NVIDIA announced a $4 billion investment in photonics suppliers Lumentum and Coherent on March 2, 2026, signaling that copper interconnects are hitting their physical limits in AI datacenters. The chipmaker committed $2 billion to each company through multiyear agreements including purchase guarantees and capacity rights for optical components. With plans to scale GPU clusters from 72 to 576 GPUs per system by 2027, NVIDIA needs interconnect technology that can handle the bandwidth—and copper can’t deliver anymore.
The Deal: Securing the Photonics Supply Chain
NVIDIA’s investment isn’t just capital—it’s a supply chain play. The agreements include multibillion-dollar purchase commitments and guaranteed access to advanced laser and optical networking products. Both companies will expand U.S.-based R&D facilities and manufacturing capacity, reducing geopolitical risk. The market validated the move immediately: Lumentum shares jumped 12% and Coherent surged 15%.
Splitting the investment between two vendors is strategic risk management. NVIDIA gets competition on pricing and innovation while avoiding single-supplier dependence.
Why Copper Is Done at AI Scale
The physics are unforgiving. As bandwidth demands approach terabit-per-second speeds, copper interconnects must become shorter and thicker to maintain signal integrity. At SerDes speeds now entering 200G and 400G territory, copper is hitting fundamental physical limits. High-speed copper connections are restricted to under 2 meters, confining them within single racks.
NVIDIA’s GPU scaling plans make this critical. The company targets an 8x increase in maximum GPUs per system—from 72 to 576 by 2027—requiring multi-terabit data movement between accelerators. Copper creates ultra-dense racks consuming hundreds of kilowatts with punishing heat loads.
This isn’t a future problem. It’s a 2027 problem, which in product development timelines means it’s a today problem. NVIDIA is securing photonics supply chains now because waiting would risk the Rubin Ultra launch.
Enter Silicon Photonics
Silicon photonics uses light instead of electricity to transmit data, integrating optical components onto silicon chips using standard semiconductor manufacturing. The advantages over copper are substantial: 800G optical modules ship today, 1.6T products arrive in 2026, and multi-terabit speeds are on roadmaps. Power consumption drops to 1-2 picojoules per bit. Latency improves as electrical path lengths shrink from centimeters to millimeters in co-packaged optics designs.
Co-packaged optics (CPO) is the specific technology NVIDIA is betting on. Instead of pluggable optical transceivers, CPO integrates optical components directly onto switch and accelerator packages. Industry analysts expect CPO to become mandatory for next-generation AI datacenters, not an optional upgrade.
Lumentum and Coherent Bring Production Capability
Lumentum manufactures lasers, transceivers, and photonic components with a datacenter portfolio including 100G and 200G externally modulated lasers and building blocks for 800G and 1.6T solutions. Coherent brings silicon photonics expertise with their 2x400G-FR4 Lite transceiver and 800G-DR8 modules, showing strong AI network demand in Q1 fiscal 2026. NVIDIA’s official announcement emphasized their production readiness and ability to scale manufacturing quickly.
The Timeline That Explains Everything
CPO deployments start in 2026-2027, with high-volume production expected by 2028. NVIDIA’s Rubin Ultra GPU architecture launches in 2027. The connection is clear: NVIDIA needs the photonics supply chain mature and scaled before their next major platform ships. This $4 billion ensures critical components are available in volume when needed.
The co-packaged optics market is projected to exceed $20 billion by 2036, growing at 37% annually. Major cloud providers invest tens of billions annually in datacenter infrastructure and are actively deploying CPO technology.
Copper’s Decline, Photonics’ Rise
This move signals where datacenter bottlenecks are shifting. Compute performance keeps scaling with new GPU architectures, but interconnect bandwidth was becoming the limiting factor. Photonics removes that constraint for the next several technology generations. Copper remains dominant for shorter, lower-bandwidth connections, but its role in high-performance AI infrastructure is ending.
Other hyperscalers face the same physics problems at scale. NVIDIA securing capacity rights with Lumentum and Coherent means AWS, Google Cloud, and Microsoft Azure may need similar strategic investments or accept longer lead times and higher prices. First-mover advantage in supply chain positioning matters when component demand is about to explode.
NVIDIA just made a $4 billion bet that light beats electricity for moving data in AI datacenters. Given copper’s physical limits and the industry’s deployment timeline, it’s less a bet and more an inevitability.
