Amazon Adds 2M More Nvidia GPUs After 5-Month Demand Surge

Amazon and Nvidia announced an expanded partnership on August 26, adding another 2 million Nvidia GPU chips to AWS infrastructure—tripling the original commitment made just five months earlier to deploy more than 1 million GPUs. Nvidia's stated explanation was unambiguous: demand since that earlier agreement "exceeded those expectations." The incoming chips span three generations: Blackwell Ultra, Rubin, and Rubin Ultra, with delivery scheduled across 2027 and 2028. Neither company disclosed financial terms, though Nvidia characterised the deal as worth tens of billions of dollars.

The deal extends well beyond raw GPU units. Nvidia CFO Colette Kress confirmed that an unspecified number of Vera CPUs—some integrated with Rubin, others standalone—will accompany the GPU shipment. Nvidia's full physical AI stack is also being integrated across AWS: Omniverse (simulation and digital twin), Cosmos (world foundation models), Isaac (robotics development), and Jetson (edge AI compute). On the model side, AWS will serve Nvidia's Nemotron family of open models through both Amazon Bedrock and SageMaker. The partnership also extends to Amazon's warehouse robots, with Amazon adopting Nvidia's physical AI stack to power that fleet.

GPU Generations and the Vera CPU

The three GPU families represent distinct positions in Nvidia's roadmap. Blackwell Ultra is an evolutionary step on the Blackwell architecture already shipping to hyperscalers. Rubin is the next full architectural generation; Nvidia confirmed production shipments began in Q3 2026, meaning investor scrutiny of Rubin's initial revenue contribution is already underway. Rubin Ultra is the high-end tier of that generation. Delivery to AWS spans 2027 and 2028, so engineering teams should not expect broad availability before mid-2027 at the earliest.

The Vera CPU pairing is strategically significant. Jensen Huang characterised Vera CPUs as a new $200 billion total addressable market when previewing the product in May. Kress added that every major hyperscaler, neocloud, AI lab, and system OEM is expected to deploy Vera, with Oracle and SpaceXAI named as lead partners already receiving shipments. The CPU-GPU interconnect architecture this enables is central to systems-level efficiency gains that are increasingly rivalling raw scaling.

Capacity and Supply Context

For teams managing training budgets or SLA commitments on AWS, the five-month acceleration from 1 million to 3 million total committed GPUs carries a specific implication: AWS was capacity-constrained faster than its own forecasting predicted. Nvidia's supply-side posture reinforces this. The company has committed $279 billion to secure manufacturing and supply capacity—up from $119 billion last quarter—including $92 billion projected for the remainder of fiscal 2026 and another $87 billion for fiscal 2028. That trajectory means Nvidia is itself in an aggressive capacity-lock posture, which historically compresses spot availability and softens discounting for buyers who did not pre-commit.

Amazon is not entirely dependent on Nvidia. Its Trainium chips—positioned as a direct alternative to Nvidia's H100 or Blackwell chips for deep learning workloads—are now being offered to external data center operators, per AWS AI chief Peter DeSantis. Amazon's custom silicon business crossed a $25 billion annualized revenue run rate on its last earnings call, supported by $225 billion in total commitments from AI labs including Anthropic and OpenAI. But the scale of the Nvidia re-order makes clear that Trainium is a complement to, not a replacement for, Nvidia capacity in the near term.

Nvidia's Financial Backdrop

Nvidia posted Q2 revenue of $96.2 billion, beating analyst estimates. Data center revenue reached $89 billion—up 117% year-over-year—and the company guided Q3 revenue to $108 billion, some of which will come from initial Rubin shipments. Those figures provide the financial context for why Nvidia's supply chain dominance continues to shape every significant infrastructure decision in the industry.

The broader implication for builders evaluating AWS AI workloads: hyperscaler demand is outpacing recently-revised internal forecasts, compressing the window to lock in favourable capacity terms before 2027 delivery queues fill. For teams assessing whether to architect around managed AWS AI services or pursue pipeline-level efficiency as an alternative to raw compute acquisition, Rubin-generation availability on AWS is a 2027 story at best—and reserved capacity for that window is already being consumed at the hyperscaler level now.