Silicon Photonics: The Quiet Infrastructure Layer Behind the AI Boom
Behind every headline-grabbing AI model launch sits a far less visible but equally consequential infrastructure challenge: how to move staggering volumes of data between GPUs, switches, and servers fast enough, and without burning unsustainable amounts of power, to keep AI training clusters running. Silicon photonics — the technology of integrating optical functions such as modulation, routing, and detection directly onto CMOS-compatible silicon substrates — has emerged as the leading answer to this challenge, and 2026 is the year it has moved decisively from promising technology to mission-critical infrastructure. Conventional electrical and pluggable optical interconnects are increasingly unable to keep pace with the bandwidth density that AI training clusters demand, and the industry-wide pivot toward co-packaged optics, on-chip lasers, and tighter integration between optical engines and switching silicon reflects an unmistakable consensus: light, not copper, is the only viable path forward for next-generation AI data centre interconnects.
NVIDIA's Strategic Bet: Why the World's Most Valuable Chip Company Is Investing in Optics
Perhaps no single development better illustrates silicon photonics’ new strategic centrality than NVIDIA’s direct moves into the optics supply chain. In March 2026, NVIDIA announced a multiyear strategic agreement with Lumentum Holdings to accelerate innovation in advanced optics technologies — spanning joint research and development specifically intended to enable next-generation AI infrastructure and systems designs. The rationale is rooted in hard physics and economics: NVIDIA’s own published data shows that co-packaged optics deliver 3.5 times better power efficiency than legacy pluggable transceivers — a margin of improvement that matters enormously at the scale of modern AI training clusters, where power delivery and thermal management, not raw silicon availability, are increasingly the binding constraint on how large a cluster can practically be built. With NVIDIA’s Rubin platform advancing the frontier of AI training hardware, the company’s direct strategic engagement with the optics supply chain signals that interconnect technology has become as commercially consequential to NVIDIA’s roadmap as the GPU silicon itself.
Lumentum's Breakout Year: A Single-Customer Purchase Commitment Like No Other
Lumentum Holdings has emerged as one of the silicon photonics ecosystem’s most direct commercial beneficiaries of the AI infrastructure build-out. The company’s chief executive has publicly characterised 2026 as a breakout year for laser chip sales, and Lumentum has confirmed receipt of the largest single purchase commitment for ultra-high-power co-packaged optics lasers in the company’s history. Lumentum holds a dominant position in electro-absorption modulated lasers — a component critical to 800-gigabit and 1.6-terabit optical transceivers — and with no viable second source currently available for certain laser variants, analysts widely anticipate double-digit price increases on key 200-gigabit EML components in 2026, a dynamic that flows directly through to Lumentum’s earnings. This pricing power, rare in component supply chains that typically face relentless commoditisation, reflects just how acute the supply constraint on advanced photonic components has become relative to AI infrastructure demand.
Broadcom and the Networking Silicon Powering the Optics Transition
Broadcom’s position at the centre of data centre networking silicon provides one of the clearest financial signals of how rapidly the AI-driven optics transition is scaling. The company’s AI-related semiconductor revenue reached $20 billion in fiscal year 2025, up 65% year-over-year, with an AI-related order backlog exceeding $73 billion — and the company has guided for AI revenue of approximately $8.2 billion in the first quarter of fiscal 2026 alone, double the prior year’s comparable quarter. Broadcom’s Tomahawk and Jericho switch silicon families have become the de facto industry standard for hyperscale data centre networking, and the company has been investing directly in its own optical connectivity solutions to ensure its switching silicon and the optical engines that feed it are co-designed for the bandwidth and power efficiency that co-packaged architectures demand. The scale of this AI infrastructure investment cycle is itself extraordinary: major technology companies collectively committed over $700 billion to AI infrastructure across 2025 and 2026, with networking representing roughly 15 to 20% of total data centre capital expenditure — a substantial and growing pool of capital flowing directly toward the photonics and optical interconnect ecosystem.
The Technology Roadmap: From Pluggable Modules to Co-Packaged Optics
The industry’s technology roadmap for 2026 is anchored by two parallel developments. First, high-volume qualification programmes for 800-gigabit and 1.6-terabit optical modules are now well underway across major data centre and networking equipment suppliers, representing the current-generation workhorse interconnect technology for AI training clusters. Second, and more transformatively, co-packaged optics architectures are gaining traction as the dominant form factor for next-generation network switches, with the optical engine integrated directly alongside the switching ASIC rather than connected via a separate pluggable module — an architectural shift that meaningfully reduces both latency and power consumption by minimising the physical and electrical distance data must travel between the optical and electronic domains. Industry analysis projects co-packaged optics will reach volume production in commercial AI systems within roughly 18 to 24 months from early 2026, implying broad commercial deployment by late 2027, even as conventional pluggable modules continue to coexist with CPO architectures for several more years given their advantages in flexibility, field serviceability, and ecosystem maturity.
Geopolitics Enters the Optics Supply Chain
Silicon photonics is not insulated from the broader geopolitical forces reshaping the semiconductor industry. Foundry-level standardisation efforts at IMEC, Tower Semiconductor, and TSMC are lowering development costs for fabless photonics designers and broadening the base of companies able to bring silicon photonic products to market without owning their own fabrication facilities — a democratising trend for the industry’s innovation base. Simultaneously, however, geopolitical export controls on advanced photonic design tools and epitaxial materials are actively reshaping global trade flows within the sector, with non-domestic fabless designers in several markets facing genuine supply uncertainty as a direct consequence. Constancy Researchers identifies the combination of foundry standardisation lowering entry barriers and export control regimes simultaneously raising barriers for specific geographies as one of the more structurally complex dynamics facing the silicon photonics supply chain through the remainder of the decade.
Beyond Data Centres: Automotive LiDAR and the Broader Application Frontier
While AI data centre interconnects represent the single largest and fastest-growing demand driver for silicon photonics today, the technology’s addressable market extends well beyond the data centre. The commercialisation of silicon photonic LiDAR for autonomous vehicles represents one of the most significant adjacent growth vectors, leveraging the same fundamental advantages — high bandwidth density, low power consumption, and small physical footprint — that make silicon photonics attractive for data centre interconnects. Telecom network upgrades supporting 5G Advanced deployment and early 6G development, alongside emerging sensing applications across industrial and consumer electronics, round out a diversified set of growth vectors that reduce the technology’s dependence on any single end market, even as AI infrastructure spending remains overwhelmingly the dominant near-term growth driver shaping company strategy and capital allocation across the sector.
Competitive Landscape & Key Players: A Supply Chain Built on Strategic Partnership
The silicon photonics competitive landscape is increasingly defined not by any single company’s standalone technology but by the strategic partnerships being formed across the value chain. Intel Corporation has maintained the largest standalone market share in silicon photonics, leveraging its deep CMOS fabrication expertise, while Cisco Systems, Broadcom, and Lumentum collectively hold a substantial share of the broader market, reflecting the concentration of value among companies that combine photonic component manufacturing with deep relationships across the hyperscale and networking equipment customer base. The shift toward strategic partnerships rather than vertically integrated solo development — exemplified by the NVIDIA-Lumentum collaboration — reflects a broader recognition that no single company currently possesses the full breadth of expertise spanning laser sources, photonic integration, packaging, and systems-level switch design required to deliver next-generation co-packaged optics at the pace AI infrastructure demand requires. Constancy Researchers assesses that the companies best positioned through the remainder of the decade will be those that secure durable strategic partnerships spanning the full optical engine value chain, rather than those attempting to compete on isolated point-component advantages alone.
What Does the Silicon Photonics Inflection Point Mean for the Decade Ahead?
Constancy Researchers’ assessment is that silicon photonics has crossed a decisive threshold in 2026, transitioning from a promising but commercially uncertain technology into a critical, capacity-constrained bottleneck technology for the entire AI infrastructure build-out. NVIDIA’s direct strategic investment in the optics supply chain, Lumentum’s unprecedented pricing power on key laser components, and Broadcom’s surging AI-related networking silicon revenue collectively confirm that optical interconnect technology has become as commercially and strategically consequential as the GPU silicon it serves — not a peripheral supporting technology, but a central determinant of how quickly and efficiently the next generation of AI training clusters can be built. The transition toward co-packaged optics over the next 18 to 24 months will be the defining technical and commercial milestone for the sector, and the companies that successfully navigate the manufacturing complexity, thermal management challenges, and geopolitically fragmenting supply chain associated with that transition will capture a disproportionate share of one of the most strategically important technology markets of the AI era.
Looking to navigate geopolitical risks and identify new industry opportunities?
Connect with Constancy Researchers for data-driven insights, geopolitical intelligence, and strategic consulting tailored to your business needs.
