The AI build-out conversation tends to start and stop with chips. But the companies quietly landing multi-year, multi-billion-dollar contracts right now aren't making semiconductors — they're making the glass, connectors, and routing systems that hold the entire data center ecosystem together.
The AI build-out conversation tends to start and stop with chips. But the companies quietly landing multi-year, multi-billion-dollar contracts right now aren't making semiconductors — they're making the glass, connectors, and routing systems that hold the entire data center ecosystem together.
Nvidia's $4.7 trillion market capitalization was built on GPUs, but the company's next growth vector may depend on something far less visible: the optical interconnects that link those GPUs into functional AI factories. In March, Nvidia invested $2 billion each in optical-component makers Coherent and Lumentum, and another $2 billion in Marvell Technology, a custom chip and networking specialist — a $6 billion bet that signals where the bottleneck in AI infrastructure is shifting.
"The GPU is only as fast as the network feeding it," said Jensen Huang, Nvidia's chief executive officer, during the company's GTC conference in March. "As models grow to trillions of parameters, the interconnect becomes the system."
The math explains the urgency. Nvidia's H100 GPU delivers 990 TFLOPS of FP16 performance, but that compute is wasted when data stalls between clusters. AI networking — the specialized switches, optical transceivers, and cabling that stitch GPU pods together — has become the limiting factor for hyperscale deployments. Without ultra-low latency and lossless throughput, multi-billion-dollar GPU clusters operate below capacity. Nvidia's own InfiniBand platform and its Spectrum-X Ethernet switches address this, but the company is now investing upstream to secure supply of the optical components that make those systems work.
Why Optics Matter More Than Ever
The shift from electrical to optical interconnects is accelerating as AI clusters scale beyond 100,000 GPUs. Copper cabling cannot sustain the bandwidth and distance requirements of next-generation data centers. Silicon photonics — the technology that Coherent and Lumentum specialize in — transmits data using light rather than electricity, cutting power consumption by as much as 40% while enabling speeds of 800 gigabits per second and beyond. For hyperscalers spending $50 billion or more annually on data center build-outs, those efficiency gains translate directly to the bottom line.
Nvidia's investments are not isolated bets. The company is building an end-to-end networking stack that includes its BlueField data processing units, which offload networking and security tasks from CPUs and GPUs, alongside its InfiniBand and Spectrum-X switching platforms. By integrating optical components from Coherent and Lumentum and custom accelerators from Marvell, Nvidia is positioning itself as the sole provider of complete AI factories — integrated systems that capture compute, high-speed interconnects, and photonics in a single architecture.
Who Wins, Who Loses
The optical infrastructure build-out creates winners beyond Nvidia. Coherent and Lumentum, which supply the lasers and photonic circuits that power high-speed data transmission, stand to benefit as hyperscalers — Amazon Web Services, Microsoft Azure, Google Cloud — accelerate their own networking investments. Marvell, already a leader in custom AI accelerators and Ethernet switches, gains a deeper partnership with the world's most valuable chip company.
The risk falls on traditional networking incumbents. Broadcom, which competes with Nvidia in Ethernet switching and data center connectivity, faces a rival that now controls both the compute and the interconnect layers. Cisco Systems, still dominant in enterprise networking, has less exposure to the hyperscale AI segment and could see its relevance diminish as the market consolidates around vertically integrated suppliers.
For investors, the optical theme offers a way to play the AI infrastructure build-out without paying Nvidia's multiples. Coherent trades at roughly 18 times forward earnings, Lumentum at 22 times, and Marvell at 25 times — all below Nvidia's peak valuation during the GPU frenzy. As hyperscaler capital expenditure budgets expand, with Microsoft alone committing $80 billion to data centers in fiscal 2026, the companies supplying the physical layer of AI infrastructure are capturing a growing share of that spend.
The optical interconnect market for AI data centers is projected to reach $15 billion by 2028, according to industry estimates, up from roughly $4 billion in 2025. That growth trajectory — driven by the same hyperscale demand that lifted Nvidia to a $4.7 trillion market cap — is now drawing investor attention to a part of the AI stack that was easy to overlook when the story was just about chips.
This article is for informational purposes only and does not constitute investment advice.
