Inside the Quiet Scramble to Scale Indium Phosphide Wafers for AI's Optical Future

The Physics Problem Behind the Investment

When AI training clusters push beyond a few dozen accelerators, the mathematics of signal integrity start to break down. Copper cables, the workhorse of rack-level interconnect for decades, cannot sustain the bandwidth density or reach required once systems scale into the hundreds or thousands of GPUs. Latency climbs, power consumption balloons, and physical constraints - cable thickness, bend radius, heat dissipation - become engineering roadblocks.

Optical interconnects, built around lasers, photodetectors, and modulators fabricated on indium phosphide substrates, sidestep many of these limits. Light travels farther with less loss, carries more data per wavelength, and dissipates less heat per bit moved. The challenge is no longer whether optics will replace copper at hyperscale; it is whether the supply chain can ramp fast enough to meet demand as clusters grow from experimental pods to production fleets.

Coherent, a photonics vendor that operates eight wafer fabrication facilities across the United States, is placing a $650 million bet that it can. The company announced plans to effectively double the footprint of its Sherman, Texas, plant and quadruple its output of indium phosphide wafers - the substrate material at the heart of most high-speed optical components used in data centers today.

Capital Structure and Timeline

The expansion is funded through a combination of private capital and public incentives. Coherent received a $2 billion investment from Nvidia in March, a stake that reflects the GPU designer's strategic interest in securing optical supply chains as its own roadmap tilts toward rack-scale and multi-rack AI systems. The Sherman project also draws $20 million from the Texas Semiconductor Innovation Fund and the Sherman Economic Development Corporation, plus up to $50 million from the CHIPS and Science Act, the federal program designed to reshore advanced semiconductor manufacturing.

Coherent has not disclosed a completion date for the buildout. Fab expansions of this scale typically require eighteen to thirty months from groundbreaking to volume production, depending on equipment lead times and cleanroom certification schedules. The company has indicated that roughly 1,000 new jobs will be created, with approximately 550 in advanced manufacturing, engineering, and technical roles - a signal that the plant will support both high-volume production and ongoing process development.

Why Indium Phosphide Matters

Indium phosphide is a III-V compound semiconductor prized for its direct bandgap and high electron mobility, properties that make it well-suited for optoelectronic devices operating at the wavelengths used in fiber-optic communication. Unlike silicon photonics, which integrates optical and electronic functions on a single platform but requires hybrid laser assemblies, InP can emit, modulate, and detect light natively. This makes it the material of choice for laser diodes, avalanche photodetectors, and electro-absorption modulators - components that sit inside pluggable transceivers, active optical cables, and co-packaged optics modules.

As AI workloads push interconnect requirements from 400 gigabits per second toward 800 Gbps, 1.6 Tbps, and beyond, the number of optical channels per system climbs. A single training pod with 512 accelerators might employ hundreds of transceivers; a multi-rack supercluster with 8,000 GPUs could require several thousand. Each transceiver contains multiple InP-based chips. The arithmetic is straightforward: more clusters, more optical lanes, more wafer starts.

Nvidia's Broader Optics Play

Coherent is one node in a larger strategy. At Computex earlier this month, Nvidia CEO Jensen Huang publicly described optics vendor Marvell as a candidate to become the next trillion-dollar company, a remark that underscored how central photonics has become to the GPU maker's long-term architecture. Nvidia has deployed $6 billion across three optical supply-chain investments this year alone.

In addition to the Coherent stake, Nvidia invested $2 billion in Lumentum, a manufacturer of pluggable transceivers, optical circuit switches, and laser modules used in hyperscale data centers. Less than a month later, the company committed another $2 billion to Marvell, in part to accelerate development of silicon photonics and custom interconnect ASICs. The pattern is clear: Nvidia is not waiting for the optical supply chain to catch up organically. It is funding capacity expansions, technology roadmaps, and manufacturing scale in parallel.

At DailyTechWire, we have tracked similar moves by other hyperscalers - Microsoft's partnerships with optical startups, Google's internal photonics development, Amazon's investment in co-packaged optics pilots - but Nvidia's capital deployment stands out for its speed and scale. The company appears to be treating optical interconnect as a strategic chokepoint, one that could constrain its own growth if left to market forces alone.

The Demand Curve and Its Uncertainties

The case for optical interconnect rests on two assumptions: that AI training and inference workloads will continue to scale in cluster size, and that copper will hit fundamental limits before alternative architectures - such as disaggregated memory or chiplet-based designs - reduce interconnect pressure. Both assumptions are plausible, but neither is guaranteed.

Training runs for frontier models have grown from hundreds of GPUs in 2020 to tens of thousands in 2025. If that trajectory continues, optical interconnect becomes non-negotiable. But if model architectures shift toward mixture-of-experts designs, sparse training, or other techniques that reduce all-to-all communication, the growth curve for optical components could flatten sooner than supply-chain investments anticipate.

There is also the question of packaging. Co-packaged optics, in which lasers and modulators are integrated directly onto the switch or GPU package, promise to reduce module count and power consumption. If co-packaged optics mature faster than expected, demand for pluggable transceivers - and the InP wafers inside them - could shift in composition if not in volume.

Coherent's quadrupling of wafer capacity is a hedge against the high end of the demand forecast. It reflects confidence that even if architectural changes moderate growth, the absolute volume of optical components required by AI infrastructure will remain steep enough to absorb the added supply.

Regional Context and Workforce Implications

The Sherman expansion also fits into a broader reshoring effort. The United States currently lags behind Taiwan, South Korea, and parts of Europe in advanced packaging and compound semiconductor manufacturing. The CHIPS Act, passed in 2022, was designed to reverse that imbalance by subsidizing domestic fabs and encouraging private co-investment.

Coherent's project is among the first large-scale compound semiconductor expansions to draw on CHIPS funding. The 550 technical and engineering roles the company expects to create will require a workforce trained in III-V materials science, epitaxial growth, cleanroom protocols, and high-volume optoelectronics manufacturing - skill sets that are scarce in the Texas labor market today. How Coherent recruits, trains, and retains that workforce will serve as a case study for other reshoring projects in adjacent sectors.

Sherman, a city of roughly 45,000 people about sixty miles north of Dallas, has historically been a manufacturing hub for automotive and industrial components. The Coherent expansion represents a pivot toward advanced semiconductors, with implications for local technical education, housing, and infrastructure. The $20 million in state and local funding is intended to offset some of those costs, but the transition will take years.

What Comes Next

If the Sherman buildout proceeds on schedule, Coherent will be positioned to supply a significant fraction of the InP wafers required by North American optical module vendors through 2028. Whether that capacity will be fully utilized depends on variables the company cannot control: the pace of AI cluster deployments, the maturity of competing interconnect technologies, and the broader trajectory of hyperscale capital expenditure.

For now, the market is pricing in optimism. Nvidia's willingness to commit $2 billion to a single photonics supplier, combined with similar investments in Lumentum and Marvell, suggests the company sees optical interconnect not as a hedge but as a certainty. Coherent's expansion is both a response to that signal and a bet that the signal is accurate.

The next eighteen months will clarify whether the optical supply chain can scale in lockstep with AI infrastructure - or whether bottlenecks in wafer capacity, packaging, or testing become the next constraint on cluster growth. For Coherent, the question is not whether optics will replace copper, but whether the company can produce enough indium phosphide to keep pace with the replacement.