AWS plans to add 2M Nvidia Blackwell Ultra, Rubin, and Rubin Ultra GPUs to its data center fleet in 2027 and 2028, in addition to 1M GPUs announced in March
Bloomberg
Context & Ripple Effects
AWS had already set a 2027 procurement baseline through its planned purchase of 1 million Nvidia GPUs, alongside other Nvidia chips. The additional deployment plan extends that commitment across multiple accelerator generations rather than concentrating it in a single product cycle.
The sequencing follows Nvidia's annual AI-accelerator roadmap, which set out Blackwell Ultra and Rubin as successive platforms. It gives AWS a multiyear basis for planning data-center capacity around Nvidia's release cadence.
First-order effects
AWS expands its planned Nvidia GPU fleet by 2 million units for 2027 and 2028, adding Blackwell Ultra, Rubin, and Rubin Ultra systems beyond the 1 million GPUs announced in March.
Nvidia gains planned demand from AWS across three successive accelerator families, rather than a commitment limited to one generation.
Second-order effects
AWS must align its data-center buildout and deployment schedule with Nvidia's product transitions, making the supplier's generation timing more consequential to AWS capacity planning.
Nvidia's ability to supply successive platforms becomes a central execution dependency for AWS's larger fleet plan, increasing the value of a coordinated hardware roadmap.
Third-order effects
If hyperscalers continue reserving capacity across multiple chip generations, AI compute procurement will increasingly resemble a multiyear capacity market rather than opportunistic purchases of a single accelerator.
The arrangement points toward cloud infrastructure being planned around integrated, recurring hardware roadmaps, with leading suppliers gaining influence over when new capacity reaches customers.
The trend:AI infrastructure buyers are moving toward multigeneration GPU commitments that tie cloud-capacity expansion to semiconductor roadmaps.
BIG NEWS from @awscloud and @nvidia 🤝 We're announcing a major expansion of our strategic collaboration for next-gen AI infrastructure. ⬇️ Building on 16 years of joint innovation, AWS will deploy 2 million additional NVIDIA GPUs across its global infrastructure in
AWS and NVIDIA just announced 2M *more* GPUs for 2027 and 2028. Blackwell Ultra, Rubin, Rubin Ultra. Plus Vera CPUs on AWS, 100K GPUs for federal AI factories, and NVHBM going into Trainium racks. Is 2M is a big number? It is. But here's what matters more. AWS committed to 1M+
Excited to announce a major expansion of our partnership with @JensenHuang and @nvidia . 2M additional GPUs, AI factories with 100K GPUs for the U.S. Government, and more. Demand for reliable AI infrastructure continues to accelerate and we're building to keep pace. @awscloud
NEW On @MadMoneyOnCNBC, $NVDA CEO Jensen Huang added color on the AWS deal: beyond the 2 million GPUs, Amazon's also buying “millions of CPUs” plus using Nvidia tech for robotics. $AMD $INTC
NVIDIA and @awscloud are expanding the partnership across the full stack — GPUs, CPUs, networking, open models and software — to make agentic and physical AI real at a pace and scale that only we can deliver together. What that looks like: ➡️ 2 million additional NVIDIA GPUs
The moral of the $NVDA deal with $AMZN to deploy 2 million additional GPU + Vera should be just another indicator of how silly the zero sum debates are. It won't be 100% NVIDIA and it won't be 100% Trainium. Supply is constrained everywhere and demand for AI is insatiable. It's
NVIDIA 💚AWS We're announcing a major expansion to our partnership with AWS. More GPUs, better integration with AWS, and more collaboration on everything from NVLink to Amazon Bedrock. Nemotron models have benefited greatly from AWS infrastructure and I am so glad to see us