Co-packaged optics: the fabric for million-GPU clusters

A million-GPU cluster has a hidden enemy, and it isn't the chips. It's the wiring between them. At that scale, the lasers that shuttle data from GPU to GPU burn a startling share of the power budget and fail often enough to matter. NVIDIA's fix, co-packaged optics, is the quiet breakthrough that makes million-GPU clusters physically possible. And, like every other layer of this story, it comes back to power.

The breakthrough, in numbers

At GTC 2025, NVIDIA unveiled Spectrum-X and Quantum-X silicon-photonics switches, built to connect millions of GPUs across an AI factory. By integrating the optics into the switch package itself, they use 4x fewer lasers and deliver 3.5x more power efficiency, with 63x better signal integrity and 10x the network resiliency of traditional pluggable optics (NVIDIA).

The bandwidth is staggering: the Spectrum-X Photonics platform offers 1.6 terabits per second per port, while Quantum-X promises roughly 2x the speed and 5x the scalability of the prior generation. The switches were unveiled in 2025 and have been moving toward deployment through 2026, backed by a silicon-photonics supply chain including TSMC, Corning, Foxconn and Lumentum (HPCwire).

Why the network became a power problem

Here's the part that surprises people. When you scale a cluster to hundreds of thousands or millions of GPUs, the interconnect stops being a detail and becomes a dominant consumer of power and a frequent source of failure. Traditional pluggable optical modules each carry their own lasers, and at fabric scale those lasers add up to megawatts of consumption and a long list of components that can break. The wiring, not the compute, starts to gate how big you can go.

Co-packaged optics attack that directly. By moving the optics next to the switch silicon and slashing the laser count, they reclaim the power and reliability the fabric was quietly bleeding. It is, at heart, another efficiency play, and like all of them it exists because power and reliability are the binding constraints once you scale far enough.

Efficiency scales the cluster; cheap power makes it pay

Co-packaged optics make a million-GPU cluster possible. They do not make it cheap. A more efficient fabric still plugs into a building, and the bigger the cluster the larger the absolute power bill, even after every efficiency gain. Photonics shaves the network's slice; it does nothing for the price of the kilowatt-hour feeding the whole machine.

So the two levers are complementary, not competing. The networking breakthrough lets you build the cluster. The power economics decide whether running it is profitable. You want both: the efficient fabric to scale, and the cheap power underneath so the scale doesn't bankrupt you. Efficiency without cheap power is a faster way to lose money; cheap power without efficiency leaves performance on the table.