Amazon AWS plans to purchase more than 3 million additional chips by 2029.

AWS has reached an agreement with Nvidia to add 2 million Blackwell Ultra, Rubin and Rubin Ultra GPUs by March 2026, on top of its previous commitment to add more than 1 million units.

By 2029, Amazon will procure a total of more than 3 million Nvidia GPUs, equivalent to roughly 1.9 GPUs for each of its approximately 1.58 million employees.

Nvidia's Chief Financial Officer said deliveries will begin this quarter, not in 2027, but neither company has disclosed the specific definition of the GPUs or the transaction price.

At the end of 2025, Amazon had about 1.58 million full-time and part-time employees, the vast majority of whom worked in logistics centers and distribution departments, with only a small portion working in its AWS data center division (estimated at about 246,000 people).

Under the recent agreement reached by Amazon Web Services with Nvidia, the company expects to receive approximately twice that number of new Nvidia GPUs by the middle of 2028.

The latest agreement was announced on August 26, the same afternoon that Nvidia released its quarterly earnings report. Under the agreement, AWS plans to deploy an additional 2 million Blackwell Ultra, Rubin and Rubin Ultra GPUs across its global infrastructure in 2027 and 2028. Previously, AWS had committed to purchasing more than 1 million Nvidia GPUs at Nvidia's GTC 2026 conference in March, and Nvidia's Ian Buck told Reuters that these GPUs would be delivered between 2026 and 2027.

This is an enormous deal whose value even exceeds the total number of Amazon's employees.

All in all, this means more than 3 million GPUs will be used in total between the data center giant and the chip design company (which is also the company with the highest market capitalization in the world as of the time of writing). Three million GPUs correspond to 1.58 million employees, meaning that each employee owns nearly 1.9 Nvidia GPUs on average.

This is only Amazon's commitment to Nvidia; it also has its own custom Tranium3 GPUs, which the company says are "almost fully reserved" or "almost completely sold out" as of mid-2026.

Meanwhile, Amazon is not adding to its own headcount; between October 2025 and January 2026, the company cut about 30,000 corporate positions, and its head of human resources Beth Galetti described the restructuring as "reducing layers, enhancing ownership and eliminating bureaucracy".

While the GPUs that AWS currently procures from Nvidia mainly consist of H100 and H200 GPUs, with a quantity less than Microsoft's 485,000 units, this is by no means insignificant. According to Omdia's estimates, Amazon purchased 196,000 Hopper GPUs in 2024 alone.

Interestingly, the definition of GPU matters: the joint statement from Nvidia and Amazon says 2 million GPUs will be launched in 2027 and 2028, but depending on how the GPUs are counted, this could mean two very different things.

At the GTC 2025 conference, Nvidia stated that it would count Rubin series GPUs by the number of computing chips rather than the number of packages, so the Vera Rubin graphics card was first unveiled under the name NVL144. But at the 2026 CES, Nvidia changed its strategy, and the flagship product is now shipped under the name Vera Rubin NVL72, which contains 72 dual-chip packages. To make things more complicated, the Rubin Ultra graphics card, expected to be released in 2027, contains four computing chips per package.

Whether each company counts each individual chip as a separate GPU or counts the entire package as a GPU can lead to a significant deviation in the final figures. In addition, the transaction also involves national security: about 100,000 GPUs will be used for AI workloads at the Level 6 impact tier of the U.S. government.

According to CloudTech News, Amazon's constraint lies in supply rather than demand; AWS's second-quarter revenue rose 37% year-on-year to $42.2 billion, and its contracted backlog reached $496 billion, up from $364 billion three months earlier.

On Amazon's earnings call in July, CEO Andy Jassy said that this year's production capacity will not be able to meet demand, and the same will likely be true next year, adding that "in fact, our demand for 2028 is already incredibly strong at the moment."

In addition to its own custom ASIC chips including Tranium3, Amazon is preparing to invest heavily in AI infrastructure in the coming fiscal years to cope with the surge in AI demand. Moreover, Amazon must be properly prepared: to become one of the largest hyperscale data center operators (if not the largest) for the foreseeable future, Amazon's free cash flow has already turned negative.

This article is sourced from the WeChat official account "Tech Business" and is republished with authorization from 36Kr.