Copper cables are physically melting at 224 Gbps per lane. This is not a metaphor. At terabit-per-second speeds, a phenomenon called the skin effect forces electrical current to the surface of the wire, generating heat so intense that cables degrade and fail. The physical medium that once stitched together data centers is now the single biggest obstacle to building larger AI training clusters. Silicon can go faster. Copper cannot.
On June 16, 2026, in Sherman, Texas, NVIDIA and Coherent Corp. broke ground on the answer: the world's first volume-production 6-inch indium phosphide semiconductor fabrication facility. The expansion will double manufacturing space and quadruple wafer production capacity, creating more than 1,000 jobs, including over 550 direct advanced manufacturing, engineering, and technical roles. It is backed by a proposed $50 million CHIPS Act grant from the U.S. Department of Commerce and a $2 billion strategic investment from NVIDIA, announced on March 2, 2026.
The Physics Problem No One Can Code Around
Copper traces carry electrical signals by pushing electrons through a conductor. At the speeds required to feed modern AI training clusters, the skin effect forces current to flow only on the surface of the wire. The result is a hard physical limit: push more data, generate more heat. The only workaround is thicker, shorter cables that consume more power. That workaround is now exhausted.
Optical interconnects built on indium phosphide do not have this problem. Photons do not generate resistive heat. They carry more data over longer distances using less energy. The bottleneck is not physics. It is manufacturing capacity. Until Sherman, no facility on earth produced 6-inch indium phosphide wafers at volume. That is what just changed.
"Optical interconnects, not GPU silicon, are now the binding constraint on how large an AI system can grow," according to TechTimes. The statement is precise. A single NVIDIA system can already span 576 GPUs across eight racks. The compute exists. The pathway between the compute does not.
A $4 Billion Bet That the Component Model Is Dead
NVIDIA did not place a purchase order. It placed a bet on control. The Coherent agreement is a multiyear, non-exclusive deal that includes a multibillion-dollar purchase commitment and future access and capacity rights for advanced laser and optical networking products, per the SEC filing. In parallel, NVIDIA invested another $2 billion in Lumentum Holdings, as reported by Futurum Group. That is a $4 billion optical procurement strategy executed in a single quarter.
The mechanism is capacity reservation at a scale the industry has never seen. NVIDIA is locking in the indium phosphide supply chain before the next generation of AI clusters demands it. The math is straightforward. If your GPU roadmap shows a 4x increase in interconnect bandwidth by 2027, and the only material that can carry that signal is produced by two companies, you buy capacity rights from both of them or you stop shipping systems.
"With Coherent, NVIDIA is pioneering next-generation silicon photonics to enable AI infrastructure at unprecedented scale, speed and energy efficiency," Jensen Huang said at the groundbreaking. Coherent CEO Jim Anderson described the partnership as an expansion of a 20-year relationship. "Semiconductor photonic devices are essential building blocks of AI infrastructure," Anderson said, "enabling the high-speed connectivity required to move unprecedented amounts of data between processors, memory, and systems."
The Hyperscalers Are Now Captive Buyers. They Will Not Stay That Way.
Here is what the consensus is missing. The Coherent/NVIDIA model proves that controlling the photonics supply chain is now as critical as owning the GPU fab. The implication is not subtle. Within 12 to 24 months, at least one major hyperscaler will announce its own indium phosphide optical interconnect partnership or acquisition.
Microsoft, Google, and Amazon each operate AI training clusters that are approaching the same copper ceiling. They currently buy optical interconnects as components from a market that now has one dominant buyer with capacity rights. That is an untenable position. If your primary GPU supplier also controls the allocation of the only interconnect material that can scale to your next-generation cluster, you are no longer a customer. You are a dependent.
The capital expenditure math forces action. A hyperscaler spending $50 billion annually on infrastructure cannot accept a single-source constraint on the component that determines cluster size. The Sherman deal signals that the component model—where optical interconnects are bought off the shelf like power supplies—is over. The alternative is vertical integration. Expect at least one of the big three to acquire or partner with a photonics fabricator within two years. The only question is whether they pay a premium now or a larger premium later.
Lumentum faces the mirror image of this pressure. The company holds a parallel $2 billion NVIDIA investment and must match Coherent's capacity expansion or risk losing share in NVIDIA's optical procurement. The CHIPS Act money flowing to Sherman is a signal that the U.S. government will subsidize photonics manufacturing at the same priority level as leading-edge logic. Lumentum either scales its own InP production or cedes the market. The roughly $50 billion CHIPS Act program was designed to bring silicon chip manufacturing back to U.S. soil. The Sherman grant redirects those funds toward photonics, a material class that was an afterthought in the original legislation. The reindustrialization argument just shifted from "make chips in America" to "make the optical backbone of AI in America." NVIDIA has stated a commitment to produce up to $500 billion of AI infrastructure in the U.S. through industry partnerships, according to the NVIDIA blog. Sherman is the first concrete proof point.
What the Infrastructure Operator Needs to Hear
If you are building or procuring AI infrastructure, the $2 billion bet is a leading indicator. Optical interconnects will be the next constrained resource, and the constraint will tighten before it loosens. The Sherman facility is a groundbreaking, not a ribbon-cutting. Volume production takes time. The 2027 terabit-scale cluster wave will arrive before the full capacity comes online.
Plan procurement accordingly. The companies that secure optical interconnect supply in 2026 will be the ones training frontier models in 2028. Everyone else will be waiting on allocation.
The Material That Can Handle the Heat
"Coherent is a world-class company, and the work you do is vital to our future, vital to the future of artificial intelligence and vital to reindustrializing the United States," Huang said at the ceremony.
The copper cables melting in data centers today are the reason Sherman, Texas, is now the site of the world's first volume indium phosphide fab. The $2 billion is not a bet on a company. It is a bet on a material that does not fail under the heat of the next million GPUs.
The future of AI infrastructure is not written in silicon. It is written in light.
