AWS locks in two million more Nvidia GPUs through 2028

Amazon Web Services has never retired an A100 server, and that six-year-old chip—still running in its data centers—is the backdrop for the commitment the company announced this week: an additional two million Nvidia GPUs deployed across its global infrastructure through 2027 and 2028. The old hardware and the new order capture the strain underneath the AI buildout: AWS is simultaneously running its oldest gear into the ground and signing up for Nvidia's newest roadmaps—Blackwell Ultra, Rubin, and Rubin Ultra—as if the one-million-GPU fleet it promised for 2026 had already been swallowed by demand.

The prior benchmark is the easiest way to read the scale: at Nvidia GTC 2026, AWS said it would bring online more than one million Nvidia GPUs in 2026, but demand has since outpaced that number. The new two million includes Blackwell Ultra, Rubin, and Rubin Ultra, which means the commitment stretches into a period when Nvidia's next-generation architecture is the volume product—the same unbuilt chip roadmap that has carried other valuations in this cycle. In the same week, Nvidia reported $96.2 billion in Q2 2026 revenue, up 18 percent from the previous quarter and 106 percent from a year earlier, with GAAP and non-GAAP gross margins both holding at 75 percent.

Those margins are why Huang can keep saying demand is running ahead of every forecast, and the AWS commitment is the kind of backlog that supports the claim. Garman framed the deal as a matter of customer choice—customers want the freedom to pick the best tools for their AI workloads, he said, and confidence that everything works seamlessly together—and that is the same story the two chief executives are telling, with the reported numbers on their side.

The government AI factories alone—100,000 GPUs on secure AWS infrastructure—carry their own power, security, and siting requirements, and that should worry an infrastructure underwriting desk. As this publication has argued, the AI buildout's bottleneck has shifted from land and chips to electrons; consent, not capital, now decides what gets built. A two-million-GPU fleet does not deploy itself; every GPU ends up inside a rack, the rack inside a building, the building behind a substation, and the substation somewhere in an interconnection queue. The commitment through 2028 implies AWS is buying power rights and data center capacity on the same horizon as the silicon.

Garman's throwaway line about the A100s deserves a second look: the cloud giant is still running six-year-old A100s, and if it has never retired an A100 server, that says more about utilization than sentiment. The one-million-GPU plan for 2026 was supposed to be the relief valve; the two-million follow-on suggests that valve was insufficient. In that light, the new commitment reads less like a procurement announcement than an acknowledgment that the constraint on AI infrastructure is no longer the fab but everything downstream of the chip—power, land, and the patience of the communities where these campuses land.

None of this makes the commitment defensive: Huang called the AWS partnership 'one of the great growth engines of the AI era,' and Garman said AWS has invested deeply with Nvidia to make its cloud the best place to run Nvidia AI technologies. The strategic alignment is real, but the unproven part is whether the downstream assets can be delivered at the same pace as the GPUs. A grid slot is worth more than the hardware behind it, and AWS has now placed a wager on that idea that runs through 2028.