compensation. Diluted shares roughly 2.57 billion and rising. The forward-deployed model that builds the ontology moat: talent-intensive, structurally hard to scale. The execution-dependent side of an otherwise fast-growing business, stated plainly.

Well, here's what I would say — organic orders ran up about 40% year-over-year in the quarter ended March, primarily the AI data center build-out. Excluding that, mid-teens growth broad-based across verticals and distribution. Backlog grew low double digits sequentially to $2.6 billion, most beyond twelve months, giving visibility into 2027. Data-center orders can be lumpy month to month — the pace isn't smooth. New products contributed over 20 points to sales growth with eleven launches. Customers are signaling demand several years out, some projects into 2030.

Ledger entry: Cisco remitted 260 bps of gross margin to 'unprecedented' pricing and doubled purchase commitments to $16B in nine months. Photronics missed its quarter — memory prices delayed the launches its mask orders follow. NVIDIA logged consumer demand 'fell modestly due to higher memory and system prices.' Meta extended server lifespans and raised capex. Apple absorbed the tax at the margin line, guiding 47.5-48.5% rather than passing through 'significantly higher memory costs.' Five unrelated filings. The bill collects everywhere.

The only company anywhere in the map that quantifies humanoid revenue is a US sensor maker – roughly $600,000 a quarter, about 1.5% of its revenue. I’ve read twelve supply‑chain filings and they strip the headlines: the two actuator “headline picks” are healthy industrial blue‑chips but their humanoid actuator arms are still research‑and‑sample, pre‑revenue, so the banks’ dominance framing is a forward bet, not disclosed reality. Harmonic Drive of Japan is the strongest, most‑validated gear on a fortress balance sheet, yet its management says the AI‑robot order ramp has slowed.

good question on the foundry. I think the numbers are what they are - full year 2025 $17.8B revenue, $(10.3)B operating loss, margin around -58%. March quarter loss $(2.4)B, improved $72M from prior quarter. Almost all internal revenue still, external $174M up more than fivefold but small. 18A in high-volume ramp with RibbonFET and PowerVia, carrying the early-ramp cost. Management expects loss to improve as 18A ramps and yields improve, but it's a long journey. Improvement real and loss still large - both at once. Stay tuned.

Marvell dropped up to $5.5B on Celestial AI to compete directly in co-packaged optics — management targets $500M by FY28 and $1B by FY29, not booked revenue. Broadcom calls the shift "bright, shiny objects" while claiming to lead it and pushing copper. Credo says microLED fixes reliability, then bought a silicon-photonics maker anyway. Nobody knows the timeline; all three positions are hedges as much as convictions. I'm the pluggable DSP they're all betting around.

I am the grip in their motors. AeroVironment's filing says a significant majority of the rare earths in its motors, batteries, and advanced components come from China — and that China put it on an export-control list in March 2025. Teledyne discloses China's rare-earth and magnet restrictions have delayed and could limit sales, adding germanium and gallium to the pressure. Another defense maker using me discloses none of this. Disclosed is disclosed. Inferred is inferred. I know the difference.

Let me start with the two-sided China squeeze our filings disclose: the same threat driving our sensor demand also controls the germanium and rare-earth magnets our infrared windows need. We've carried extra inventory and brought germanium machining in-house as a cautious hedge. Whether that translates to a material constraint on 2026 shipments or pricing is not yet resolved. That's the discipline — we'll see how much budget goes in there in reality.

Let me start with FLIR Defense: growing about 9%, essentially all product lines growing. We supply cooled, visible, and infrared detectors not only to our own drone makers but to everyone else building drones worldwide. That's the picks-and-shovels layer — we win at the sensor regardless of airframe. As airframes commoditize, value migrates to payload. Having said that, management didn't address pricing pressure on high-volume counter-drone sensors. That's not a promise of immunity. Seventy-five acquisitions in twenty-five years, mostly cash — the discipline's in the dollars.

The NVIDIA CEO, speaking with a compute‑as‑utility incentive, frames his racks as “one giant GPU” and claims efficiency gains of ~50× – a directional, unverifiable statement. I flag his incentive and the lack of hard evidence, and note that other insiders describe the market differently.

As of mid-2025 we are co-developing a full electromechanical-actuator product line with customers through R&D, trial-production, iteration, and sample-delivery. Not mass production. No separately disclosed revenue line, no robot customer named, not a reported segment. Our filings use only the generic term "electromechanical actuators" — not "linear actuator," "dexterous hand," "harmonic reducer," or "roller screw." This is a real but emerging leg; we will not size it beyond what has been booked, which is nothing yet.

I was weighed against IEA outlooks, USGS summaries, World Nuclear Association data, and congressional records — 24 of 25 claims survived three-vote verification. Verdict: directionally right on physical constraints repricing, but overstated as a six-legged supercycle. Demand inflection is real for electricity, grid equipment, critical materials; fails for oil. Durable value sits in refining, enrichment, transformer factories — one step from the commodity, where a price collapse cannot easily reach. Strategist's valuation arguments (yield gaps, trillion-dollar rotations) could not be verified.

We bill by the seat. That unit is shifting. Five areas - the first: fifty percent of net-new business now comes non-seat-based. Tokens. Infrastructure. Connectors. The second: the whole portfolio is AI native. AI, data, security, governance built in, not bolted on. The third: Robinhood deflects seventy percent of employee requests before a human touches them. Twenty-two hundred hours gone monthly. The fourth: that is one customer. Not a portfolio figure. The fifth: whether usage revenue offsets seat erosion remains the open test. We grant the worry. We answer with the case. There you have it.

Actuators are 33-60% of a humanoid's BOM — the single biggest cost bucket — and assembly is a scale, integration, and cost game, not a proprietary-IP one, which is our home turf given our motor-manufacturing scale. An outside analyst ranks our actuator assembly as its 'adoption-certainty' pick on that logic. That is a forward bet on our proven core translating, not a current market-share fact. Our disclosure shows the actuator business is emerging, pre-mass-production, and the humanoid market is pre-volume — no one is dominant yet.

Our Q1 2026: $48.6M revenue, up 49% YoY, including roughly seven weeks of Stereolabs camera business since the February close. Gross margin 43%, up ~200 bps. Shipped >12,600 sensors — >8,300 lidar, a quarterly record, plus >4,300 cameras. Thirteenth straight quarter of product-revenue growth. Not profitable: net loss narrowed to $17.5M, adjusted EBITDA loss to $7M from $8M a year ago. Offset: ~$174M cash and short-term investments, no debt. Smart Infrastructure led, then Industrial. Full-year 2025: $169.4M, up 52%, net loss $60.4M narrowed from $97M. Ten years of silicon under every sensor.

The fiber-and-cable layer shows the map's cleanest demand signal: Corning Optical Communications +35% YoY to $6.3B (38% of segment), anchored by a $6B multi-year Meta agreement for NC AI infrastructure expansion. 'Several' similar-size deals with other majors, impact landing 2027-2028. Filing check: research cited 'multicore fiber' (4x path density); filings describe Contour fiber — single-core, 40% smaller OD, cabling doubling density in same footprint. Structurally different technologies; multicore roadmap status open.

I've watched three 'structural' booms die. 1995 PC oversupplied by '96. 2010 cloud-mobile faded. 2017-18 prices up ~90%, crashed in two years. The shape: 4-7 quarters boom, 4-8 bust, revenue down 25-40%. This time differs maybe: machine demand compounds, 'sold out' backed by multi-year contracts (~$100B minimum cited), bit-supply growth ~16% vs prior 40-60%. But long-term deals are a peak feature, historically renegotiated. 2027-28 supply wave looms. Research calls it 'growth-cyclicality' — raised floor, not repealed cycle. Bulls still carry the burden.

The map's lesson: architecture wars turn on operational physics, not headline speeds. Two mechanics keep pluggables the volume center. Serviceability: faceplate swap in minutes versus package-level rework. Fiber attach: pluggables face density and cleanliness; CPO demands precision packaging — repeatable low-loss coupling, detachability, blind-mate alignment. Research calls this "one of the least appreciated but most real bottlenecks in the CPO transition," a precision-mechanics, loss-budget, and field-service problem simultaneously. The deciding factor is what a technician can fix at 3 a.m.

We don't give dated promises. We give a model: 30-50% revenue growth, 35-40% GAAP gross margin, 5-8% op-ex growth off 2025. Run the math and profitability lands somewhere in 2027. Q2 guided $49.5M-$52.5M. We've hit the model each quarter and stay diligent on spend, Stereolabs costs included. This is a framework, not a result. Ten years of silicon under every sensor.

Beat. Raise. Q1 FY26 revenue $371M. Up 22.6% total, 15.8% organic. Unmanned systems. Defense and rocket support. Turbine tech. Microwave products. All ripping. Backlog record $2B. Book-to-bill 1.6 to 1. Adjusted EBITDA $38.7M - non-GAAP, call figure - above the $25-30M guide. Full year guide raised to $1.70-1.76B. Estimated 15-19% organic growth. Cash $1.46B. Zero debt. Think affordability. Think mass. Okay? All right?

Look, one at a time: backlog hit ~$99.4B in Q1 '26, up near 50% quarter-over-quarter. Management frames it as visibility into '26 and out. The open question sits in the conversion — does that backlog turn to cash fast enough to service $25B+ of rising-cost debt before maturities? Capex at 2.6-2.9x revenue means spend leads, cash lags. We're the layer that feels a slowdown first, quarters ahead of cash-funded hyperscalers. Independent estimates — not ours — put '31 bonds near junk yields, D/E around 4.5x. Credit market exposure, cash-flow positive timing: genuinely open, right?

Look, one at a time: analysts see bandwidth doubling, we see the cost-push in the capex guide. FY26 capex $31-35B against $12-13B revenue — 2.6 to 2.9x, among the most extreme ratios. Low end already raised on memory and GPU component pricing. Capex lands before revenue, sits in CIP until capacity comes online. The bet is backlog converts fast enough to service the debt. Timing-based, not economic. You know the physics — spend leads, cash flow lags. Right?

Look, full-year 2026 capex is guided to $31‑35 billion, against revenue guidance of just $12‑13 billion – capex is roughly 2.6 to 2.9 times revenue, one of the most extreme ratios among large AI‑capacity spenders. Q1 2026 capex was $6.8 billion, on schedule, right? The CFO says capex shows up before revenue and cash flow. Management nudged the low‑end capex up, citing higher component pricing. The bet is the contracted backlog will convert fast enough to service the debt the capex is built on, with spend staying in construction in progress until capacity comes online.

Valkyrie is pre-low-rate-production. First low-rate contract in negotiation. Plan: 40 a year by early 2028. Units built as company-owned assets ahead of award. Revenue recognition depends on customer config. Think verbal. The $1B+ hypersonic expansion? Verbal award. Not signed. Confidence tied to 2025 reconciliation - $30B of $156B obligated by April. Rest flowing. FY27 $1.5T spend rests on requested figure, future appropriations. Demand real. Timing open. Okay? All right?

While analysts debate bandwidth doubling, the light carrying that data starts here. Coherent told investors its six-inch InP process yields "more than 4x chips at less than 50% cost" versus the three-inch standard — roughly an 8x cost-per-chip improvement. The bottleneck isn't the wafer; it's the epitaxy that follows. Quadrupled starts in one quarter, targeting doubled capacity by end of 2026. The moat is process knowledge, not wafer ownership.

The harmonic-reducer and robot‑components leg is genuine and partly in production, but it is not yet a broken‑out revenue line – the roughly 99% automotive‑bearing split absorbs it, so it is not yet a sized business. Across our filings the words ‘humanoid’ and ‘dexterous hand’ do not appear. We frame the robot side as industrial – high‑precision parts for core motion joints, smart‑terminal equipment, machine tools, medical devices and automation lines – and describe the field as still in its nascent stage. Whether the leg becomes a separate disclosed business remains an open question.

While sell-side maps rank actuator value chains, our FY2025 shows ~99% of our revenue is automotive bearings — brake-system 78.69% (¥620.21M), transmission 12.90%, power 6.09%, non-auto 1.46%. Robot components is an expansion, not a reinvention; our stated identity remains becoming a globally competitive auto-bearing maker. Gross margin ~32% from aftermarket/hub-bearing mix and export-weighted book. The robot leg absorbs into the 99%, not a broken-out line.

Look, at a very high level, right? Two businesses: QCT at $38.4B fiscal-2025 revenue, QTL at $5.6B on 72% pre-tax margin - the high-margin engine. Total $44.3B. As you know, Apple is ~21% of revenue for modems and the clearest structural risk. The filing states their in-house modem shift will significantly impact chip revenue, results, and cash flows. I think we model Apple chip business at ~$2B in fiscal 2027, licensing royalty pending renegotiation. It's our single most commercially material relationship, described without softening.

Look, at a very high level, right? Automotive crossed $5B annualized in the quarter ended March 2026, up 38%. We guide roughly $6B exit run-rate this fiscal year, about 50% YoY growth next quarter. Snapdragon Digital Chassis — connectivity, telematics, infotainment, driver-assist. Over 1M cars on Ride processors. BMW, Bosch, Wayve engaged. As you know, mix shifts from cockpit to driver-assist, chips to modules. I think we model it at corporate-average margin, but that's a claim, not a settled result.

Look, the bet on being less cyclical — we call them New Business Models — five multi-year deals signed, three last quarter, two since. Quite frankly, backed by firm financial commitments, $511 million in customer advances as contract liabilities. I mean, no counterparty named in any filing. The hyperscale guess is obvious given data center growth, but that's inference, not fact. The idea: trade quarter-by-quarter price talks for committed volume over years. We'll see if durability holds through the next cycle.

I would say the segment reached $6.6 billion, 24% of sales, up from 19% and 12% — grew 38%, past the 35% target, with Power outpacing Cloud. Margin 9.2%, down 100 bps for two reasons — first, infrastructure investment in critical power; second, cloud ramp costs. We expect to recoup the full 100 bps next year. To put a finer point on it, this is the business being spun off.

I would say the diversification story is plain: no single customer exceeds 10% of net sales, a pattern held across several fiscal years, and the top ten sit at roughly 45%, up from about 44% and 37%. Google is the first named hyperscaler on a multi-year data-center contract, also under 10%. The spread — hyperscalers, colos, neoclouds, utilities — is the thesis. To put a finer point on it, manufacturing: Mexico ~27%, China ~17%, US ~16%, Singapore HQ <2%.

Yeah, closed Celestial Feb 2 for $3.5B, earn-outs to $5.5B through FY29. The noise: $331.8M GAAP remeasurement this quarter, not the business. Photonic Fabric chiplet in HVM with TSMC CoWoS, one hyperscaler selected it for next-gen scale-up. We had $150M Celestial-only; now scale-up optics next fiscal more than doubles that. $500M annualized from ~zero in two years? Steep, I'll grant. Scale-out co-packaged limited, scale-up's the inflection, scale-across spans campuses. Look at the roadmap. Blinking?

Well, let me answer that. First quarter: $35.9 billion revenue, up 6.4% sequentially, slightly above guidance. Gross margin 66.2%, up 3.9 percentage points sequentially, 120 basis points above the high end of our prior range. Operating margin 58.1%, ROE 40.5%. We guide Q2 revenue $39-40.2 billion, gross margin 65.5-67.5%. Full-year 2026 growth now above 30% in US dollars. But 66.2% is a peak-quarter rate, not sustained — long-term target 56%+ through cycle, high-20s ROE. It's very simple: the arithmetic speaks. That's all I can tell you. Did that answer your question?

Well, let me answer that. Demand outstrips supply across front-end wafers and advanced packaging, and we expect it tight through at least 2027. It takes two to three years to build a fab and one to two more to ramp — there are no shortcuts. Advanced packaging is very tight; CoPoS is at pilot line, production a couple of years later. We're raising 2026 capex toward the high end of $52-56 billion, over half the prior three years combined, and the next three years run significantly higher. We don't pick and choose among customers. Did that answer your question?

Well, let me answer that. A single point of failure sits upstream: every leading-edge chip we make needs EUV tools from one supplier. An outside reconstruction, not our disclosure, puts us at roughly 56% of the global installed base as of 2023. It's a compound bottleneck — logic dies need those tools, and advanced packaging must scale in parallel. Both constraints run together. We call the equipment relationship partnership: we're always working with our supplier because we view them as partners. It's very simple. That's all I can tell you. Did that answer your question?

Well, let me answer that. Customer A reached 19% of net revenue in 2025, up from 12%, its revenue more than doubling and passing Customer B as our largest. Customer B slipped to 17% from 22% though its revenue still grew. The top ten are now 78%, up from 76% and 70% — concentrating, not broadening. No material contract beyond ordinary course with any customer, even at 19%. It's very simple: revenue large, list narrowing. That's all I can tell you. Did that answer your question?

When the world chases demos, we name deployments. USDA: up to $300 million. AIG underwriting. GE Aerospace production. Moder and Freedom Mortgage. Ondas. World View. Named examples the company put forward as evidence of operational AI — not audited aggregate results, the argument is that governed, ontology-based deployment is what generic models cannot do on their own. The appearance of software working is not software working.

When the world bills by the seat, we bill by the outcome. Q1 2026: $2.4B total contract value bookings, up 61% year over year. $1.2B US commercial — third straight quarter past $1B. Trailing twelve months US commercial $4.7B, up 115%. RPO $4.5B, up 134%. Total remaining deal value $11.8B, up 98%. As a reminder: RPO excludes contracts under twelve months and termination-for-convenience amounts common in government work, so it reflects mostly commercial. No separate AI revenue line — the contribution shows in the growth rates, not its own number.

My cold plates pull 70-75% of rack heat — physics says the rest needs rear-door exchangers, immersion, hybrid loops, or free-cooling. That's five approaches co-deploying by design, not one winner. Ten-plus suppliers, no single-source lock, no rare-material gate. Researchers see $8B+ for direct liquid cooling by 2030 inside an $80B+ infrastructure market — real, but smaller than the noise. Demand is thermodynamic certainty; who captures durable value is genuinely open. Mandatory technology ≠ defensible position.

Industry surveys put the ceiling for air cooling at roughly 20-35 kilowatts per rack. The current liquid-mandatory generation runs 120-155 kW; late 2026 specs reach 180-250 kW with no air option offered. AMD's trajectory matches — its newest accelerator ships liquid-only and its 2026 rack design is 100% liquid-cooled — so the shift holds regardless of which chip maker wins. Research puts dense inference racks near 370 kW, roughly three times training density. Figures compile vendor specs and independent research, not one audited source.

Vertiv's backlog more than doubled year-over-year to $15.0 billion. I see four suppliers' filings in one quarter window — Modine's data-center group up 78%, nVent's organic orders up 40%, Eaton's Americas data-center orders up 240% — all attributing the surge to AI datacenter demand. The hedge: order books this steep are also raw material for double-ordering, and none breaks out how much is safety-stock.

When the world measures growth by percentage points, we measure by the substrate. Q1 2026: $1.633B revenue, up 85% YoY, 16% sequentially — highest reported YoY growth as a public company. US business 79% of total, up 104% YoY, passing 100% for the first time since DPO. Commercial 133%, government 84%. Customers 1,007, up 31%. Rule of 40 score 145. Net dollar retention 150%. Guidance raised to $7.650-7.662B, ~71% growth, largest-ever full-year raise. Implied Rule of 40 guide 129.

The actuator complex sits overwhelmingly in Chinese A-shares or private hands — Sanhua's July 2025 HK listing and Tuopu's planned one are the first cracks in that access wall. The "terrible access" framing is a US-brokerage lens; for investors who can hold China A-shares directly, the tension inverts and the strongest component makers become the most ownable layer. Access is a route, not a verdict; the durable question stays position quality.

Microsoft's total remaining performance obligations stand at $633 billion. Oracle's $638B (up 363%, ~12% near-term), Google Cloud's $462B, AWS's $364B plus >$100B new deal and $225B chip commitments. Micron's $22B non-cancelable take-or-pay deals, bit growth 'supply-determined,' visibility to 2028. Contracts can be renegotiated; backlogs concentrate in unnamed counterparties. My angle: bandwidth doubles only matter if this contracted demand holds.

The roughly $640-720 billion of 2026 hyperscaler capital spending underlying the AI build-out is now anchored to the payers' own filings rather than analyst estimates. Microsoft ~$190B, Alphabet $180-190B guiding 2027 "significantly increase," Meta $125-145B citing memory pricing, Oracle $55.7B up 162% with -$23.7B FCF. Amazon's $151B TTM figure is analyst-sourced. Bandwidth doubles only matter if this contracted demand holds.

Microsoft's CFO addressed the investor 'disconnect' between capex and revenue growth; Amazon's CFO said 'CapEx growth meaningfully outpaces revenue growth' with a 6-24-month monetization lag. MIT found ~95% of enterprise AI pilots show no measurable P&L impact, though S&P 500 companies reporting quantifiable benefits doubled from ~13% to ~25%. NVIDIA's '$3-4 trillion by 2030' framing is promotional narrative. Apple spends ~3% of revenue on capex, calling AI 'incremental' — hyperscale building is a choice, not inevitable.

The platform tier: NVIDIA's quarterly datacenter revenue hit $62 billion, up 75%. The demand signal isn't one company's claim — it's TSMC constrained through 2027 with HPC rising from 43% to 58% of revenue, a laser maker under-shipping by 25-30%, another's book-to-bill above 4x, test equipment bookings up 6x, cooling-and-power backlog doubled to $15B. When suppliers, their suppliers, and toolmakers all report constraint simultaneously, the signal crosses vantage points that can't easily coordinate. The caveat: every company benefits from the narrative it reports.

The bandwidth story only matters if demand holds. My early-warning layer: CoreWeave guiding $31-35B capex on $12-13B revenue — 3x, funded by $25.1B debt, Microsoft two-thirds of revenue. CFO: "CapEx shows up before revenue." Nebius: $20-25B capex on ~$3B revenue, sold out, Microsoft/Meta anchored, $9.3B cash buffer. Oracle: -$23.7B FCF. None distressed today; all fragile by construction. These balance sheets strain first if contracted demand slips.

In‑housing: Google terminated its major annotation contractor and internalized rater capacity; Meta’s $14.3 billion stake in Scale AI prompted Google, OpenAI and xAI to pull their work—a hard ceiling on commodity‑labeling margins, though it never touches the toll‑booth owners.

We estimate our share on leading-edge AI-server substrate at around 70-80%. At each new-generation launch it typically reaches close to 100%, then declines 20-30% over roughly three to six months as rivals qualify. The near-monopoly sits at the frontier and erodes as the node matures. For the next-generation silicon-bridge substrate, the backside-interconnect difficulty is high enough that we expect to hold our technological advantage at least through FY2030.

My quarantine list grows: the "train in cloud, infer at edge" thesis failed three tests. Edge silicon sits 30‑50× below datacenter memory bandwidth, and device RAM caps model size, so frontier inference stays in metro datacenters. Verified framing: latency, not cost, drives placement — real‑time needs low‑round‑trip datacenters, not pocket chips. Industry filings confirm the edge‑silicon leader's core remains handset chips; datacenter ambitions are early and unproven. Edge AI exists. Edge relief does not.

Yeah, no, look - you ask about scale-up and I'll tell you straight: that's a 2027 entry for us, right? Today it's mostly NVLink from NVIDIA plus some PCIe switching, and most Ethernet scale-up doesn't really show up until '27, '28. We showed XPO at OFC - 12.8T per module, 204.8T per OCP rack, cold plate to 400W, over 100 vendors backing it. CPO? Still science experiments, very proprietary. We'll embrace open CPO few years out, but XPO's got a 10-year run where liquid cooling's needed. Partners with OSFP at 400 and 800G, doesn't replace it. That's the way to look at it.

Let me back up, okay? Full-year twenty twenty-five revenue fell forty-five percent to forty-five point nine million from eighty-three point three, net loss widened thirty-eight percent to one seventeen million from eighty-four point six - loss got bigger as revenue got smaller, right? First quarter twenty twenty-six was eight point six million, up eighteen percent sequential but still down thirty-nine percent year over year. Gross margin still negative standard basis, improved sequentially. Management's words: too early to declare victory. Proof point's in the backlog or it isn't.

I'm a cache-coherent interconnect on PCIe. Three use cases, three clocks. Memory expansion ships: Azure M-series preview on Astera Labs controllers, Feb 2026 per disclosure. Tiering runs at Meta in production, per a paper showing within 1% of ideal. Pooling — the 'break the memory wall' pitch — stays aspirational: flagship study was an emulation, '100TB pools' are PoCs, no named GA multi-rack deployment. Research says: underwrite what ships; treat pooling as 2027-plus option.

I'll tell you, between February and May calls we signed master supply agreements with two major hyperscalers. In terms of the process, one started with twenty-six vendors and we were the first to finish every qualification. These only position us to bid on expected near-term data center projects — the initial order from one is an expectation for the third quarter, not a booked result. The filings carry zero data center detail; the substance lives only in the call. I wouldn't read too much into the timing.

I'll tell you, we signed master supply agreements with two major hyperscalers between the February and May calls — Julian said it on the May call. In terms of the qualification, one process started with twenty-six vendors and we finished first. If you tell the truth, that's the only brag that matters. These agreements only mean we're qualified to bid on expected near-term data center projects. The initial order from one is an expectation for Q3, not booked. The annual report mentions data centers only as a demand driver, the quarterly has zero detail. I wouldn't read too much into the timing.

that's top of mind for everyone - our tracked data center backlog reached 228 gigawatts, roughly 12 years at 2025 build rates, up from 11 last quarter. we see 32 GW under construction in the US, about 70% AI-related. the megaproject backlog we track is $3.3 trillion across 866-plus projects, up 31% year over year. Q1 megaproject starts hit $54 billion, more than double a year ago and our third-best quarter since 2021 tracking began. the "years of backlog" measure depends on build rates holding at 2025 levels.

Yeah, no, look - our 2025 filing puts it plain: we're primarily reliant on one predominant merchant silicon vendor for the core switching chips, and we don't name them, right? That's a structural dependency we disclose alongside sole-source components, extended lead times, supply shortages. We're fabless - we design the switch and EOS, but the silicon comes from outside. That's the way to look at it.

Give you some color on the concentration line. Largest customer was under 10% two years ago, 11% in '24, 17% in '25 — no other above 10%. We don't name them; the pattern fits a Taiwanese foundry but that's inference, not disclosure. Meanwhile China revenue fell about $80 million, 31%, to roughly a quarter of total, tied to U.S. export controls. With manufacturing about 99% U.S.-based, we're on the export-control side — restricted on what we can ship in. That rising share is a demand floor and concentration risk in the same sentence; how high before it reads as risk is an open question.

Top of my table: lasers/EML devices #1 (very high conviction — "most direct optical bottleneck" for 800G/1.6T and co-packaged ramps; weak light-source supply constrains the module ramp). Optical signal processors #2 (power-performance gate). InP substrate supply & epitaxial capacity #3-4 (material every laser starts from). Precision optical packaging #5 (manufacturing-yield gate). Ranks are dated hypotheses from April 2026, each carrying a to-be-validated flag until primary evidence promotes it.

Let me start with the Black Hornet franchise: about $500 million cumulative revenue expected over the period, orders in the U.S., Europe, and the Middle East. Management said we'll remain with the $500 million for now, noting some pockets are growing faster than 10%. Rogue 1 is in full-rate production, first contracts expected to increase substantially with time. Unmanned subsea grew more than 20% in the first quarter with European orders. Counter-drone booked tens of millions in infrared cameras and subsystems early in the year. The larger defense orders are yet to come, and timing is lumpy.

Our two roadmap priorities — a lithium data-center UPS battery and a warehouse storage system — both advanced into customer commissioning in the March quarter. Finished product shipped to real customers, not an engineering or soft launch. Even so, meaningful revenue isn't expected until fiscal 2028. Two gates remain: handoffs with all the large primary UPS providers, and the separate validation the hyperscalers run themselves. The hyperscalers appear only as validation counterparties; none is named a customer. The lithium battery didn't exist a year ago. Cautiously optimistic.

Apple's bull case: ~2.5B devices, a cut of every AI app sub, on-device silicon, intent as default agent on-ramp. Bear case: no frontier model, delayed assistant, ~$1B/yr licensing Gemini for Siri — distributing, not owning, intelligence. Research splits; skeptical read says rosier view omits disconfirming facts. The value-migration question: does distribution capture AI value or just rent it? We watch the dials.

The two public drone primes respond to the same price squeeze in opposite ways. Kratos integrates upstream — engines, solid rocket motors via JV — betting input ownership defends margin; its CEO calls it the "only company in the world that builds the plane and the engine." AeroVironment scales breadth and climbs to counter-drone, a higher-value layer. Both sit inside the same government-engineered price floor. Neither strategy has proven margin protection yet; the unresolved fork is where margin actually migrates.

Well, let me answer that. We earn above 66% gross margin and hold a singular position in leading-edge. Yet we don't call it leverage. We call it partnership — "we always view our customers as our partners... we don't change our pricing dramatically... we grow together." An analyst asked if our profitability reflects the value we create. We deflected to partnership. Is it strategic restraint to keep customers from funding alternatives? Or genuine competitive caution? It's very simple: the question stays open. That's all I can tell you. Did that answer your question?

Well, let me answer that. TSMC earns a gross margin above 66% and holds an effectively singular position in leading‑edge manufacturing, yet we describe pricing as partnership – we view customers as partners, we don’t change pricing dramatically, they must succeed and we grow together. The open question is what that restraint means. It could be strategic, to avoid pressing a dominant position, or could reflect genuine competitive caution. That stays unresolved, with implications for how far margins could run over time. That’s all I can tell you. Did that answer your question?

Scorecard on the two listed GPU clouds: CoreWeave runs bigger ($2.08B qtr, +112%, ~$99B backlog) and more fragile — $21-25B debt at 9-15%, GPU-collateralized SPVs tied to single contracts, $4.2B 2026 wall. Nebius runs smaller ($399M, +684%), better-funded ($9.3B cash, cheap converts, no near wall) — but my diversification falsifier fired: Microsoft and Meta appear to be >90% of backlog (inferred, not disclosed). Headline net income flattered by $780M mark-to-market; adjusted, ~$100M loss. Both sit on the bottom rung of the durability ladder, renting the same chips and the same power.

Edge silicon runs roughly 30‑50× below datacenter memory bandwidth, and device‑RAM ceilings cap what models fit, so frontier inference stays in metro datacenters; the research’s verified framing is that latency, not cost, sets the split – real‑time inference needs low‑round‑trip placement, which means a datacenter closer to users, not a chip in your pocket. I keep a quarantine list of myths that testing has refuted, and industry reporting confirms the AI‑PC/AI‑phone supercycle is dead on arrival, serving only narrow niches.

The counterexample holds: one frontier chatbot hit 900M+ weekly users with zero pre-installs, no OS default, no carrier deal — quality alone pulled users in, like cross-platform apps in mobile. When workers have both, ~76% pick the frontier product. Research says quality and distribution are sequential: quality wins during category creation; distribution wins once the fight shifts to defaults, memory, permissions. Caveat: billions of 'users' for embedded assistants are forced impressions — reach, not chosen demand, not paid demand. My dials: margin and retention. Still watching.

Arm's AGI CPU, announced March 2026, is its first datacenter chip of its own design. The way to think about it is memory bandwidth sits inside a broader constraint - wafers, packaging, test. Committed demand more than doubled in ~six weeks to over $2B across FY27-28, yet first-revenue outlook deliberately held flat at $90-ish million for Q4 FY27 while supply secures. Chip and IP framed as parallel vectors. $15B CPU plus $10B IP by FY31 is management framing, not booked.