AWS and NVIDIA triple their GPU deal: two million chips, claimed performance, not a word about energy

On 26 August 2026, in a joint press release, AWS and NVIDIA announced an expansion of their strategic partnership on AI infrastructure. The two companies say they plan to install 2 million additional GPUs — specifically the Blackwell Ultra, Rubin and Rubin Ultra models — across AWS's global infrastructure over 2027 and 2028. The announcement builds on an earlier commitment covering more than 1 million GPUs starting in 2026. According to a reconstruction by TechCrunch, the May 2026 order has thus been raised to a total of roughly 3 million units. The financial terms were not disclosed; TechCrunch's estimate, which puts the deal in the tens of billions of dollars, is so far not confirmed by the companies' official figures, and neither firm has said how many of the GPUs are already under contract, nor given a precise delivery schedule.

The cooperation goes beyond the supply of graphics compute, reaching into areas where Amazon has long been building its own answers. The agreement also calls for bringing infrastructure based on NVIDIA Vera CPUs to AWS, aimed at agentic AI workloads that need high-performance CPU compute alongside GPU acceleration. The parties state their intention to build «AI factories» for the US government, promising 100,000 GPUs on secure AWS infrastructure for federal and national-security workloads, including Impact Level 6 (IL6) and above — though the release names no delivery date. On the integration side, NVIDIA and Annapurna Labs, an Amazon subsidiary, are extending NVLink Fusion support to NVIDIA's custom high-bandwidth NVHBM memory, with the aim of letting Amazon's Trainium chips and the GPUs coexist inside the same rack-scale architecture. This convergence is happening while Amazon keeps developing its own silicon, Trainium and Graviton, which according to TechCrunch have passed an annualised revenue run rate of 25 billion dollars. On the software and robotics front, NVIDIA's open Nemotron models are coming to Amazon Bedrock and SageMaker, while Amazon Robotics adopts NVIDIA's platform for physical AI, including Jetson, Omniverse and Isaac.

The official release is thick with efficiency metrics, none of them verified by independent third parties. The vendors state that the new NVIDIA RTX PRO 4500 Blackwell Server Edition GPUs for Amazon EC2 G7 instances deliver 4.6 times higher AI inference performance and 2.1 times higher graphics performance than the previous generation. Again according to the two partners, data processing on Amazon EMR runs 3.7 times faster with 30% better price/performance; finally, the companies say vector indexing on Amazon OpenSearch is 9 times faster at a quarter of the cost. Without the starting costs, though, «a quarter of the cost» and «30% better» say nothing about what anyone actually saves.

The announcement landed on the same day NVIDIA published its results for the second quarter of fiscal 2027, ended 26 July 2026. Official figures from NVIDIA's Investor Relations pages show revenue of 96.2 billion dollars, up 18% on the previous quarter and 106% year on year, with the Data Center segment posting 89.0 billion dollars. Gross margin stands at 75.0%, with diluted earnings per share of 2.46 dollars GAAP and 2.22 dollars non-GAAP. Against that backdrop of rapid expansion, AWS CEO Matt Garman said in the official release: «Customers want the freedom to choose the best tools for their AI workloads, and they want confidence that everything works seamlessly together». In the same note, NVIDIA founder and CEO Jensen Huang said: «NVIDIA and AWS have built one of the largest growth engines of the AI era, and demand is running faster than any forecast».

The announcement falls on the day of the quarterly results, and the two sets of numbers are read together; but physics does not follow the rules of the stock market: sooner or later someone will have to explain what energy is going to run all these algorithm factories. — Olya