Technology and AI
Two million more processors in two years, and a hundred thousand of them behind a government security boundary
By Staff Writer | 28 August 2026
Amazon Web Services and Nvidia have committed to deploy two million additional graphics processors across the AWS estate in 2027 and 2028. A separate line in the same announcement covers 100,000 processors on secure infrastructure for United States federal and national security workloads at Impact Level 6 and above.
The joint announcement of 26 August is the second capacity commitment the two companies have made this year, and the first one has not run its course. At the March conference AWS said it would add more than one million Nvidia processors starting in 2026. The new document says demand has exceeded those expectations, and adds two million more across 2027 and 2028. The parts named are Blackwell Ultra, Rubin and Rubin Ultra, so the commitment spans three generations rather than one.
Numbers on that scale are easier to grasp as a construction programme than as a purchase order. Two million accelerators is a multi-year sequence of buildings, substations, transformers, cooling plant and grid connections, procured and consented in several jurisdictions at once, with the equipment specification changing between phases. The processors are the part that gets announced. The part that decides whether the announcement holds is everything underneath them.
The government line is the one to read twice
Set out among the commitments is a plan to build what the announcement calls AI factories for the United States government, including 100,000 processors on secure AWS infrastructure for federal and national security workloads. The document specifies workloads classified at Impact Level 6 and above, which is the accreditation tier for information up to secret. That is a different procurement world from commercial cloud capacity: separate accreditation, separate physical and personnel controls, and a separate audit trail on every component.
NVIDIA and AWS have built one of the great growth engines of the AI era, and demand is running ahead of every forecast.
Jensen Huang, founder and Chief Executive of Nvidia
Demand running ahead of forecast is the reason given for the second commitment arriving before the first was delivered. It is also, read the other way, an admission that the forecasts these programmes are planned against have not been reliable for at least a year. Anyone pricing a long-dated supply agreement into this market has to decide whose forecast the price risk sits with.
Beyond the accelerators
Several of the other commitments matter more to the engineering than to the headline. Nvidia's Vera server processors are to be brought onto AWS, giving a second option for workloads that need heavy conventional compute alongside accelerators. The NVLink Fusion interconnect support announced for Amazon's own Trainium chips at the November 2025 conference is being extended to Nvidia's custom high-bandwidth memory, so Amazon's in-house silicon and Nvidia's parts can sit inside a common rack architecture rather than in separate estates. That is the practical answer to the question of whether a hyperscaler building its own chips is a competitor or a customer. On this evidence it is both, in the same rack.
The performance claims are the vendors' own and are stated as comparisons rather than absolutes: the G7 instance generation is put at 4.6 times the inference performance and 2.1 times the graphics performance of the G6 generation before it; data processing with the named library at up to 3.7 times faster with 30 per cent better price performance than processor-based configurations; vector indexing at up to 9 times faster for a quarter of the cost. None of those is independently tested here and none should be relied on without a benchmark against the workload in question.
Customers want the freedom to choose the best tools for their AI workloads, and they want confidence that everything works seamlessly together.
Matt Garman, Chief Executive of Amazon Web Services
Amazon Robotics also appears in the announcement, adopting Nvidia's simulation and robotics platform for warehouse automation, with the training and validation running on the same accelerated instances. The two companies date their joint work to 16 years. What is new is that the announcements have stopped being about products and started being about capacity, and capacity is a civil engineering question.