AI data centers hit interconnect limits
- Nvidia, Broadcom and Marvell are all recasting AI networking around optics in 2026, as bigger GPU clusters run into copper reach, power and reliability limits. - Broadcom says its Tomahawk 6-Davisson co-packaged switch delivers 102.4 terabits per second and cuts optical interconnect power use by more than 70%. - Nvidia plans photonics switches for 2026 as AI spending widens beyond GPUs to networks, fiber and packaging. (nvidia.com)
AI data centers are running into a new limit: moving data between chips fast enough, not just buying more chips. (nvidia.com) (marvell.com) Inside an AI cluster, the network is the wiring that lets thousands of graphics processors act like one machine. As those clusters spread across more racks, copper links lose signal, burn more power and become harder to cool. (marvell.com) (broadcom.com) That is why vendors are pushing co-packaged optics, which moves optical parts next to the switch chip instead of plugging them in at the edge. The aim is to shorten electrical traces before turning signals into light. (nvidia.com) (broadcom.com) Nvidia said its Spectrum-X Ethernet Photonics and Quantum-X Photonics switches are scheduled for commercial availability in 2026. The company says the systems deliver up to 409.6 terabits per second, 800-gigabit ports and power savings of up to 3.5 times over older designs. (nvidia.com 1) (nvidia.com 2) At GTC in March 2026, Nvidia also described new rack-scale systems linked through Kyber for copper and co-packaged-optics scale-up, plus optical scale-out networking. The message was that the network is now part of the compute system, not a separate box. (blogs.nvidia.com) Broadcom has been making the same case from the switch side. Its Tomahawk 6-Davisson, announced as shipping, is a 102.4-terabit-per-second Ethernet switch with co-packaged optics that Broadcom says cuts optical interconnect power consumption by more than 70%. (broadcom.com) Marvell says the pressure is strongest in “scale-up” fabrics, the short-reach links that tie accelerators together inside AI servers and racks. The company says copper’s physical limits are driving demand for both co-packaged optics and co-packaged copper, depending on distance. (marvell.com 1) (marvell.com 2) Industry analysts are framing 2026 as the year optical interconnect stopped being a telecom niche and became a data-center priority. Yole Group said after OFC 2026 that competition is shifting from transceivers toward packaging, integration and system design. (yolegroup.com) That shift reaches beyond networking vendors. Broadcom’s Davisson uses TSMC photonic-engine technology in its package, and Nvidia has named partners including TSMC, Coherent, Corning, Foxconn, Lumentum and SENKO in its photonics supply chain. (broadcom.com) (nvidia.com) The result is that AI infrastructure spending is widening from graphics processors and memory into optics, fiber, lasers and advanced packaging. In 2026, the fastest way to build a bigger AI cluster may be to move more of its network at the speed of light. (semiengineering.com) (nvidia.com)
