Amazon just tripled its order of Nvidia chips over 'surging demand' | TechCrunch

Amazon has tripled its order of Nvidia GPUs for AWS, committing to add 2 million more chips on top of the 1 million it had already purchased just five months ago — a 3× reorder that signals the AI infrastructure arms race is still accelerating, not plateauing. The new chips span Nvidia's Blackwell Ultra, Rubin, and Rubin Ultra generations, with deployment scheduled across 2027 and 2028. Nvidia's Q2 revenue hit $96.2 billion, with data center alone accounting for $89 billion — up 117% year-over-year — and the company guided Q3 to $108 billion.

The Claim

The core of the announcement is scale and speed: Amazon went from 1 million committed GPUs to 3 million in roughly five months, citing "surging demand" from startups, AI labs, and government customers pouring workloads onto AWS. The new order covers not just Blackwell Ultra — the current high-end Nvidia chip — but also Rubin and Rubin Ultra, the next-generation architectures that entered production in Q3 2026. That puts Amazon in line to receive bleeding-edge silicon the moment supply ramps up, rather than waiting in queue behind slower buyers.

The deal extends well beyond GPU procurement. Nvidia's Vera CPU line — which ships both standalone and integrated into Rubin systems — is now heading to AWS alongside every other major hyperscaler. Nvidia's Nemotron family of open models is coming to Amazon Bedrock and SageMaker, adding a distribution path that previously ran only through OpenRouter and Nvidia's own NIM service. And Nvidia's full physical AI stack — Omniverse, Cosmos, Isaac, and Jetson — is being adopted for Amazon's warehouse robotics fleet, meaning the partnership touches Amazon's logistics operation, not just its cloud business.

On the supply side, Nvidia disclosed it has committed $279 billion to its own manufacturing and supply chain, up from $119 billion the previous quarter. That jump is the company's attempt to get ahead of the demand it is already booking. Multiple reports confirm the TechCrunch account of the deal's scope, and a secondary source aggregating the story noted that the expanded partnership "indicates significant capital commitment beyond hardware procurement as enterprises scale AI capabilities" — a fair summary of what both sides are signaling publicly.

What We See

Our read is that the 3× reorder in five months is the more meaningful data point than the raw GPU count. Buying 2 million more chips is large, but large orders are routine for hyperscalers. What is not routine is revising a commitment upward by that magnitude over a five-month window. That pace suggests Amazon's internal demand forecasting failed to anticipate how fast enterprise AI workloads were materializing — which, in turn, means the customers placing those workloads are themselves growing faster than expected.

The Nemotron-on-Bedrock piece deserves more attention than it is getting in the initial coverage. Nvidia's Nemotron open model family moving into Bedrock and SageMaker transforms Nvidia from a hardware vendor into a model distribution partner for Amazon. That is a structural shift: Nvidia now has a software presence inside Amazon's managed AI stack, not just a hardware slot in its data centers. For developers, this matters because Bedrock is where many teams already run their inference — adding Nemotron there means less friction to try it, which could meaningfully accelerate enterprise adoption of open models.

The Trainium development is equally notable and somewhat underreported. Amazon is now selling its own Trainium chips — the custom AI accelerators it built to reduce its Nvidia dependence — to third parties. That makes Trainium a direct, if currently distant, competitor to Nvidia's H100 and Blackwell lines. Amazon is simultaneously writing its largest Nvidia check ever and seeding the market for its own alternative. The fact that both moves are happening at the same moment reflects a mature procurement strategy: lock in supply from the market leader while building a hedge.

Where It Falls Short

The "surging demand" framing in the announcement deserves scrutiny. Nvidia's data center revenue growing 117% year-over-year is real, and the Q3 guidance of $108 billion is not a number a company invents. But the composition of that demand — how much comes from productive inference workloads versus speculative capacity buildouts that may never fully utilize the hardware — is not disclosed. Hyperscalers have historically over-ordered infrastructure during hype cycles and then written down underutilized capacity; the 2026 GPU order wave could follow the same pattern, though there is no public evidence yet that utilization is lagging.

The 2027–2028 deployment timeline for the new chips also introduces meaningful uncertainty. Rubin and Rubin Ultra are next-generation architectures that only entered production in Q3 2026. Nvidia has a reasonably strong track record of hitting its own supply milestones in recent years, but a two-year delivery window is long enough that demand conditions, competing chip options, and Amazon's own Trainium roadmap could all shift substantially before the hardware arrives. The deal is a commitment, not a guarantee, and neither Amazon nor Nvidia disclosed what, if any, cancellation or adjustment provisions it includes. For developers watching GPU-backed API pricing, the honest answer is that this order secures Amazon's infrastructure position for the back half of the decade — but it does not tell you what your inference bill will look like in 2027.

Sources