AWS and NVIDIA Expand Partnership to 3 Million Total GPUs Including Blackwell Ultra and Rubin Architectures
Summary
- • AWS and NVIDIA announced 2M additional GPUs (Blackwell Ultra, Rubin, Rubin Ultra) for 2027–2028, bringing total NVIDIA GPU commitment on AWS to 3 million
- • The expansion follows the March 2026 GTC deal for 1M+ GPUs; surging demand prompted the accelerated and enlarged procurement
- • Partnership also adds NVIDIA Vera CPU infrastructure, NVLink Fusion with custom high-bandwidth memory, and Nemotron open models on Amazon Bedrock and SageMaker
- • AWS will build dedicated AI factories with 100,000 NVIDIA GPUs at Impact Level 6+ security for US federal and national-security workloads
Details
3M total NVIDIA GPUs committed to AWS
AWS's total NVIDIA GPU commitment now reaches 3 million: 1M+ from the March 2026 GTC announcement plus 2 million new units announced August 26, 2026, for deployment in 2027–2028.
March 2026 GTC deal exceeded by demand
The original 1M+ GPU deal announced at NVIDIA GTC in March 2026 was driven by customer demand that has since exceeded initial projections, prompting the accelerated expanded procurement.
Blackwell Ultra, Rubin, and Rubin Ultra GPU families
The 2 million new GPUs span NVIDIA's current and next-generation architectures: Blackwell Ultra (2027) and Rubin and Rubin Ultra (2028), covering two full product generation cycles.
NVIDIA Vera CPU-based infrastructure on AWS
Partnership extends beyond GPUs to NVIDIA's Vera CPU architecture, enabling full-stack NVIDIA compute — both GPU and CPU — natively on AWS cloud infrastructure.
NVLink Fusion with custom high-bandwidth memory
The collaboration extends NVIDIA NVLink Fusion interconnect with custom high-bandwidth memory (HBM) integration, enabling tighter GPU-CPU coupling for demanding AI workloads on AWS.
Nemotron open models on Bedrock and SageMaker
NVIDIA's Nemotron open-source model family will be available through AWS's managed ML services — Amazon Bedrock and SageMaker — broadening access for enterprise customers.
100,000-GPU AI factory for US government (IL6+)
AWS will build dedicated AI factory infrastructure with 100,000 NVIDIA GPUs at Impact Level 6 and above, serving federal and national-security workloads with sovereign-grade cloud security.
Jensen Huang: 'unprecedented pace and scale' for agentic AI
NVIDIA CEO Jensen Huang stated the expanded partnership aims 'to make agentic and physical AI real at an unprecedented pace and scale,' framing the deal as foundational to next-generation AI workloads.
Sources: AboutAmazon, Bloomberg, GlobeNewswire (NVIDIA), WSJ; details via Grok live web research (August 26, 2026).
What This Means
AWS's commitment to 3 million NVIDIA GPUs — tripling in under six months — underscores the intensity of hyperscaler AI compute demand and cements NVIDIA's dominance as the cloud infrastructure backbone for frontier AI. Securing Rubin-generation GPU supply through 2028 means AWS is locking in next-generation hardware before it ships, establishing a durable competitive moat in AI cloud capacity. The expansion beyond GPUs into Vera CPUs, NVLink Fusion, and managed open-model services signals a shift from ad-hoc procurement to deeply integrated full-stack NVIDIA compute environments. The 100,000-GPU sovereign AI factory for US government workloads at Impact Level 6 further entrenches AWS as the default cloud for federal AI — a strategic position with significant national-security and long-term commercial implications.
Sentiment
Excited about hyperscaler demand surge and NVIDIA's lock-in, with detailed analysis of scale and implications
“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... One customer. Two years. That's the multi-year visibility I keep hearing NVIDIA doesn't have.”
“+2 MILLION GPUs and use some NVDA CPU infrastructure... At GTC in March, AWS said it would add >1 million NVIDIA GPUs starting in 2026. This new 2 million is incremental, implying >3 million gross additions across 2026–2028... >4 GW IT... >$100B to NVDA”
“JUST IN: AWS TO DEPLOY ADDITIONAL 2 MILLION $NVDA GPUS. Deployment will be NVIDIA Blackwell Ultra, Rubin and Rubin Ultra GPUs across its global infrastructure in 2027–2028, including AI factories”
“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)”
Split
~80/20 positive/neutral on demand intensity and NVIDIA dominance; limited debate on alternatives or risks in the posts found.
Sources
- AWS and NVIDIA deepen strategic collaboration to accelerate AI from pilot to productionAws
- AWS and NVIDIA to Deliver 2 Million Additional GPUs and Next-Generation Infrastructure for Agentic and Physical AI - NVIDIA NewsroomNvidianews
- 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 MarchBloomberg
- Amazon just tripled its order of Nvidia chips over ‘surging demand’TechCrunch