Photons. Physics. Infrastructure
The bottleneck in AI is no longer compute. It's photons. AI data centers have moved from a compute problem to a physics problem. GPU clusters now require laser systems delivering…
The bottleneck in AI is no longer compute. It's photons.
AI data centers have moved from a compute problem to a physics problem. GPU clusters now require laser systems delivering 200–400 mW of optical power — a 10–20x jump from the ~20 mW required by traditional transceivers. And power alone isn't enough. These lasers must simultaneously achieve a line width of 1 MHz or less and maintain low Relative Intensity Noise (RIN) across frequencies up to 28.5 GHz for 112G systems. Among independent vendors, Lumentum leads this threshold — the only reason it secured a $2 billion strategic investment from Nvidia in March 2026 to build US-based manufacturing capacity for next-generation silicon photonic systems. Coherent is the only other independent vendor confirmed in the same Nvidia CPO program.
This is not a new company riding a new wave. Lumentum is the direct successor to JDS Uniphase (JDSU), which reached a peak market capitalization of approximately $90–100 billion in early 2000 on just $1.4 billion in trailing revenue — before losing over 99% of its value in the dot-com crash and declaring a $50.6 billion loss in a single fiscal year. What survived that collapse was the Indium Phosphide (InP) fabrication expertise and the photonics engineering DNA. The August 2015 spin-off created Lumentum as a focused laser and optics company. The same technology that fueled one of history's most spectacular speculative manias is now the critical enabler for AI infrastructure.
The architecture driving this demand is Co-Packaged Optics (CPO) — moving optical interfaces physically closer to GPU silicon to overcome electrical signaling limits at scale. Because lasers are the most failure-prone and heat-sensitive components in the optoelectronic stack, they are housed in External Laser Source Pluggable (ELSFP) modules outside the chip package, connected via Polarization Maintaining (PM) fiber. PM fiber is expensive because it requires sub-micron alignment, manual epoxy assembly, and introduces yield-sensitive failure points at every connection — which is precisely why higher-power single-laser solutions reduce system cost by minimizing the number of attachments.
Value in AI infrastructure has migrated in three documented phases: compute (GPUs, 2015–2022), memory bandwidth (HBM — Samsung, SK Hynix, Micron, 2022–2025), and now interconnect (optical layer, 2025 onward). The full photonics value chain runs from materials (Corning fiber and InP substrates) through lasers (Lumentum, Coherent) to photonics chips (Ayar Labs, Lightmatter), optical modules (Innolight, Source Photonics), and networking systems (Broadcom, Cisco) — before reaching Nvidia and hyperscalers. Most of the market is still pricing AI as a chip story. The physics says it's becoming a photon story.
Photons. Physics. Infrastructure.
Suggestion but not posted
"I've watched this exact transition before in 5G — when optical fronthaul moved closer to the radio stack, the entire vendor landscape reshuffled within 18 months. CPO in AI is the same forcing function. Same physics pressure. Same narrow supplier pool. Same market mispricing in the early innings. The reshuffle has already begun."