THE SIGNAL
Nvidia Is Buying The Light Between The Chips
The race for AI compute looks like a fight over GPUs. The real contest is over the optical wiring that lets thousands of those GPUs behave as one machine, and Nvidia is quietly buying the factory that makes it.
The Bottleneck Was Never The Chip
What happened: Coherent, an optics manufacturer that Nvidia invested in earlier this year, announced a $650 million expansion of its plant in Sherman, Texas, to quadruple production of indium phosphide wafers. Indium phosphide is the niche semiconductor material used to build the lasers and optical parts that move data between AI chips as pulses of light instead of electrical signals over copper. The expansion is backed by public money: roughly $20 million from Texas state and local funds, and up to $50 million from the federal CHIPS and Science Act. Coherent says the project will add about 1,000 jobs.
What's really going on: As AI systems grow from a few dozen chips in a rack to hundreds or thousands wired together as a single computer, copper can no longer carry the data fast enough across those distances, so the connections have to switch to optics. That turns indium phosphide, a material most people have never heard of, into a hard physical ceiling on how large any AI cluster can be built. Nvidia, which already sells the chips, put a reported $2 billion into Coherent earlier in 2026, and is now watching public subsidy fund the capacity that secures its own supply. The company is extending its hold from the processor to the wiring between processors, the one layer a competitor might have used to build an alternative.
Why most people are missing this: They assume the scarce resource in AI is the GPU, when the harder limit is the optical material that decides whether thousands of GPUs can be joined into one machine at all.
The Take: Nvidia did not just buy more supply, it bought the chokepoint, and got three levels of government to help pay for it.
Why it matters: Once the interconnect layer consolidates around a single supplier that Nvidia holds a stake in, the cost of building a competitive AI cluster without Nvidia rises at every level, not only the chip. Public funding meant to widen domestic chip capacity may instead harden one company's grip on the part of the stack almost no one was watching.
The Pattern
The tension is between vertical control, where one company quietly owns every layer that matters, and an open supply chain, where buyers can still mix parts from rivals. Vertical control is winning, because the layers people ignore are the easiest to corner without anyone objecting. The move that decides the next decade is not the chip everyone tracks. It is the material almost no one can name.
What This Signals
Optical interconnect capacity, not raw chip count, becomes the number that sets how large the next generation of AI clusters can actually be built.
Nvidia's stake in the company expanding that capacity turns a supplier relationship into a structural dependency, and a structural dependency is far harder to undo than a temporary pricing edge.
Government subsidy meant to spread semiconductor production may be funding consolidation instead, paying to expand the one supplier its largest customer already has a claim on.
Quick Byte
Indium is one of the rarest metals refined at industrial scale, recovered almost entirely as a byproduct of zinc mining, with no dedicated mines of its own. An entire generation of AI infrastructure is being built on a material the market only produces by accident.