Loading page…
The partnership reaches beyond GPU supply into CPUs, networking and AWS’s own Trainium chips. The largest capacity commitment remains a future deployment, making execution on the shared infrastructure the central test.
Listen to this story
AWS is turning its Nvidia relationship into a multi-year capacity and systems bet: after an earlier plan for more than 1 million GPUs from 2026, it now targets 2 million additional Nvidia GPUs in 2027–28 and a 100,000-GPU secure buildout for U.S. federal workloads. The architecture is meant to combine Nvidia GPUs with future Trainium chips via NVLink Fusion, rather than choose one accelerator.
The 100,000-GPU federal figure covers planned Impact Level 6-and-above infrastructure, not currently rentable capacity.
Next-generation AWS Trainium is expected to support Nvidia NVLink Fusion; Annapurna Labs is working on Nvidia HBM technology.
Nvidia Vera CPU systems are being developed for code execution, tool use, sandboxing, analytics, pipelines, and agent orchestration on AWS.
AWS plans to deploy 2 million additional Nvidia GPUs across its global infrastructure in 2027 and 2028 and plans U.S. government AI factories with 100,000 GPUs on secure AWS infrastructure. The commitment pairs a large future capacity target with a deeper effort to join Nvidia hardware to AWS’s own AI chips and cloud systems.
The technical center of the deal is an effort to make AWS’s Trainium accelerators work more closely with Nvidia technology. AWS says next-generation Trainium chips will support Nvidia NVLink Fusion, a high-speed chip interconnect, while Amazon’s Annapurna Labs works with Nvidia’s custom high-bandwidth-memory technology. The intended architecture would integrate Trainium and GPUs in a common rack-scale design.
The companies are not presenting Nvidia GPUs and Trainium as mutually exclusive choices. AWS says the expanded capacity is intended for agentic AI, scientific discovery, enterprise automation and physical AI. For agentic systems, AWS and Nvidia are also working to bring Vera CPU-based infrastructure to AWS for code execution, tool use, sandboxing, analytics, data pipelines and orchestration.
AWS had already said at Nvidia GTC 2026 that it would add more than 1 million Nvidia GPUs starting in 2026. It now says demand exceeded those expectations and has set out the additional 2 million-GPU plan for the following two years. The announcement is therefore a forward deployment commitment rather than capacity customers can assume is available now.
The government portion gives the partnership a distinct planned deployment path. AWS and Nvidia plan to provide Nvidia’s AI stack through secure AWS infrastructure for federal and national-security workloads classified at Impact Level 6 and above. The 100,000-GPU figure describes the planned federal buildout, not capacity available today.
For AWS, the near-term evidence is existing software and instance integrations; the central test is execution on the planned infrastructure. Vera-based AWS systems remain under development, the federal AI factories are planned, and the 2 million additional GPUs are scheduled across 2027 and 2028. The partnership’s business effect will depend on those components arriving as an integrated service at the promised scale.
Editorial analysis
The strategic signal is not simply that AWS wants more Nvidia chips. AWS is also preserving a role for Trainium by connecting it to Nvidia’s NVLink Fusion and memory work, while adding Vera CPUs for CPU-heavy parts of agent systems. That makes this a broader interoperability bet than a conventional cloud-capacity reservation. The next evidence to watch is product availability and deployment milestones: the 2 million-GPU commitment is scheduled for 2027–2028, while the practical question is whether Nvidia GPUs and Trainium arrive as a workable shared rack-scale option.
Loading discussion...
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.