NVIDIA & Coherent Expand 6-Inch InP Fab, Locking AI Optical Interconnect Supply Chain
Summary
Key Takeaways
Coherent's expanded 6-inch indium phosphide (InP) fab in Sherman, Texas—the world's first at this scale—will mass-produce lasers, optical components, and pluggable modules for AI interconnects. NVIDIA CEO Jensen Huang explained that at Vera Rubin Ultra NVL576 scale (576 GPUs across eight racks), copper traces cannot carry signals efficiently due to reach and power constraints; silicon photonics is the only viable solution, with optical links paying a one-time conversion penalty but offering near-free distance thereafter. NVIDIA is investing $2B in Coherent for R&D and capacity, alongside multi-billion purchase commitments. The fab will create 550+ direct jobs, backed by ~$50M from the CHIPS Act and ~$17M in Texas state incentives. Huang noted that compound semiconductors like InP and GaAs are critical for AI networking but have historically thin domestic supply chains—this expansion aims to close that gap.
Why It Matters
NVIDIA's move is a defensive play against AMD and Intel's AI cluster interconnects. By injecting $2B into Coherent and locking multi-billion purchases, NVIDIA effectively privatizes the silicon photonics supply chain—any NVIDIA GPU cluster must use Coherent optics, creating hardware-level vendor lock-in. Competitors will struggle to source equivalent InP-based optics; global 6-inch InP capacity is currently unique to Coherent. The press release downplays risks: InP wafer yield and ramp-up are far slower than silicon CMOS; capacity may bottleneck NVIDIA's NVL576 production timeline. Also, tail latency in optical modules can degrade with temperature and aging, and NVIDIA hasn't disclosed PFC/ECN congestion control adaptation for optical links. Enterprises face cross-generation optical module incompatibility—NVIDIA could change wavelengths or modulation in future architectures, forcing full network replacement.
PRO Decision
【Vendors: AMD, Intel, Broadcom】 Accelerate open optical interconnect ecosystems. AMD should partner with Lumentum or Marvell to develop standards-based silicon photonics modules (e.g., OIF 224G/Lane) and push UALink to support optical media, breaking NVIDIA's Coherent monopoly. Intel can leverage its silicon photonics R&D and IFS foundry to offer integrated optical engines compatible with PCIe/CXL. 【Enterprises: CIOs and Architects】 Perform zero-trust audits: demand TCO models for optical modules (power, maintenance, lifespan) and verify tail latency and congestion control performance. Include optical interoperability clauses in contracts to allow future replacement with standardized modules (QSFP-DD/OSFP). Evaluate hybrid copper+optical architectures to avoid single-point dependency. 【Investors】 Recognize this as a move to turn optical interconnects from commodity into proprietary lock-in. Short-term benefits Coherent, but long-term, the lock-in may spur rivals (AMD+UALink) to accelerate open alternatives, eroding NVIDIA's interconnect dominance. Monitor Coherent's ramp-up risk and InP yield; if below expectations, NVIDIA's GPU shipments will be constrained.
Get 3-5 key AI infrastructure signals weekly →
💬 Comments (0)