AWS and Nvidia announced on 26 August 2026 an expansion of their partnership by 2 million more GPUs bound for Amazon data centres across 2027–2028. It triples the May order, which covered just over a million chips. Financial terms were not disclosed.
Key takeaways
- 2 million additional GPUs deploying across 2027–2028
- Blackwell Ultra, Rubin and Rubin Ultra chips plus Vera CPU-based infrastructure
- 100,000 GPUs reserved for US federal and national-security workloads at IL6 and above
- Analysts put the deal in the tens of billions of dollars
- Amazon Robotics is adopting Nvidia’s physical AI platform
Three chip generations at once
The order does not cover a single chip family. It spans Blackwell Ultra, the current generation, plus Rubin and Rubin Ultra — successors only now entering production. On top of that comes an expansion of the existing Blackwell fleet, including RTX PRO 4500 Blackwell Server Edition for Amazon EC2 G7 instances.
| Chip | Status | Role in the deal |
|---|---|---|
| Blackwell Ultra | Current generation | Expansion of the existing fleet |
| Rubin | Entering production | Bulk of 2027–2028 deliveries |
| Rubin Ultra | Rubin’s successor | Final deployment phase |
| Vera | CPUs | Infrastructure accompanying the GPUs |
Spreading the purchase across three generations makes operational sense. AWS is not waiting on one launch date — it adds compute as each chip comes off the line.
The deal reaches beyond silicon
Alongside the GPUs come Vera CPUs, NVLink Fusion with NVHBM memory and Spectrum networking tuned for GPU clusters. Open Nemotron models arrive on Amazon Bedrock and Amazon SageMaker AI, while the cuDF and cuVS libraries accelerate data processing.
The physical strand is the most interesting one here. Amazon Robotics is taking on NVIDIA’s robotics stack — Jetson, NVIDIA Omniverse and NVIDIA Isaac. Amazon’s warehouses become another proving ground for that platform.
NVIDIA and AWS have built one of the great growth engines of the AI era, and demand is running ahead of every forecast.
Jensen Huang, founder and CEO of Nvidia.
Why it matters
The scale of the purchase says more about confidence in demand than about the technology. Committing three years ahead, to chips not yet in volume production, means AWS treats a compute shortage as a bigger risk than a surplus. For rival clouds, it signals that the window for locking in cheap supply is closing.
What’s next?
- Deployments are scheduled for the third quarter of 2027 and through 2028
- The 100,000 government GPUs require IL6?IL6: A US Department of Defense security accreditation level for the most sensitive government workloads.-level infrastructure, meaning separate isolated regions
- No deal value was given, so its effect on AWS margins cannot be estimated
