Light speed: optical networking as AI's next frontier

The shift in four numbers

Why networking has become the critical variable

The first phase of the AI infrastructure build-out was defined by the race for accelerator chips. The next phase will be defined by the race to connect them. Every GPU requires connections to function in a cluster — and as AI systems scale from thousands to hundreds of thousands to millions of accelerators, the networking layer becomes as critical as the compute layer itself.

A single AI accelerator operating in isolation delivers a fraction of its potential value. Networking is what transforms individual chips into cooperative computing systems — enabling the seamless data exchange, low latency, and high bandwidth that large model training and inference require. Without it, compute investment is stranded.

The investment community has been slower to focus on this layer than on chips and data-centre construction. That is beginning to change. The specifications of next-generation AI server platforms — from current Blackwell-class systems through the coming Rubin and Rubin Ultra generations — imply networking dollar content per computing unit growing by an order of magnitude or more. This is not a marginal upgrade. It is a structural shift in where capital flows within the AI infrastructure stack.

The AI build-out spent its first phase arguing about who makes the best chip. Its next phase will be decided by who moves data fastest, at the lowest power, at the largest scale.

Critically, this growth is broad-based — it does not depend on any single technology configuration prevailing. Whether clusters scale primarily through scale-out switching or scale-up supernode architectures, whether co-packaged optics or pluggable modules dominate, all configurations are expected to benefit from strong structural demand for higher bandwidth, lower latency, and better power efficiency. The question for investors is not whether optical networking grows, but where within it value concentrates.

Understanding the two dimensions of AI networking

AI cluster networking operates across two distinct dimensions, each with its own technology requirements, component mix, and growth trajectory. Understanding the distinction is essential to mapping the investment opportunity.

The key insight is that scale-up networking — historically dominated by copper and largely overlooked as an investment theme — is undergoing a complete architectural transformation. As GPUs migrate from single-rack configurations to multi-rack supernode designs with hundreds of accelerators operating as a single logical unit, the networking requirements within those configurations become dramatically more demanding. This is the origin of the 29× dollar-content increase projected from the current generation to Rubin Ultra.

Scale-out networking, meanwhile, continues to evolve in its own right — with attach rates per GPU roughly doubling between generations and data rates migrating from 800G to 1.6T to 3.2T. Even where scale-out appears to be a "known" market, the magnitude of content growth per computing unit is substantially underappreciated.

Four generations, one direction: networking content compounds

The trajectory from current-generation platforms to Rubin Ultra represents one of the most dramatic step-changes in infrastructure component demand in the history of the technology industry. Each generation raises the bandwidth ceiling, expands the optical footprint, and introduces new connection modalities.

The connection-media hierarchy: PCB, copper, and optics

AI data-centre connections are not a single technology — they are a hierarchy of solutions optimised for different distances, speeds, and cost points. Understanding where each medium excels and where it fails determines how the networking stack evolves across generations.

The decisive trend is that optics are migrating inward — from long-distance scale-out connections toward the shorter, denser, higher-bandwidth environment of scale-up networking. This migration is what drives the disproportionate growth in networking dollar content across generations. It is not that scale-out optics contracts; it is that scale-up optics is a category that barely existed in the previous generation and is becoming one of the largest line items in the AI infrastructure bill of materials.

Co-packaged optics: the technology defining the next phase

Co-packaged optics — CPO — is the most consequential new technology entering commercial deployment in AI networking in 2026. It places optical engines as close as physically possible to the chip, shortening electrical signal paths from several centimetres to millimetre scale and fundamentally changing the power, latency, and bandwidth characteristics of the connection.

Conventional pluggable optical modules sit at the edge of the switch chassis — connected via a multi-centimetre electrical path that limits bandwidth, consumes power through digital signal processors and retimers, and introduces latency. CPO moves the optical engine to the package substrate, directly adjacent to the switch ASIC or accelerator chip. The shorter path eliminates conversion stages, cuts power consumption, and enables the high-bandwidth, short-distance connections that next-generation AI clusters demand.

The economic case for CPO is most compelling at high bandwidth densities where pluggable modules reach their physical limits. At 6.4T and 12.8T speeds, the constraints of pluggable optical form factors — power, size, heat — become prohibitive, making CPO not merely preferable but practically necessary.

CPO does carry real trade-offs. It requires supply-chain migration at the system level rather than simple module replacement. Maintenance complexity is higher: while a failed pluggable module can be swapped in minutes without disrupting the switch, a failed CPO optical engine in a switch-integrated design may require replacing the switch PCB. Reliability engineering for CPO systems is consequently more demanding, and the coexistence of pluggable modules and CPO is expected to persist across all major use cases for the foreseeable future.

Silicon photonics: displacing the incumbent technology

Within pluggable optical modules — which remain the dominant form factor for scale-out AI networking and will continue to co-exist with CPO — a technology transition is under way from electro-absorption-modulated lasers (EML) to silicon photonics (SiPh). This shift has significant implications for bill-of-materials costs, gross margins, and supply-chain positioning.

Silicon photonics offers three structural advantages over traditional EML-based modules: higher integration density and smaller physical footprint; lower power consumption through reduced component count; and lower manufacturing cost at scale. These advantages compound as data rates rise — silicon photonics offers a 32% bill-of-materials cost advantage and a 20% price advantage at 1.6T versus EML equivalents.

For suppliers, the silicon photonics transition is margin-positive. As the product mix shifts toward higher-speed SiPh-based modules — benefiting from lower laser input costs — optical module supplier gross margins are projected to expand toward 48–55%, driven by this mix migration.

Optical circuit switches: the all-optical frontier

Optical circuit switches — OCS — represent the most architecturally disruptive technology in the AI networking stack. Unlike conventional switches that require optical-electrical-optical signal conversion at every hop, OCS operates entirely in the optical domain, routing light signals directly from input to output fibre without any electronic conversion.

The strategic appeal for AI data centres is compelling. An OCS switch passes 800G, 1.6T, and 3.2T signals identically — it does not "see" the data rate, only the light. This means an AI cluster can upgrade its compute platform without replacing its switching infrastructure. In an environment where data rates double every 12–18 months, the ability to field-upgrade clusters without switch replacement is operationally and economically significant.

OCS commercial momentum is building rapidly. Leading optical component suppliers have disclosed OCS order backlogs in the hundreds of millions of dollars as of early 2026, with engagement across more than ten hyperscale customers apiece. The technology has already been deployed at scale in leading hyperscaler AI supercomputer platforms. The industry consortium driving OCS standardisation includes most major AI platform vendors and optical component companies.

Speed migration: 800G → 1.6T → 3.2T

Data-rate migration is the engine of blended average selling price growth in optical modules. As higher-speed modules displace lower-speed ones in deployment, the revenue pool expands even without volume growth — and volume is growing rapidly as well.

The 1.6T migration is the critical transition for the 2026–2027 period — moving from mainstream 800G to a generation that doubles per-port bandwidth. The Vera Rubin platform doubles the optical module attach ratio per GPU from 1:2–3 to 1:4–6, compounding the volume effect of the speed upgrade. 3.2T, while still in early verification stages in 2026, will be the standard for Rubin Ultra scale-out connections in 2027–2028.

This migration cycle is expected to be lengthened by China cloud infrastructure deployments, which tend to lag Western hyperscaler platforms by 12–18 months but represent substantial incremental demand for each speed tier as they upgrade.

Light-source supply: a structural bottleneck through 2027

The most acute near-term constraint in the optical networking supply chain is not the optical module itself — it is the laser light source that powers it. Supply of both EML and continuous-wave (CW) laser components remains very tight in 2026 and is expected to persist through 2027 before reaching a more balanced state in 2H 2028.

The supply constraint is driven by three reinforcing factors. First, InP (indium phosphide) substrate — the foundational material for both EML and CW lasers, used in photo-detectors (receivers) and laser diodes (transmitters) — has limited production capacity globally and is subject to geopolitical tension. Second, qualification and ramp-up timelines for new capacity are measured in years, not months. Third, demand is simultaneously increasing across multiple dimensions: more ports per rack, higher data rates per port, and more racks deployed.

The practical implication for investors is that the extended period of supply tightness creates a structural pricing advantage for laser and optical-engine suppliers that cannot easily or quickly be competed away. This is the "buy out the store before the storm" dynamic characteristic of AI supply chains — and laser components are one of the last major items in the stack where that dynamic remains fully operative.

TAM trajectory: from $15bn to $154bn

Networking is no longer the supporting act