AWS triples its Nvidia commitment to 3M GPUs — the 1M order ran out early

AWS and Nvidia announced on 26 August 2026 that they will deploy 2 million additional Nvidia GPUs across AWS infrastructure in 2027–2028, comprising Blackwell Ultra, Rubin and Rubin Ultra parts, for agentic AI, scientific computing and physical AI workloads. That takes AWS's committed total from more than 1 million GPUs to more than 3 million. The earlier commitment — announced at Nvidia's GTC 2026 for deployment starting in 2026 — was consumed faster than planned. The expanded collaboration also spans CPUs, networking, memory, open models, data processing and robotics, and includes building AI factories for the US government with 100,000 GPUs on secure AWS infrastructure.

Why it matters

The interesting number is not 2 million, it is that the previous million ran out early. Hyperscaler capacity announcements are usually forward-looking commitments with soft dates and no way to check them; this one comes with an implicit admission that a stated 2026 plan was under-scoped against realised demand. That is the closest thing to a demand datapoint the accelerator market produces, because it is a customer revising upward against its own published figure rather than a vendor projecting.

It also puts a scale on where the compute buildout is heading. Three million GPUs committed by a single cloud, with delivery pushed into 2027–2028 on parts (Rubin Ultra) that do not exist yet, is a bet placed two product generations ahead — and it is being placed at the same time as Nvidia's $105B lease backstop for OpenAI's Ohio campus, which is the same company underwriting demand at the other end. Nvidia is now simultaneously the supplier of record for the largest cloud commitment and a guarantor for a rival buildout. That is a healthy order book and a concentrated one, and the two readings are not in conflict.

What to watch

Whether AWS reports actual deployed GPU counts against the 3 million commitment rather than restating the commitment, when Rubin Ultra silicon slips or holds, and whether the 100,000-GPU government capacity gets a named program.