Amazon just expanded a massive GPU commitment with Nvidia, what it actually means
Nvidia told investors on its quarterly earnings call that Amazon Web Services (AWS) will receive another 2 million Nvidia GPUs, naming Blackwell Ultra, Rubin and Rubin Ultra as part of the deal, with shipments targeted for 2027-2028. The announcement builds on a prior commitment reported five months earlier for more than 1 million GPUs that began rolling out this year, and Nvidia said “demand has exceeded those expectations.”
That headline number matters because it is not just more silicon. Nvidia said the expanded agreement covers chips plus software, networking, model support and robotics tools, and that some Vera CPUs will ship to AWS both integrated with Rubin GPUs and as standalone units (comments attributed to Nvidia management on the call).
What the companies said, in plain language
- Nvidia stated it will supply an additional 2 million GPUs to AWS (Blackwell Ultra, Rubin, Rubin Ultra), expected across 2027 and 2028, and that this follows an earlier commitment for more than 1 million GPUs announced several months before the call.
- Nvidia said it will also provide Vera CPUs (some paired with Rubin GPUs and some standalone), networking hardware, model tooling and data‑processing software, and its robotics stack, Omniverse (simulation), Cosmos/NeMo toolsets, Isaac (robotics) and Jetson (edge modules), for AWS to use.
- Nvidia told investors that demand has exceeded prior expectations. Colette Kress, Nvidia’s CFO, characterized Vera as targeted for deployment across hyperscalers, AI labs and OEMs.
- AWS will surface Nvidia’s open‑model family via its managed services (Bedrock and SageMaker), and the two firms said the expanded partnership will touch cloud, enterprise and robotics use cases.
Financial specifics were not disclosed on the call. Nvidia also spoke about securing manufacturing capacity and supply to meet demand. Amazon continues to advance its own in‑house silicon programs (Trainium and Graviton) even as it expands this relationship with Nvidia.
Why AWS would buy millions of Nvidia GPUs while building its own chips
On the surface it looks contradictory: why buy huge volumes from a supplier when you are growing your own silicon business? Three business realities explain the mix.
- Workload fit: High‑end training workloads and many production LLMs run on Nvidia’s mature GPU stack and CUDA‑optimized toolchains. For massive model pretraining and peak throughput, Nvidia hardware is often the pragmatic choice.
- Software and ecosystem: Many third‑party models, libraries and enterprise stacks are validated primarily on Nvidia platforms. Offering Nvidia instances reduces friction for customers who expect compatibility.
- Supply management and hedging: The market for advanced GPUs is capacity constrained. Large forward commitments and a mixed procurement strategy reduce risk, letting AWS secure capacity for surges while it scales and monetizes its own Trainium and Graviton offerings for other workloads.
Beyond chips: why the software and robotics pieces matter
This agreement isn’t purely about raw FLOPS. Nvidia’s pitch, and what AWS appears to be buying into, is an integrated stack: high‑performance GPUs and CPUs in the cloud, model toolkits for training and serving (NeMo/related model families), simulation and digital‑twin platforms (Omniverse), robotics development (Isaac), and edge modules (Jetson).
Coupling simulation, cloud training and edge hardware shortens the loop from model design to safe, repeatable deployment in physical environments like warehouses. For enterprises building autonomous systems, that cloud-to-edge continuity can shave months off development cycles and reduce costly real‑world mistakes.
Three strategic industry implications
- Demand appears structural. Management comments point to sustained demand from startups, enterprises, AI labs and governments. A hyperscaler preparing to absorb millions of GPUs signals a long‑term appetite for large‑scale model training and inference.
- Multi‑vendor compute menus will persist. Hyperscalers will continue to offer both vendor hardware (Nvidia) for performance‑sensitive workloads and in‑house silicon (Trainium/Graviton) for cost‑sensitive or differentiated offerings.
- Robotics and edge AI will accelerate. When cloud providers combine large training fleets, simulation platforms and accessible edge modules, enterprises can iterate faster on physical AI, from logistics automation to inspection robots.
What remains unclear
- Exact commercial terms and unit pricing for the GPU commitment were not disclosed.
- Precise delivery and availability schedules, and how the 2027-2028 shipment targets align with statements about earlier rollouts, require the call transcript or press releases to reconcile.
- The exact volumes and timing for Vera CPU deliveries to AWS were not specified on the call.
- Operational details on how Nvidia’s robotics stack will be rolled into Amazon’s warehouse programs (pilot vs. large rollout) were not provided.
- How AWS will price and position Nvidia‑backed instances versus Trainium/Graviton offerings in competitive terms remains a commercial planning question.
Practical next steps for business leaders (30/60/90 day checklist)
- 30 days, Audit your CUDA exposure: Inventory models, libraries and pipelines that are CUDA‑dependent. Identify high‑risk components that would be costly to port.
- 60 days, Benchmark cost vs. performance: Run targeted cost‑performance tests: pretraining on high‑end Nvidia instances, inference on both Nvidia and alternative instances (Trainium/Graviton) to see where each is competitive for your workloads.
- 90 days, Secure flexible procurement: Negotiate cloud contracts with options for reserved capacity and flexibility clauses to capture spot/reserved pricing as market capacity evolves. Build multi‑vendor deployment plans to avoid single‑supplier bottlenecks.
- Ongoing, Invest in simulation-first robotics dev: If you’re deploying physical AI, prioritize simulation and digital‑twin workflows to reduce deployment risk and accelerate validation cycles when moving from cloud prototypes to edge robots.
Short, direct answers to the key questions
-
Did Amazon actually triple its Nvidia order?
According to Nvidia’s remarks on its earnings call, AWS added another 2 million Nvidia GPUs after an earlier, separate commitment for over 1 million GPUs made months earlier. Taken together, the incremental commitment represents a very large expansion of planned Nvidia capacity for AWS.
-
Will this make Nvidia the only supplier AWS uses for training?
No. AWS is scaling both external Nvidia capacity and its own silicon programs (Trainium and Graviton). Expect AWS to match workload type to the most appropriate silicon, Nvidia for peak training performance and ecosystem compatibility; in‑house chips for cost‑sensitive inference and other workloads.
-
How big is the deal in dollar terms?
Neither company disclosed financial terms on the call. Public reporting and analyst commentary have offered estimates; without confirmed pricing or contract details, any dollar figure should be treated as an estimate rather than a disclosed fact.
-
Does this change the outlook for robotics and edge AI?
The deeper cloud+simulation+edge integration being described implies a faster cadence for robotics development and deployment. If AWS adopts Nvidia’s simulation and edge stack broadly, enterprises building robots can expect shorter validation cycles and fewer integration surprises.
Bottom line
Whether you focus on infrastructure, product or operations, treat the announcement as a procurement and timing signal: hyperscalers are planning for sustained, large‑scale model workloads and are securing both compute and the surrounding software ecosystem. That does not end competition between vendor silicon and hyperscaler‑built chips. It makes multi‑vendor strategies and portability planning more urgent. Big forward orders do more than buy chips, they shape where teams build, which models scale, and how quickly businesses can move from prototype to production.