Silicon photonics gains AI momentum

Silicon photonics is the idea of moving data with light instead of electricity — but on manufacturing processes close enough to mainstream chipmaking that hyperscalers can actually buy it at scale. That matters because AI systems are no longer mostly compute problems. They are communication problems. The bottleneck is the cost, power, and heat of shuttling enormous amounts of data between GPUs, switches, and memory-rich packages. What changed is that this stopped being a science-project story and became a product-and-roadmap story, with NVIDIA shipping a concrete networking plan in March 2025 and TSMC showing how optics gets packaged much closer to the silicon. ### What is silicon photonics, exactly? It is optical communication built with chip-style fabrication. Instead of sending high-speed signals over copper traces and cables, the system converts them into light, routes that light through photonic structures, and converts it back where needed. The appeal is simple — optical links lose less energy over distance, create less heat, and scale bandwidth better as AI clusters get larger. (nvidianews.nvidia.com) ### Why is AI suddenly the forcing function? Because giant training and inference clusters behave like one distributed machine. A modern AI deployment is full of accelerators waiting on data from other accelerators. As those clusters spread across racks and rows, copper starts to hurt twice — once in power draw and again in signal integrity. NVIDIA’s pitch for its Spectrum-X and Quantum-X photonics switches is basically that AI factories need to connect millions of GPUs without the network becoming the power hog or failure point. (tsmc.com) ### What did NVIDIA actually do? NVIDIA announced Spectrum-X Photonics and Quantum-X Photonics on March 18, 2025. The headline spec was 1.6 terabits per second per port, with claims of 3.5x energy savings and 10x resilience for large AI deployments. Just as important, NVIDIA framed this as a supply-chain stack, not a single chip — naming TSMC, Coherent, Corning, Foxconn, Lumentum, and SENKO as part of the build. That is what makes the announcement feel real. (nvidianews.nvidia.com) ### Where does TSMC fit in? TSMC is the manufacturing bridge between photonics demos and data-center hardware. Its COUPE platform — Compact Universal Photonic Engine — is built to stack a photonic die with an electrical control die, so the optical engine sits much closer to the compute package. TSMC says its 65nm silicon photonics process is already in volume production, and it has shown 200 Gbps optical modulation plus better than 99% 3D stacking yield on engineering samples. (nvidianews.nvidia.com) At its 2026 symposium, TSMC also highlighted COUPE as part of the path toward co-packaged optics for AI scaling. ### Why does “co-packaged optics” matter so much? Because the old model puts optics at the edge of the box as pluggable modules. That works, but every extra electrical hop burns power and adds complexity. Co-packaged optics moves the optical engine right next to the switch silicon. Basically, it is the difference between shouting across a room and talking into someone’s ear. Shorter electrical paths mean lower loss, better efficiency, and a cleaner route to higher bandwidth. (tsmc.com) ### Is this only an NVIDIA story? No — and that is another reason the momentum looks real. Lightmatter has been pushing a more aggressive version of the same idea with its Passage M1000, a 3D photonic interposer platform claiming 114 Tbps total bandwidth across a 4,000 mm² footprint. That is not the same product category as NVIDIA’s switches, but it points in the same direction: optics moving inward, from the cable to the package itself. (semiengineering.com) ### What is still hard? Lasers, packaging, thermals, testing, and cost. Optical systems are great at moving bits, but they introduce their own manufacturing headaches. The catch is that AI infrastructure only adopts new interconnect technology when reliability and serviceability are good enough for giant fleets. That is why partnership depth matters so much here — nobody gets to production with just a clever photonics die. (lightmatter.co) ### So what is the real takeaway? Silicon photonics is no longer just a nice future answer to AI bandwidth. It is becoming the architecture roadmap. First the optics moved into bigger pluggables. Now they are moving beside the switch. The next step is deeper package-level integration — and if that lands, the economics of scaling AI clusters change in a very big way. (nvidianews.nvidia.com)