NVIDIA locks AWS 1M‑chip deal
NVIDIA agreed to sell Amazon Web Services up to 1 million AI chips by the end of 2027, a move that secures hyperscaler capacity and risks tightening availability for smaller buyers. That kind of forward allocation can deepen the competitive advantage of cloud giants and pressure startups to prioritize cost‑efficient model choices. (seekingalpha.com) (tipranks.com)
Ian Buck, NVIDIA’s vice president for hyperscale and high‑performance computing, told Reuters that shipments under the AWS agreement will begin this year and continue through the end of 2027. (money.usnews.com) NVIDIA confirmed the AWS order spans more than GPUs, explicitly covering Spectrum networking silicon and inference processors tied to its December deal for Groq technology. (finance.yahoo.com) AWS used NVIDIA’s GTC 2026 stage to say it plans to deploy more than one million NVIDIA GPUs across global regions starting in 2026. (aws.amazon.com) Datacenter-focused reporting says AWS expects those deployments within roughly the next 12 months as it rolls out broader Rubin/Blackwell and inference capacity. (datacenterdynamics.com) NVIDIA’s Rubin/Vera platform is being positioned as a multi‑silicon stack — the company’s Rubin announcement described six new chips integrated across GPUs, CPUs, DPUs and networking to cut inference token costs. (investor.nvidia.com) NVIDIA CEO Jensen Huang projects the Blackwell and Rubin families represent roughly a $1 trillion revenue opportunity through 2027, a framing analysts note the AWS commitment maps directly onto. (bloomberg.com) NVIDIA executives have said AWS will deploy a mix of specialized silicon for inference — “you cannot rely on just one chip” — underscoring that the agreement calls for multiple chip types to be combined in production stacks. (thetechnologyexpress.com) Market commentators flagged the scale of the commitment as among NVIDIA’s largest single‑client supply arrangements to date and warned that locking multi‑year volume with a hyperscaler can concentrate supply and pressure pricing and lead times for smaller data‑center customers. (computing.net)
