Co-packaged optics goes commercial as AI networks hit a power wall
Nvidia ships Quantum-X Photonics, brings co-packaged optics to Ethernet in H2 2026; light replaces copper to wire million-GPU clusters at 3.5x the efficiency
Summary
Copper interconnect is hitting reach and power limits as AI clusters scale, and co-packaged optics (CPO) is moving from demo to commercial deployment in 2026. Nvidia is bringing Quantum-X InfiniBand Photonics switches to market early in 2026 and Spectrum-X Ethernet Photonics in the second half, integrating optical engines directly onto the switch ASIC for up to 3.5x better power efficiency and 10x resiliency, the precondition for wiring million-GPU Rubin-era clusters. The strategy splits the field: Nvidia embeds photonics deep in a system-integrated architecture, while Broadcom pursues a modular, Ethernet-leveraged path; photonics startup Lightmatter is the wildcard, also inside NVLink Fusion. Optical interconnect borrows the same silicon-photonics base as photonic quantum.
By the numbers
- ~3.5x, power-efficiency gain from co-packaged vs pluggable optics.
- ~10x, network-resiliency improvement (fewer optical components).
- 400 Tb/s, top Spectrum-X Photonics switch throughput (512x 800Gb/s).
- H2 2026, Spectrum-X Ethernet Photonics commercial availability.
- 5 years, analyst estimate for all AI data-centre interconnects going optical.
Why it matters
At million-GPU scale, the optics, not the GPUs, set the power and reliability ceiling. Pluggable transceivers burn too much energy and fail too often; co-packaging light onto the switch is how AI networks keep scaling. It is also a new margin battleground: Nvidia, Broadcom and a wave of photonics startups are fighting for a layer that barely existed two years ago.
What to watch
- Spectrum-X Ethernet Photonics shipping on schedule in H2 2026.
- Nvidia's system-integrated vs Broadcom's modular CPO winning hyperscaler designs.
- Optical-engine yield and field reliability as the real scaling test.