Lasers, Not GPUs: The Real AI Data Center Bottleneck

For the last two years, every data center capacity conversation has started with the same question: can you get GPUs, and can you get power? A third constraint is now forcing its way into that conversation, and it is one most buyers have not modeled at all. The optical transceivers and laser components that physically link GPUs, switches, and racks together inside an AI cluster are now the tightest link in the supply chain — tighter, by some measures, than the chips themselves.

TrendForce estimates the global AI optical transceiver market will grow from $16.5 billion in 2025 to $26 billion in 2026, a jump of more than 57% in a single year, as shipments of 800G-and-faster modules climb from roughly 24 million units to nearly 63 million. Demand is rising that fast because it has to: hyperscale data center traffic in North America has been growing more than 30% a year, and every additional GPU added to a cluster needs several optical links to talk to its neighbors. The problem is that laser production hasn’t scaled anywhere near as fast as GPU production has. LightCounting estimated in August 2026 that demand for high-speed optical components was outrunning supply by roughly 30%, and Lumentum’s own CEO has said publicly that the shortage in indium phosphide (InP) lasers has now surpassed the DRAM and NAND shortages that rattled the server market earlier in this buildout.

Why lasers are harder to scale than chips

The bottleneck traces back to a narrow, specialized part of the supply chain: indium phosphide laser fabrication. Unlike silicon, InP epitaxial growth is done by a small handful of specialty manufacturers with lower throughput and yield, and it can’t simply be subcontracted to a new foundry the way digital logic can. Roughly five companies — Lumentum, Coherent, Mitsubishi, Sumitomo, and Broadcom — account for essentially all of the EML (electro-absorption modulated laser) and continuous-wave laser diode supply that 800G and 1.6T transceivers depend on. Analysts have compared the situation directly to the CoWoS advanced-packaging crunch that constrained Nvidia’s H100 ramp in 2023 and 2024: a single specialized process step, controlled by very few suppliers, gating an entire industry’s build-out.

The physics is unforgiving, too. Copper interconnects become impractical well before 10 meters at today’s speeds — Nvidia’s own 800G twin-port OSFP family tops out around 3 meters of passive copper and 4 to 5 meters of active copper, and vendors like Credo stretch active copper cables to about 7 meters at best. Beyond that, every server-to-switch and switch-to-switch link in a modern AI cluster has to be optical. As Nvidia’s NVLink scale-up domains grow — from 72 GPUs in Blackwell-generation systems toward larger Rubin-generation configurations — and as clusters scale out to tens of thousands of GPUs, the number of optical links required multiplies into the hundreds of thousands per large deployment, and every one of them draws from the same constrained InP laser supply. For buyers who have already read our breakdown of the GPU and memory supply crunch, this is the next link in that same chain, not a separate problem.

Nvidia is buying the bottleneck, not just the chips

Nvidia’s response tells you how serious the company thinks this is. On March 2, 2026, Nvidia announced it would invest $2 billion each — $4 billion total — into Coherent Corp. and Lumentum Holdings, alongside multibillion-dollar, multiyear purchase commitments, to fund U.S.-based laser and optical-networking manufacturing capacity. Jensen Huang framed it as advancing “the world’s most sophisticated silicon photonics,” but the more revealing detail is what the suppliers themselves said: Coherent disclosed in February 2026 that most of its calendar 2026 capacity was already booked, 2027 was filling quickly, and its data-center book-to-bill ratio had climbed above 4x. Coherent says it’s about 80% of the way to doubling its internal InP wafer capacity by the fourth quarter of 2026 — progress, but not fast enough to close a gap that size overnight.

Buyers are already feeling it downstream. Fusion Worldwide, a secondary-market electronics broker, reported in September 2026 that lead times on some Nvidia-specification transceiver SKUs had stretched past 50 weeks. Large buyers are responding by placing orders far ahead of need and paying substantial deposits to hold allocation; smaller buyers are getting quote-specific, often longer, lead times and rising prices as a result. Hyperscalers are also hedging by widening their supplier pools from two or three module makers to five or seven, spreading demand across second-tier vendors rather than depending on a single source.

The power and lock-in angle buyers can’t ignore

Optics isn’t just a timeline risk — it’s a power and architecture decision, too. A single modern AI rack can draw up to 600 kW, and optical networking alone can account for roughly 10% of that draw. That’s one reason Nvidia is pushing co-packaged optics (CPO) platforms like Quantum-X and Spectrum-X Photonics, which the company says can cut optical power consumption by up to 3.5x and improve resiliency by up to 10x compared with pluggable modules. The catch for buyers is that CPO, at least as Nvidia is currently positioning it, risks locking a cluster design into a single vendor’s photonics roadmap. Meta and Microsoft have both pushed publicly for an open, multi-vendor pluggable CPO ecosystem instead, precisely to avoid the kind of single-source dependency that’s causing today’s lead-time pain in the first place. There’s also a geopolitical layer underneath all of it: China controls a significant share of the world’s raw indium supply, and export restrictions on related materials are already adding pressure to an already tight laser market — a constraint that new U.S. manufacturing investment alone won’t fully resolve in the near term.

What this means for anyone planning capacity right now

None of this is a reason to delay an AI cluster build, but it is a reason to change how the build is planned. Optical transceivers and the lasers inside them now belong on the same critical-path schedule as GPUs and power equipment — not an afterthought ordered once racks and switches arrive. A cluster with GPUs, power, and cooling ready but no transceivers in hand simply doesn’t go live; as one industry sourcing guide bluntly put it, a cluster without cables cannot be accepted. Buyers negotiating capacity contracts should ask providers directly which optical suppliers and speed grades their build depends on, how much of 2026 and 2027 capacity is already locked in with long-term agreements, and what the fallback plan is if a shipment is delayed or a module fails burn-in testing and has to re-enter the same order queue. Where the network design allows it, specifying 400G instead of the newest 1.6T modules, or using copper for short in-row runs under roughly 5 to 7 meters, can meaningfully reduce exposure to the tightest part of the market this year. And because pricing power has shifted decisively toward the module makers, buyers should expect optics to show up as a larger, less negotiable line item in any new capacity deal signed over the next several quarters — budgeting for it accordingly, rather than discovering it late, is the difference between a cluster that lights up on schedule and one that doesn’t.

The GPU shortage taught buyers to secure chip allocation early. The lesson for 2026 and 2027 is that the same discipline now has to extend to the lasers and optics that connect those chips — because a rack full of GPUs with no way to talk to each other is not a working data center. For a structured way to vet a provider’s supply commitments, power position, and contract terms before you sign, start with the TechInfraHub Data Center Buyer’s Toolkit.

Written by

Raajeev Ratra

Data Center Infrastructure Expert | 15+ Years in DC Design, Operations & Project Management

Raajeev is a seasoned data center professional with hands-on experience in hyperscale facilities, colocation design, power & cooling infrastructure, and global DC operations. He shares practical insights to help engineers and IT leaders build better infrastructure.

Connect on LinkedIn →