The news
On August 26, 2026, Amazon Web Services (AWS), the cloud unit of Amazon (AMZN), and Nvidia (NVDA) announced an expanded AWS Nvidia deal: AWS plans 2 million more Nvidia GPUs in 2027-2028, and the companies plan AI factories for the U.S. government.
According to the joint release, issued from Seattle and Santa Clara, California, AWS plans to deploy the additional GPUs across its global infrastructure, spanning Nvidia's Blackwell Ultra, Rubin and Rubin Ultra generations. The plan builds on one AWS announced at Nvidia's GTC 2026 conference to add more than 1 million Nvidia GPUs starting in 2026. The companies said demand had since exceeded those expectations. TechCrunch described the move as Amazon tripling its Nvidia chip order.
For the government, the companies said they plan to build AI factories that deliver Nvidia's AI stack, including plans for 100,000 GPUs on AWS's secure infrastructure for federal and national-security workloads. The release said these would support workloads classified at Impact Level 6 (IL6) and above, a security tier for classified government work. It gave no location, cost or completion date.
The deal reaches beyond GPUs. AWS and Nvidia said they were working to bring infrastructure based on Nvidia's Vera CPU to AWS as another option for agentic AI, meaning AI software that carries out multistep tasks, which needs high-performance CPU compute alongside accelerators. No launch date was given. Nvidia and Amazon's Annapurna Labs chip unit were also extending earlier work on Nvidia's NVLink Fusion chip interconnect for future Trainium chips, Amazon's in-house AI processors, to cover Nvidia's new custom high-bandwidth memory (NVHBM). The companies said that would give Trainium faster, more power-efficient memory.
AWS also said it was the first major cloud provider to offer compute instances accelerated by Nvidia's RTX PRO 4500 Blackwell Server Edition GPU, through Amazon EC2 G7 instances. It said G7 delivered 4.6 times the AI inference performance and 2.1 times the graphics performance of the previous G6 generation. Nvidia's Nemotron open models were available on Amazon Bedrock as fully managed, serverless models, according to the release. AWS CEO Matt Garman said customers wanted "the freedom to choose the best tools for their AI workloads."
The announcement came the same day Nvidia reported results for its fiscal second quarter ended July 26, 2026. Its earnings release, filed with the Securities and Exchange Commission, showed revenue of $96.2 billion, including $89.0 billion from its Data Center business, and an outlook for third-quarter revenue of $108.0 billion, plus or minus 2%. That release did not mention AWS.
The numbers
- Additional Nvidia GPUs AWS plans for 2027-2028
- 2 million
- Earlier AWS plan, announced at GTC 2026
- More than 1 million GPUs starting in 2026
- GPUs planned for U.S. government AI factories
- 100,000
- G7 vs. G6 performance, per AWS
- 4.6x AI inference, 2.1x graphics
- Nvidia fiscal Q2 revenue (quarter ended July 26, 2026)
- $96.2 billion
- Nvidia supply commitments, per TechCrunch
- $279 billion, up from $119 billion
Why CEOs should care
For technology buyers, the 2 million GPUs are a 2027-2028 plan, not capacity available today. Companies negotiating multi-year AWS commitments should ask their account teams which GPU generations, Blackwell Ultra, Rubin or Rubin Ultra, will reach which regions and when, and get those terms in writing. The companies disclosed no pricing, so more supply should not be read as a promise of lower rates.
For CFOs and CTOs, the deal shows Amazon backing both Nvidia and its own Trainium chips, and tying them closer through NVLink Fusion and NVHBM memory. That widens choice, but it also makes portability a budgeting question. Ask how easily models and data pipelines move between Nvidia-based instances and Trainium, and whether committed spend can shift between them. Teams running Apache Spark jobs or vector search can test AWS's claims of up to 3.7 times faster GPU-accelerated processing on Amazon EMR and up to 9 times faster vector indexing on Amazon OpenSearch Service against their own workloads.
For CISOs and public-sector leaders, AI factories built for IL6-and-above workloads could give agencies and defense contractors a way to run AI on classified data inside AWS. Until the companies name dates, locations and accreditation status, treat them as plans. Boards should ask whether their company's AI roadmap assumes capacity from plans that run through 2028, and what the fallback is if deliveries slip.
The bigger picture
Cloud providers and chipmakers are locking in supply years ahead. TechCrunch reported that Nvidia has committed $279 billion to secure supply and manufacturing capacity, up from $119 billion the prior quarter. It also quoted Nvidia Chief Financial Officer Colette Kress as saying Vera shipments were under way to lead partners, which TechCrunch said include Oracle and SpaceXAI. AWS's Vera plan adds it to that rollout.
Amazon is hedging at the same time. TechCrunch reported that Amazon's custom chip business is growing, driven by $225 billion in total commitments from AI labs such as Anthropic and OpenAI, and that Amazon AI chief Peter DeSantis has said AWS is in talks to sell its Trainium chips. On the government side, ExecutiveBiz noted AWS's earlier plan to invest $50 billion in federal AI and supercomputing infrastructure across its Top Secret, Secret and GovCloud offerings.
What happened next
Through September 29, 2026, Tech CEO Daily found no further announcement from AWS or Nvidia setting a date for Vera-based infrastructure on AWS or naming locations for the government AI factories. Coverage on August 27 by Engineering.com and ExecutiveBiz restated the companies' terms without adding new ones.
What to watch: Nvidia's fiscal third-quarter results against its $108.0 billion revenue outlook; Amazon's third-quarter 2026 results, for updates on capital spending and GPU deployment; and AWS's re:Invent conference, where AWS announced NVLink Fusion support for Trainium in 2025 and could give timelines for Vera and G7 availability.
What “Fact-checked” means
Fact-checking means testing a story’s facts against the evidence before it is published. This story went through at least two separate checks before this version was published.
- What we checked
- Its names, figures, dates, job titles, quotes and who said what were checked against the story’s sources, including its main source where it could be opened. The headline was checked for accuracy and overstatement.
- How
- A first check reviewed the whole story. If it passed, a second, skeptical check went back to the sources to look for mistakes in the most important facts. If a check flagged the story, it was edited to fix the problems found, and a separate re-check then reviewed the whole story again.
- Who
- The checks are made with our newsroom’s technology tools, as steps kept separate from the writing, under rules set by our editor, Hussein Mukhtar. A story the checks still flag is not published automatically; it is held for the editor, who decides whether it is fixed, published or dropped.
- Archive story
- It was written after the event. Its dates were checked against the date of the event, including the dates of any later developments it reports.
- If something is wrong
- “Fact-checked” does not mean error-free. If a material error is found after publication, we correct the story and add a note saying what changed. Report an error
