Photonic interconnect

AI has become a data-movement machine

The industry measures AI infrastructure in FLOPs. Increasingly, the scarce resource is the ability to keep those FLOPs fed.

Copper is not disappearing. It remains cheaper, simpler and better over short distances. But each generation of AI infrastructure joins more accelerators across more racks. At the edge of that expanding machine, the electrical link becomes too lossy, too hot and too difficult to route. The marginal connection has to become optical.

Reach

At higher lane rates, copper’s useful distance collapses. The equalization required to recover the signal consumes power and adds complexity. What works inside a rack does not cleanly extend across a row of racks.

Power

Every watt spent recovering an electrical signal is a watt unavailable to compute. As cluster power approaches the limits of a building—and then a grid connection—interconnect efficiency stops being a component detail.

Density

Larger scale-up domains require more links, more faceplate bandwidth and more routing. Copper cables become bulky; pluggable optics push the limits of the switch faceplate; board traces between an ASIC and a module become an electrical problem of their own.

The result is not a universal replacement cycle. It is a moving boundary.

A hybrid machine

The future AI system is neither all-copper nor all-optical. It uses each medium where its physics are cheapest.

This distinction matters because it changes the bet. We do not need copper demand to collapse. We need the AI system to keep getting larger.

Three transitions hiding inside one story

“Co-packaged optics” is often used as shorthand for the entire optical buildout. It is only one layer of it.

01 — Copper to optical links

This is the highest-conviction transition. It happens whenever a required connection crosses the bandwidth–distance–power boundary.

NVIDIA’s Rubin Ultra NVL576 is the architectural proof: eight racks joined into one 576-GPU NVLink domain using both copper and direct optical connections. The system is hybrid because the physics are hybrid.

02 — Pluggable to co-packaged optics

In a traditional optical switch, an electrical signal travels across the board before a pluggable module converts it to light. At higher speeds, that electrical journey becomes costly.

CPO moves the optical engine next to the switching ASIC. The electrical path shrinks; bandwidth density improves; power falls. But serviceability gets harder and a failure sits closer to an expensive chip. External laser modules, redundancy and improved packaging are attempts to recover that operational flexibility.

NVIDIA says its Spectrum-X Ethernet Photonics switches are in production. That validates CPO in scale-out networking. It does not mean every optical link immediately becomes co-packaged.

03 — Optical networking to optical scale-up

Data centers have used optics for years. The new frontier is light entering the tightly coupled scale-up domain: the fabric that makes hundreds of accelerators behave more like one machine.

This is the strategically important transition. Scale-out moves jobs among machines. Scale-up changes the size of the machine itself.

What you need to believe

• Scale-up domains keep growingFrontier training and inference continue to benefit from larger, more tightly connected accelerator domains rather than fragmenting entirely into small independent systems.
• Copper’s boundary is physicalBetter cables, retimers and signaling extend copper, but do not remove the loss, reach and power trade-off at each higher lane rate.
• Optical content grows faster than unit efficiencyNew architectures may use fewer lasers per unit of bandwidth, particularly centralized continuous-wave sources. Total system bandwidth and link count must grow quickly enough for supplier value per AI system to rise anyway.
• Qualified manufacturing remains scarceCapital alone cannot instantly reproduce epitaxy, six-inch InP yield, reliable high-power lasers, advanced packaging or hyperscaler qualification.
• Suppliers retain some of the economicsNVIDIA and hyperscalers do not use multisourcing, purchase commitments and customer-funded capacity to capture all of the benefit themselves.

The first four make photonics inevitable. The fifth makes it investable.

NVIDIA validates the market—and threatens the margin

In March 2026, NVIDIA committed $2 billion each to Coherent and Lumentum, alongside purchase commitments and access to future capacity. It has since announced a separate $2 billion investment and silicon-photonics collaboration with Marvell.

This is the strongest possible demand signal short of reported revenue. The dominant buyer is underwriting the supply chain.

It is also a warning.

Demand is real

NVIDIA does not finance multiple optical suppliers unless it expects to consume their output. The roadmap has crossed from engineering interest into procurement.

Supply is constrained

High-performance lasers and InP devices require fabs, process knowledge, yield learning and lengthy qualification. The constraint cannot be solved with a purchase order alone.

The buyer wants leverage

Multiple funded suppliers reduce dependence on any one vendor. Capacity access and purchase commitments can improve utilization while limiting scarcity pricing.

The original version of this thesis treated a “locked” supply chain as proof of supplier power. The sharper interpretation is that NVIDIA is industrializing a bottleneck before the bottleneck can tax the entire platform.

The stack is not one trade

The bottleneck will move.

At one moment it may be substrate. At another it may be 200G EML yield, high-power CW lasers, packaging, connectors or installation. A robust investment should own manufacturing capability that can follow the bottleneck—not merely the component currently in shortage.

Four ways to own the boundary

It is a portfolio of qualified manufacturing platforms that can survive changes in architecture, mix and pricing.

The ramp is already in the income statement

Coherent Q3 FY26 revenue — +21% year over year; pro-forma datacenter and communications revenue grew 41%. Non-GAAP gross margin reached 39.6%. Results

Lumentum Q3 FY26 revenue — +90% year over year, with 47.9% non-GAAP gross margin and 32.2% non-GAAP operating margin. Results

Corning Q2 2026 Optical Communications sales — +32% year over year; Enterprise Networks grew 65%. Results

Fabrinet Q3 FY26 revenue — up from $872 million a year earlier, showing that assembly and pluggables have not been erased by the CPO roadmap. Results

These numbers prove demand. They do not prove the stocks are cheap.

Questions the thesis has to survive

A correct technology thesis can still be a bad trade

The previous rule—buy five to seven percent below a recent high—was price anchoring disguised as discipline.

The proper framework begins with normalized 2028–29 economics:

• Estimate datacenter revenue under bear, base and bull optical-link growth.
• Model EML, CW/CPO and VCSEL mix rather than applying one photonics growth rate.
• Normalize gross margin after current scarcity pricing fades.
• Deduct the capex, depreciation and working capital required to create capacity.
• Underwrite free cash flow and incremental return on invested capital.
• Apply a terminal multiple appropriate to the actual business—not the AI narrative.

How this breaks

• The machine stops getting largerSoftware, workload locality or distributed architectures reduce the need for large tightly coupled accelerator domains.
• Copper moves the boundary faster than expectedActive copper, retimers or a new electrical architecture preserve acceptable reach, power and density across the links expected to turn optical.
• Optical production disappointsCPO fails system-level reliability or serviceability tests, or six-inch InP and advanced packaging do not achieve planned yield.
• Capacity outruns demandCoherent, Lumentum, Sumitomo, AXT and others bring qualified capacity online faster than optical consumption grows.
• The product mix turns against the portfolioCentralized CW sources, VCSELs or another architecture reduce supplier value faster than total bandwidth grows.
• The buyer captures the rentNVIDIA and hyperscalers use multisourcing, financing and architecture control to turn scarce suppliers into low-return dedicated capacity.
• AI capex rolls overOrders and fab plans were built for a demand curve that no longer exists.
• The price assumes none of the aboveThe most common failure is not getting the future wrong. It is paying as though the future is already certain.

Watch the boundary, the mix and the cash

Different substrates win different primitives

Silicon is still the computational substrate. Photonics does not need to replace it.

Light wins a narrower primitive: moving information across distance at high bandwidth.

That is precisely why optical interconnect is investable before photonic computing. The system already needs the primitive. The conversion cost can be paid at the boundary, while mature silicon continues doing what it does best on either side.

This is the filter for the next computing substrates:

• Does the physical system expose a primitive at radically lower cost?
• Is that primitive already expensive enough to matter at system level?
• Can it be inserted without replacing the entire stack?
• Does the advantage survive conversion, packaging, control and software?
• Can a company capture the savings before incumbent compute adapts?

Photonic interconnect passes the first four. The portfolio is a bet on the fifth.

Honest summary

Light is becoming the structural medium for the marginal AI link. NVIDIA’s product roadmap, procurement and capital allocation make that increasingly difficult to dispute.

But inevitability at the architecture layer does not guarantee monopoly economics at the component layer.

Coherent is the broad platform. Lumentum is the concentrated scarcity trade. Corning is the architecture-diversified connectivity play. Fabrinet is the hedge on a slower and messier transition.

The thesis should become more confident about the physics and less romantic about the supply chain.