NVIDIA and Coherent Scale InP Optical Interconnects for AI Rack-Scale Photonic Fabric
Summary
Key Takeaways
NVIDIA and Coherent break ground on a 6-inch Indium Phosphide (InP) wafer fab in Sherman, Texas, to produce high-speed lasers, transceivers, and pluggable optical modules for AI datacenter rack-to-rack interconnects. Huang noted that in the Vera Rubin Ultra NVL576 configuration, 576 GPUs across 8 NVLink racks require silicon photonics; copper is insufficient. Optical links have a one-time electro-optical conversion cost but save power over longer distances. Coherent received $50M from CHIPS Act and ~$17M from state/local incentives. The fab will create 550+ direct jobs. InP is key for emitting and modulating light in AI connectivity.
Why It Matters
NVIDIA's move is a defensive play to encircle AMD/Intel's UALink and OCP initiatives by locking the optical supply chain. By tying Coherent's InP capacity, NVIDIA creates vendor lock-in: any NVLink rack must use its certified optics, or performance consistency is lost.
The article omits physical limitations: InP wafer yield and ramp-up uncertainty; can a 6-inch fab serve millions of GPUs? Optical tail latency in large HPC clusters degrades due to temperature drift and connector contamination. NVIDIA's centralized NVSwitch may introduce head-of-line blocking. Cost trap: pluggable optics have 3-5 year lifespan vs. copper's 10+ years, raising TCO.
PRO Decision
【Vendors】AMD/Intel should accelerate UALink open standard, partnering with Broadcom and Marvell to create multi-vendor InP/silicon photonic interconnects, breaking NVIDIA's NVLink optical certification lock. Define common mechanical/electrical interfaces in OCP to reduce dependency.
【Enterprises】CIOs must perform zero-trust audits on Vera Rubin clusters: demand independent lifetime and tail latency reports for optical modules, and penalty clauses for non-certified optics. Evaluate UALink or InfiniBand alternatives to preserve architectural flexibility.
【Investors】This PR move solidifies NVIDIA's AI moat but InP capacity constraints and optics depreciation will pressure margins. Monitor Coherent's ramp speed and whether NVIDIA is forced to open its interconnect. Beware of Broadcom/Marvell being marginalized.
Get 3-5 key AI infrastructure signals weekly →
💬 Comments (0)