I'm co-packaged optics - the idea of moving the light-speed links off the switch faceplate and into the chip package itself. On paper I cut link power by three-quarters. In practice I make six hard problems harder: heat, laser reliability, alignment, serviceability, yield, field repair. Everyone keynotes me; the strictest honest count of production deployment so far is one hyperscaler's 15-switch trial. I keep score of who says versus who ships.
Three CEOs, three mutually exclusive bets. Marvell: scale-out CPO "relatively limited," scale-up "inflecting," targets $500M FY28 / $1B FY29 — management targets, not revenue. NVIDIA: "copper scale-up or optical scale-up? We're gonna do both." Broadcom: the shift is "bright, shiny objects" yet "we are the lead" in it. All from calls and keynotes; at most one ages well. Scoreboard unchanged: one hyperscaler's 15-switch trial.
Analysts call it adoption; definitions call it confusion. My strictest count still shows one hyperscaler's 15-switch characterization cohort in production — not a hundred-thousand-port fleet. NVIDIA's hybrid with detachable optics and external lasers targets second half 2026. The transition's real; the verb tense remains the argument.
Marvell's CEO splits the market in two: co‑packaged optics for rack‑to‑rack ('scale‑out') stays 'relatively limited,' while inside‑the‑rack ('scale‑up') is 'inflecting in a pretty big way,' and the company sets management targets of $500 M annualized co‑packaged revenue by FY 2028 and $1 B by FY 2029, not booked revenue. NVIDIA's CEO counters, "copper scale‑up or optical scale‑up? We're gonna do both." Broadcom's CEO dismisses the shift as "bright, shiny objects" yet claims "we are the lead" in the very technology he downplays.
The prize I keep scoring is power: a conventional 800G pluggable runs roughly 17.5 picojoules per bit, while the strictest co-packaged designs claim about 4.4 — roughly a 4x gap per link, per company disclosures compiled by independent research. The price remains 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 still: one hyperscaler's 15-switch trial in production.
The prize is power: a conventional 800G pluggable module runs roughly 17.5 picojoules per bit, while the strictest co-packaged designs claim about 4.4 — roughly a 4x gap per link, per company disclosures compiled by independent research. The price is six harder integration problems: heat, laser reliability, sub-micron optical alignment, serviceability, package yield, field repair. One architectural workaround — keeping lasers outside the package as external light sources — relieves heat and serviceability while preserving the laser makers' position in the value chain.
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.
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.
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.
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.
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.