- AWS deal with Nvidia adds 2 million Blackwell Ultra, Rubin, and Rubin Ultra GPUs, on top of more than 1 million committed in March 2026
- Totals more than 3 million Nvidia GPUs due before 2029, a ~1.9x multiple for each of Amazon's roughly 1.58 million employees
- Nvidia's CFO has said deliveries start this quarter, not in 2027, but neither company says what it counts as a GPU or what the deal costs
Amazon ended 2025 with about 1.58 million full- and part-time employees, most in fulfillment centers and delivery operations, with a much smaller share working in its AWS data center division (estimated at around 246,000).
By mid-2028, Amazon Web Services expects to have taken delivery of roughly twice that many new Nvidia GPUs, based on a recent deal it made with Nvidia.
The latest deal was announced on August 26, the same afternoon Nvidia reported quarterly earnings, and will see AWS plan to deploy two million additional Blackwell Ultra, Rubin, and Rubin Ultra GPUs across its global infrastructure in 2027 and 2028. This is on top of over a million Nvidia GPUs that AWS committed to at Nvidia GTC 2026 in March, and which Nvidia's Ian Buck told Reuters would ship from 2026 through 2027.
A mammoth deal that dwarf's Amazon's entire payroll
Together, this all amounts to over three million GPUs in play between the data center giant and the chip designer, which is also the world's most valuable company at the time of writing. At 3 million GPUs against a workforce of 1.58 million, that's nearly 1.9 Nvidia GPUs per employee.
This is only Amazon's commitment to Nvidia; it also has its own custom Tranium3 GPUs, which the company says are "nearly fully subscribed" or "almost fully booked" as of mid-2026.
At the same time, Amazon is not increasing job roles on its end; it cut around 30,000 corporate roles between October 2025 and January 2026, a restructuring its top HR executive, Beth Galetti, described as "reducing layers, increasing ownership, and removing bureaucracy."
Building on an already solid base
While AWS's current GPU base from Nvidia, consisting mainly of H100 and H200 GPUs, is smaller than Microsoft's 485,000 count, it is hardly a rounding error, with Omdia estimating that Amazon bought 196,000 Hopper GPUs in 2024 alone.
The interesting part is determining what counts as a GPU: Nvidia and Amazon's joint statement about 2 million GPUs across calendar years 2027 and 2028 might mean two very different things depending on how you count GPUs.
At GTC 2025, Nvidia said it would count Rubin-generation GPUs by compute die rather than by package, which is why the Vera Rubin rack was first unveiled as NVL144. It reversed course before CES 2026, and the flagship now ships as Vera Rubin NVL72, counting 72 two-die packages. To make matters a bit more complicated, Rubin Ultra, due in 2027, puts four compute dies in each package.
Whether the companies count each die as a separate GPU or the whole package as a GPU could make the numbers skew heavily in either direction. The deal also includes a national security element: approximately 100,000 GPUs are earmarked for US government AI workloads at Impact Level 6.
Amazon's limitation here is supply, not demand; AWS revenue rose 37% year over year to $42.2 billion in the second quarter, and its contracted backlog reached $496 billion, up from $364 billion three months earlier, as CloudTech News reported.
On Amazon's July earnings call, CEO Andy Jassy said capacity would fall short of demand this year and likely next, adding: "In fact, the demand we already have for 2028 is striking."
In addition to its own custom ASICs, including Tranium3, Amazon is gearing up for a few fiscal years of eye-watering AI infrastructure spend and AI demand combined, and it better be right: Amazon's free cash flow has already turned negative as it looks to fund its bid to be one of, if not the largest hyperscaler in play for the foreseeable future.
Follow TechRadar on Google News and add us as a preferred source to get our expert news, reviews, and opinion in your feeds.
English (US) ·