The announcement arrived in the middle of Nvidia Corp.’s earnings call, a few sentences from Amazon’s cloud division that rewrote the day’s narrative. Amazon Web Services will add another two million Nvidia GPUs to its infrastructure, spanning the Blackwell Ultra, Rubin, and Rubin Ultra generations, with delivery spread across 2027 and 2028. The commitment, disclosed by Nvidia on the call, is worth tens of billions of dollars on a per-chip basis, analysts estimated, and it roughly triples the scale of what AWS had pledged only five months ago.
The order builds on a commitment Amazon made in March, when AWS said it would deploy more than a million Nvidia GPUs. Nvidia said demand since then “has exceeded expectations,” and the new order reflects a customer that has moved from buying chips as needed to securing them years in advance. Executives on both sides described the arrangement in supply-chain terms, delivery windows and capacity guarantees, but the strategic meaning is simpler: Amazon has become one of Nvidia’s largest buyers, a shift from its historical role as a customer that rented out Nvidia’s latest hardware by the hour.
The role reversal is the story the market is still digesting. Cloud providers have always been intermediaries, buying compute in bulk and selling it in slices. What has changed is the scale of the buying and the reasons behind it. AWS now needs chips not only to rent to customers but to guarantee its own position in an arms race where Microsoft, Google, and Meta are signing similar multi-year orders. The company that pioneered renting other people’s hardware is now stockpiling its own, and its data-center footprint is being built around commitments that run years into the future.
The numbers help explain why. Nvidia’s data center revenue reached $89 billion in a single quarter, with hyperscalers accounting for $48.7 billion of that. Those buyers are effectively placing orders that Nvidia’s factories cannot fill fast enough, which is why the company has begun guaranteeing land, power, and leases for its customers rather than just selling them chips. Amazon’s two-million-GPU order is one of the largest single commitments in that pipeline, and it tells competitors that AWS intends to be the compute supplier of first resort for the AI boom.
The deal also says something about how the AI industry is being financed. None of the hyperscalers carries enough cash to buy this much hardware outright, so the orders are structured as forward commitments, leases, and financing arrangements that push the cash burden onto the future. Nvidia itself has created financing platforms with Apollo, BlackRock, and KKR that could mobilize more than $500 billion for AI infrastructure, and it has guaranteed up to $108.5 billion in exposure on land, power, and shells for its biggest customers. Amazon’s order sits inside that system: a promise to pay for chips that do not exist yet, backed by revenue that has not been earned yet.
The deal also reshapes the competitive balance among cloud providers. Microsoft and Google have both been spending heavily on their own AI infrastructure, and Meta has committed to building out data centers at a similar pace. Amazon’s two-million-GPU order moves it from following that wave to leading it, at least in quantity, and it gives AWS a chip position that rivals will have to match or explain. Cloud customers, meanwhile, get a choice between providers with deep Nvidia fleets and providers betting on custom silicon, and the price competition between those two models will shape margins across the industry for years.
For the broader industry, the order is a signal about capacity. Blackwell Ultra, the current flagship, is still ramping, and Rubin, its successor, is expected to arrive next year, with Rubin Ultra to follow. AWS’s decision to commit to all three generations at once suggests the company expects demand for AI compute to keep growing for years, and that it wants its data centers to be able to serve the newest chips the moment they ship. The risk is the one that has haunted every capital-spending boom: that capacity built today becomes a cost problem if demand softens.
Power is the constraint that most concerns analysts. Data centers large enough to absorb two million of Nvidia’s most power-hungry chips need gigawatts of electricity, and utilities are already struggling to supply the buildouts underway. Amazon has been buying renewable energy and building its own power capacity, but the pace of AI demand is testing the limits of the grid in every region where the company operates. Executives said the order is scheduled around power availability, which is why delivery runs through 2028 rather than arriving in a single surge.
For investors, the announcement answers a question that has hung over Nvidia’s outlook: is the demand real, or is it a bubble of orders that can be canceled? Amazon’s commitment, made public in the middle of a regulated earnings call, is as firm a signal as the market gets. The company has little incentive to announce capacity it does not intend to use, analysts said, and the multi-year structure means Nvidia can plan its supply chain with unusual confidence. The two million chips are now part of both companies’ forward books, and the AI infrastructure cycle, whatever its eventual shape, will run through them.
