NVIDIA ramps AI networking
NVIDIA said its new CPO Spectrum‑X switch ASIC is now in full production and confirmed a massive supply deal with AWS for up to 1 million GPUs through 2027 — the agreement includes Spectrum networking and ConnectX infrastructure, expanding NVIDIA beyond compute into full AI stacks. (x.com)
Spectrum‑X’s photonics switch platform is built around 1.6 Tb/s per‑port CPO engines and scale options that reach hundreds of terabits of aggregate throughput on a single chassis. (nvidianews.nvidia.com) NVIDIA says the Spectrum‑X design was co‑developed with TSMC and supply‑chain partners including Corning, Foxconn and Lumentum to integrate silicon photonics and optics directly on switch silicon. (investor.nvidia.com) NVIDIA’s product literature cites co‑packaged optics power improvements of multiple‑times energy savings versus pluggable optics and claims far higher link resiliency at AI‑scale deployments. (investor.nvidia.com) Meta Platforms and Oracle have already standardized Spectrum‑X into open switching and giga‑scale AI plans, with Meta integrating Spectrum‑X into FBOSS and Oracle citing Spectrum‑X for its “giga‑scale AI factory” builds. (investor.nvidia.com) NVIDIA’s hyperscale lead added a scheduling detail in comments to Reuters: shipments tied to the broader AWS arrangement are slated to begin this year and continue through 2027, according to Ian Buck, NVIDIA’s VP for hyperscale and HPC. (money.usnews.com) The AWS arrangement goes beyond GPUs to include NVIDIA networking stack deployments — an explicit plan to deploy ConnectX‑class NICs and Spectrum‑class switches inside AWS data centers, a departure from AWS’s long history of primarily custom in‑house networking. (businesstoday.com.my) NVIDIA’s ConnectX family now advertises up to 400Gb/s per port on ConnectX‑7 and SuperNIC generations supporting up to 800Gb/s and in‑hardware offloads (GPUDirect, RoCE, inline crypto) that are designed to reduce CPU overhead in large AI inference fabrics. (nvidia.com)
