← Back to feed

Coherent has broken ground on an expanded manufacturing facility in Sherman, Texas, targeting the production of indium phosphide wafers used in AI data center interconnects. The expansion follows NVIDIA’s $2 billion investment in the company in March 2026 — a deal structured to secure photonic component supply, deepen joint R&D, and advance US-based manufacturing of a material most AI infrastructure discussions skip past entirely.

The Sherman facility will be the world’s first 6-inch InP fabrication plant. At project completion, it is expected to create more than 1,000 jobs, including 550+ direct advanced manufacturing, engineering, and technical roles.

Why Photonics Matter Now

The bottleneck in large-scale AI clusters is shifting. Training a frontier model on 100,000+ GPUs is now routine. At that scale, the constraint is not compute throughput inside the GPU — it is bandwidth between GPUs, between server racks, and between data centers.

Copper interconnects carry data at rates set by electron physics. As rack densities increase — NVIDIA’s Vera Rubin NVL72 configurations pack 72 GPUs per rack — the heat load and bandwidth requirements of copper become a hard ceiling. InP-based photonic components replace copper with laser-driven optical signals, moving data at light speed with dramatically lower power per bit.

Coherent makes the lasers, optical components, and compound semiconductors that implement this layer. Its customers include the major hyperscalers building the GPU clusters that AI training runs on. NVIDIA is both an investor and, indirectly, a customer — its GB300 and Vera Rubin rack architectures depend on photonic interconnects for the bandwidth claims in their spec sheets.

The Investment Logic

NVIDIA’s $2 billion into Coherent is not a passive equity position. The press language at the time covered three specific objectives: secure supply, advance R&D, and build US manufacturing capacity. All three are defensive.

AI semiconductor supply chains have a documented vulnerability at the photonic layer. InP wafer production has been dominated by a small number of facilities, most outside the US. A supply disruption in photonic components constrains AI cluster deployment in ways that cannot be worked around by ordering more Nvidia chips. The chips need the interconnects to function at claimed performance.

The Sherman expansion addresses that directly. US-based InP production reduces geographic concentration risk, satisfies government procurement rules for domestic content, and gives NVIDIA a contractual claim on capacity that does not depend on import logistics.

The Compute Chain

The photonic layer is the piece of the AI infrastructure buildout that does not appear in most capital allocation discussions. The headline numbers go to GPU orders (NVIDIA, AMD), hyperscaler capex (CoreWeave $8.5B credit facility, Fluidstack $1B raise), and compute real estate (Colossus, Fairwater, Stargate). The fiber and lasers that carry data between those components are assumed rather than funded.

The Coherent deal makes the assumption explicit. At Vera Rubin NVL72 densities, with 131,000-GPU clusters like Stargate operating in production, the interconnect layer becomes a supply chain bottleneck with the same strategic weight as HBM or advanced packaging. NVIDIA’s move is an acknowledgment of that — made before the bottleneck became visible in delivery schedules.

Numbers

• NVIDIA investment: $2 billion (March 2026)
• Facility location: Sherman, Texas (same semiconductor corridor as Samsung Austin and TI)
• Process: World’s first 6-inch InP wafer fabrication (prior standard: 4-inch)
• Jobs: 1,000+ total, 550+ direct technical roles
• Application: Optical transceivers and laser arrays for AI cluster interconnects
• Timeline: Construction in progress; no public completion date