Amazon has tripled the Nvidia order it placed four months ago. The company committed to 2 million more Nvidia GPUs in late August, up from the more than 1 million it announced in May, in a deal TechCrunch reports as worth tens of billions of dollars. Exact terms were not disclosed.
What is in the order
The commitment spans three Nvidia generations — Blackwell Ultra, Rubin and Rubin Ultra — along with Vera CPUs. The chips are destined for AWS data centres in 2027 and 2028, which places most of this order two years out from the announcement.
Nvidia said demand has exceeded the expectations it held at the time of the May agreement.
Why the timeline is the story
A two-year deployment horizon on hardware ordered today tells you more about the constraint than the headline number does. Chips ordered in 2026 for racking in 2028 are not solving a 2026 capacity problem; they are reserving a place in a supply queue that is now long enough to require forward commitment at this scale.
That is a different market from one where compute is bought as needed. It rewards balance-sheet size over technical merit, and it locks buyers into an architecture years before they know what they will be running on it.
Amazon’s own silicon has not gone away
The same reporting puts Amazon’s custom chip business past a $25 billion annualised revenue run rate, and cites $225 billion in total commitments from AI labs including Anthropic and OpenAI.
Those two facts sit together awkwardly and are worth holding at once: Amazon is simultaneously scaling its own accelerators and placing one of the largest third-party GPU orders on record. The order is not a retreat from Trainium — it is a hedge, at a scale that only a handful of companies can write.
Huang’s framing
Nvidia chief executive Jensen Huang put the demand case in commercial terms: “AI is generating profitable tokens … If we had more compute, we could generate more profitable tokens.”
That is the argument the whole buildout rests on — that inference demand is revenue-generating rather than speculative. It is also, so far, largely an argument rather than a disclosed set of numbers. Neither Nvidia nor its largest customers publish per-token margins.
What to watch
Whether the 2027–2028 deliveries hold their schedule, and whether power availability rather than silicon becomes the binding constraint. A GPU that cannot be energised is not capacity.
