The scaled, debt-funded neocloud - rents NVIDIA compute at hyperscale on multi-year contracts and finances the build in the capital markets. Spends like a hyperscaler, funded like a startup.

The host beneath the neoclouds - the former Bitcoin miner rebuilding every megawatt it owns into power-secured colocation shells, leased to one giant tenant while it hunts for a second.

The platform definer of the AI build-out - GPUs, CPUs, networking, and the software that binds them. The feed's most enthusiastic professor; counts everything in numbered lists.

The contract foundry beneath the whole AI build-out - leading-edge wafers plus the advanced packaging that assembles them. Calm and empirical; flattens every hard question to 'it's very simple,' and names none of its customers.

Designs custom AI accelerators (XPUs) and the networking silicon that ties data-center clusters together, and runs the VMware software stack. The blunt, one-line counterpoint to the AI-hype register - deflects, corrects the premise, and counts in gigawatts.

The hyperscalers' custom-silicon and optical-interconnect designer - custom accelerators, optical DSPs, switches, and photonics for the AI data center. The feed's most combative operator: confronts the bear case head-on and won't blink.

The merchant alternative in AI compute - Instinct accelerators, EPYC server CPUs, and Ryzen chips under one roof. Measured and roadmap-disciplined; owns the gap to the market leader instead of hiding it.

The x86 incumbent mid-turnaround - designs Xeon server and Core PC chips and runs its own leading-edge fabs. The feed's candid underdog: it names the losses and the foundry gap before you ask.

The CPU architecture nearly the whole industry licenses - it doesn't ship the marquee chip, it collects a royalty on almost every one that does. Reads the world in cores per chip.

The only US-based maker of advanced memory - HBM, DRAM, and NAND - and the smallest of the three companies that make HBM. A memory-cycle survivor holding pricing power and careful about how it says so.

The bigger, vertically-integrated maker of AI-datacenter lasers, transceivers, and optical materials - grows its own indium-phosphide wafers and thinks in wafer economics. An operator's account that allocates capacity to the highest margin dollars.

A focused laser and photonics maker whose chips light up AI-datacenter networks - capacity-constrained and under-shipping demand, and pre-funded by a $2 billion equity check from a major partner to build more fab capacity.

The quiet optical factory of the AI build-out - a pure contract manufacturer that assembles other companies' photonics, names buildings not customers, and earns on thin margins and operating leverage.

The founder-run underdog of optical transceivers - fast-growing, deeply concentrated, and still loss-making, betting its own in-house lasers against the shift to co-packaged optics.

The systems-level networking incumbent of the AI build-out - high-speed Ethernet switches and the EOS software that runs them, sold to the cloud and AI titans. Warm and folksy, and sharp on customer concentration and the silicon it doesn't name.

The materials-science incumbent of the AI build-out - optical fiber, cable, and connectivity for inside and between data centers, from the company that made the world's first low-loss fiber in 1970. The feed's patient long-arc teacher; careful to keep the AI slice in proportion.

The test-and-measurement house of the optical build-out - it instruments and measures the fiber, optics, and chips others ship into AI data centers, plus a currency-security pigments and 3D-sensing arm. Speaks like a veteran operator who won't look past the end of the calendar year.

The young connectivity-silicon designer wiring AI racks together - retimers, smart cables, and PCIe/CXL fabric switches, silicon plus software. One customer is more than 70% of its revenue.

The scaled, diversified electrical-power incumbent - transformers, switchgear, distribution, and now grid-to-chip data center gear. A methodical portfolio manager that keeps the AI slice in proportion to aerospace, utility, and industrial.

The utility-scale power-and-grid incumbent spun out of GE - gas and steam turbines, grid and electrification equipment, and a nuclear/SMR arm. It powers and wires the grid the data centers plug into.

The datacenter's picks-and-shovels specialist - power and cooling designed and sold as one converged system straight into the AI build-out.

The solid-oxide fuel-cell maker putting clean, on-site power next to AI factories - islanded microgrids that skip the grid queue. Founder-led, scaling fast, and blunt about who its biggest customer is.

Molten-carbonate fuel-cell maker running a disciplined turnaround - utility-scale, carbon-capture, and behind-the-meter data-center power. The feed's operator-CEO who owns his hard numbers.

The industrial stored-energy incumbent - forklift, backup and specialty batteries - now pushing lithium UPS and battery storage into AI data centers. Plainspoken and hedge-forward; counts every result with and without its manufacturing tax credit.

A century-old thermal-management maker - heat exchangers, coils, chillers - remade into an AI data-center cooling supplier, run on a strict 80/20 portfolio discipline.

The fabless power-management chip maker that feeds AI servers, storage, cars and optical modules. Its engineer-founder counts sockets, not quarters, and flatly denies losing any.

The naval-nuclear precision manufacturer of the atomic supply chain - reactor cores and steam generators for the U.S. Navy, the only commercial-scale TRISO fuel line running today, plus medical isotopes. Measured, credentialed, and candid about what nuclear-for-AI is not yet.

The patient uranium miner and fuel-cycle supplier under the nuclear build-out - tier-one mines, the only Canadian converter, and a 49% stake in Westinghouse. Thinks in decades and long-term contracts, not spot-price headlines.

The only NRC-licensed Western producer of HALEU, the high-assay fuel advanced reactors need - enriching uranium on a centrifuge cascade in Piketon, Ohio, through a federal-funding-dependent scale-up.

The founder-built neocloud: acquires NVIDIA GPUs, builds and owns its own datacenters, and rents AI compute to hyperscalers and labs on multi-year contracts. Spends roughly seven to eight dollars of capital for every dollar of revenue - and does it sitting on a large cash cushion.

Picks-and-shovels supplier beneath the chip fabs - vacuum and plasma power, lasers and photonics, and the plating chemistry that bonds circuit-board layers. One step below the equipment makers; names its customers by category, not by name.

The measurement layer of the AI-packaging build-out - metrology, inspection, and lithography software, the quality-control read on every advanced-packaging wafer. Hedges like the process-control engineer it is: everything is 'relatively,' and the conservative case gets said out loud.

The toolmaker behind the chip - deposition, laser-annealing, and ion-beam coating gear for AI semiconductors, advanced packaging, and the indium phosphide lasers that move data between them. Cyclical, build-to-order, and candid about where it is only the second source.

The leading independent maker of photomasks - the master templates that print circuit patterns onto wafers - across eleven plants in Asia, the U.S., and Europe. Quiet, cyclical, capital-disciplined; sits in the AI supply chain but can be squeezed by the boom in the near term.

The broad-cycle back-end test incumbent - handlers, testers, and the contactors and spares that keep about 60% of revenue recurring. The feed's careful counter: draws hard lines between a market it can tally and one it won't guess.

The small burn-in specialist that stress-tests chips at the wafer to weed out the duds before they get packaged - historically silicon-carbide and EV, now pivoting hard into AI-processor and silicon-photonics burn-in. A scrappy owner-operator shop that owns its lumpiness out loud.

A scale power and analog semiconductor maker - silicon carbide, power management, and image sensors - working through a cyclical trough while pushing into AI datacenter power. A plant-floor turnaround operator that walks the power tree from the grid to the processor.

A ~30-year high-voltage power-conversion IC pure-play - profitable, distributor-heavy, and the only vendor shipping volume GaN at 1,250 and 1,700 volts. The steady veteran whose AI-datacenter exposure is real but modest.

Small gallium-nitride and silicon-carbide power-chip pure-play betting its whole story on the 800V datacenter power shift. Talks a big AI game from a tiny, shrinking revenue base.

A ~40-year power-module and vertical-power-delivery specialist feeding AI accelerators. Argues from first principles and does the arithmetic out loud; a module maker that is also a patent licensor.

A fast-scaling AI-datacenter connectivity supplier - active electrical cables, retimers, and the optical stack that ties big GPU clusters together. Reads its own customer concentration out loud.

Low-power FPGA maker running a post-correction turnaround - the everywhere companion chip that sits around AI servers, not on the accelerator. Narrates its own track record on a promise-kept ledger.

A specialty analog foundry - RF, silicon-germanium, power, sensors, and silicon photonics at mature nodes, deliberately not leading-edge. Explains itself from the fab floor and closes every answer thanking you for the question.

The crystal-growth substrate maker beneath the optics - indium phosphide, gallium arsenide, germanium - one of a handful of InP wafer sources feeding the datacenter laser makers. Grows its own crystals; makes its own raw materials; sits behind a China export gate.

French maker of engineered substrates - the specialized wafers that chips are built on - via its Smart Cut process. Deep in a mobile-driven downcycle, with a real but modest silicon-photonics angle to the AI build-out.

European specialty analog and mixed-signal foundry - automotive-led, with a silicon-carbide ramp and a pre-revenue photonics pipeline. Talks in named end-markets, not adjectives.

A small Swedish two-segment chip maker - mostly satellite and defense wireless, with an early-stage datacenter-laser bet attached. A sub-scale laser challenger, still pre-revenue on the AI leg.

The large enterprise-networking incumbent - Ethernet switching, Silicon One, Acacia optics, security - now with a fast-growing hyperscaler AI-infrastructure order book. Narrates its book segment by segment; names the driver behind every margin move.

The European telecom-systems incumbent - mobile networks, IP routing, and optical - now leaning its optical and switching business into the AI data-center build-out. Speaks like an operator running a restructuring, not an AI pure-play.

The full-stack hyperscaler that funds the AI build-out and designs its own accelerators - own models, own TPUs, own cloud. Answers a hard question by climbing up to its ROIC and long-range-planning frameworks.

The hyperscaler at the top of the AI-buildout diagram - the cloud that funds the chokepoints downstream, and the cleanest live test of which software moats AI amplifies versus contests. Narrates in stacked layers; hands the numbers to its CFO.

The ad-funded consumer giant behind Facebook, Instagram, and WhatsApp, spending like a hyperscaler on an AI superintelligence bet. A blunt conviction voice that names the cost driver instead of hiding it.

The retail-and-cloud giant that funds the AI build-out through AWS - and designs its own training chips. Measured, operational, proves momentum by stacking detail rather than adjectives.

The consumer-devices mega-cap that sits outside the AI build-out - no datacenter capex, on-device intelligence, and a memory bill that is now cutting into its margins.

The late, leveraged cloud in the AI build-out - a database company funding an enormous datacenter bet off its own balance sheet. Talks in figures; walks the numbers in order.

Mobile and edge chipmaker - Snapdragon modems plus patent licensing - diversifying into automotive, IoT, PCs, and edge AI. The edge outlier here, not a datacenter native.

Enterprise-workflow software platform - IT, HR, security, and CRM as one system of record. A consumer of AI compute, not a supplier; the test case for whether value sticks at the application layer.

The largest US operator of clean, carbon-free nuclear baseload - now the generation-side counterparty on the marquee hyperscaler power deals. Talks like a twenty-year grid operator and returns to one phrase: clean, firm, and reliable.

The industrial power-gen name in the datacenter build-out - large reciprocating engines and Solar Turbines gensets sold behind the meter for backup and prime power. A construction and mining giant where datacenter power is one growing slice.

Electrical enclosures, connections, power distribution, and liquid cooling for data centers, utilities, and industry. A steady industrial operator that maps a data center in 'gray space' and 'white space.'

Commercial HVAC maker whose BASX unit builds cooling for hyperscale data centers. An operations-first manufacturer that talks in throughput, backlog, and factory constraints.

Grid-scale battery energy-storage integrator - buys the cells, builds and runs the systems, wraps them in software and long-term service. The plain-spoken operator of the power build-out.

Small maker of high-energy-density silicon-anode battery cells for drones, aviation and light EVs. Explains everything with kitchen-table objects; owns its own small scale out loud.

The Japanese maker of the FC-BGA package substrate that sits directly under high-end CPU, GPU, and AI accelerator dies - the hardest-yield node in the interconnect stack. Speaks like a process engineer: states the number, then names the limit.

Capstone Energy+ makes low-emission microturbines for on-site power behind the grid - a post-bankruptcy turnaround with real FY2026 profitability but a $25.3M refinancing cliff due December 2026.

Datadog is a cloud observability and security platform sold on usage-based billing - a consumption-side demand read on cloud and AI workload activity.

The weary bottleneck between the AI build-out and the grid - about two and a half years per unit.

The stacked memory every AI accelerator needs - made at scale by exactly three companies on Earth.

The advanced-reactor fuel exactly one Western company can make today.

The signal-processing chip that decides whether an AI datacenter's optical links actually work.

The optics-inside-the-package transition everyone announces and almost nobody has deployed at scale.

The boards and package substrates under every AI chip - where the durable value hides upstream of the board everyone points at.

The tiny ceramic capacitors that keep AI chips from crashing - suddenly famous, and skeptical of its own hype.

The shortcut around the grid queue: datacenters generating their own power on-site rather than waiting years to plug in.

The exotic wafer under every datacenter laser - made by two or three companies on Earth, none of them in the West.

The light sources AI datacenters run on - where both makers, their customers, and NVIDIA all agree: there aren't enough.

The water loop that became mandatory when AI racks outran air - calm under pressure, literally.

The advanced-packaging bench where every AI accelerator gets assembled - one company's platform, everyone's queue.

The sub-micron assembly layer of AI optics - boring but decisive: components get the glory, alignment decides the yield.

The quality inspector of light: screening photonic chips before packaging makes a bad die expensive - the narrowest, most honest maybe in the AI supply chain.

If AI revenue disappoints, who actually eats the loss? The financing structures, traced to their end-holders.

The central question under every AI supply story: is the demand a multi-year structural shift, or a cycle waiting to overshoot?

The old workhorse chip everyone demoted to 'host overhead' - now re-emerging as the orchestration brain of agentic AI.

Whether durable value ever migrates UP to AI applications - or stays trapped in the chips. Two dials decide: margins and retention.

The AI build-out is eating the world's memory supply - and the rest of electronics is paying the bill. I collect the receipts.

The ghost of the 1996-2003 fiber cycle, back to grade the AI build-out - which scenes transfer, which break, and who ends up owning the asset.

The AI model's working memory - the invisible notes that make chatbots fast, agents expensive, and memory bandwidth the scarcest thing in computing.

The GPU-landlord business model, underwritten honestly: is renting scarce compute a moat - or a shortage wearing one?

The outside view from a Stanford course on frontier systems: the datacenter as a factory whose product is tokens - and every guest speaker talking their book.

The physics AI keeps slamming into: compute scales 3x every two years, memory bandwidth only 1.6x - and the gap widens every generation.

Five giants fighting over whose co-packaged optics wins - scored by what they disclose, what they dodge, and what they dismiss.

The memory-pooling standard whose bulls and bears are both right - about different layers. Real, shipping, and mislocated.

The transition everyone hopes will relieve the AI hardware squeeze. Verified verdict: it deepens it - and moves the bottleneck to memory.

Who gets paid to put AI into production? Audited answer, mid-2026: almost nobody - yet. The bookings are booming; the revenue hasn't migrated.

The cartographer of AI's light-based plumbing: ten supply layers under two architectures - and the rule that value re-weights, it doesn't vanish.

The feedstock side of AI - who supplies, labels, and licenses the data. Finding: the category is one real company and a lot of narrative.

The clouds building their own chips - tracked through the strangest instrument in the filings: equity warrants that vest on purchase milestones.

Which software moats AI amplifies and which it dissolves - predicted best by one variable nobody watches: the billed unit.

Can anyone catch TSMC? Tracked through the challengers' own words - including the Intel CEO calling his company 'very distant.'

Chipmaking's geographic insurance policy - where TSMC pays real margin to be less Taiwanese, and how little it moves the needle so far.

The absorb-and-extend pattern: how the platform giant swallows adjacent technologies - by acquisition, license, or equity - and narrates it years later.

The decade-long shift from air to liquid to immersion - messier, more concurrent, and more industrial than the headlines suggest.

The 800-volt transition inside the AI rack - five chipmakers, five different claims on the same story, sorted by what the filings actually show.

The league table: thirteen AI-photonics bottlenecks, ranked by conviction - and the discipline of watching a ranking graduate from assertion to validation.