Amazon's last GPU commitment to Nvidia ran out early. The replacement is two million more chips, and it does not start until 2027.
AWS and Nvidia say they will deploy two million additional GPUs across Amazon's global infrastructure in 2027 and 2028 — Blackwell Ultra, Rubin and Rubin Ultra parts. AWS had announced more than one million GPUs starting in 2026 at Nvidia's GTC in March; demand consumed that plan faster than expected, and the new commitment takes the total introduced across AWS this year past three million. The deal also brings Nvidia's Vera CPUs to AWS, and has Amazon's Annapurna Labs extending NVLink Fusion so Nvidia's custom high-bandwidth memory can be attached to Amazon's own Trainium chips. Separately the two will build AI factories for the US government, with 100,000 GPUs on secure AWS infrastructure rated Impact Level 6 and above.
Why it mattersNvidia told investors last week that next year is bigger only because it cannot build any faster. This is that claim written as a purchase order by the largest cloud on earth — and dated 2027, which is the honest signal: the constraint is not orders, it is silicon, power and time. The Trainium detail cuts against the usual story of Amazon building its own chips to escape Nvidia; here Nvidia's memory goes inside Amazon's accelerator, which is cooperation, not substitution.
Read in the app — free, in 9 languages
Related stories
Microsoft says its own chips will become the main engine for its AI.2026-10-12Drones have shut both sites that house Yandex's AI supercomputers.
2026-10-12McKinsey says 11 million Americans will need a new line of work.
2026-10-12Amazon is in talks to buy the startup Anthropic walked away from.
2026-10-12Salesforce renamed AIForce to SIForce, though no rule required it.
2026-10-11