Amazon is adding 2 million more Nvidia GPUs to AWS data centers in 2027 and 2028, tripling a commitment it made just five months ago to deploy more than 1 million Nvidia chips starting this year. The new order covers Blackwell Ultra, Rubin, and Rubin Ultra parts and was announced Wednesday alongside Nvidia's quarterly earnings. Neither company disclosed financial terms, but at prevailing GPU unit costs the deal is worth tens of billions of dollars.
The expansion arrives in the same quarter Nvidia reported $96.2 billion in sales, with $89 billion coming from its data center segment, up 117% year-over-year. Nvidia guided to $108 billion in Q3 revenue, some of it from Rubin GPUs that began production shipments this quarter. Amazon and Nvidia said in a joint statement that "surging demand from startups, enterprises, AI labs, and even governments" drove the decision to expand the relationship.
“demand has exceeded those expectations.”— Nvidia, company statement
Nvidia said that since the March deal, "demand has exceeded those expectations." That is the throughline of the announcement: a customer that already agreed to 1 million GPUs came back within half a year and asked for two million more.
Key facts
- 01Amazon is adding 2 million Nvidia Blackwell Ultra, Rubin, and Rubin Ultra GPUs to AWS data centers in 2027 and 2028.
- 02The order triples a commitment made five months ago for more than 1 million Nvidia GPUs starting this year.
- 03Nvidia posted $96.2B in Q2 sales, with $89B in data center revenue, up 117% year-over-year.
- 04Nvidia guided to $108B in Q3 revenue and has committed $279B to secure supply, up from $119B last quarter.
- 05Amazon's custom-chip business crossed a $25B annualized run rate on $225B in total commitments from labs including Anthropic and OpenAI.
The deal is not just about GPUs. Nvidia CFO Colette Kress said Vera CPUs will also ship to AWS, "some integrated with Rubin, others standalone," and that Nvidia expects Vera to reach "every major hyperscaler, neocloud, AI lab, and system OEM," with early shipments going to lead partners including Oracle and SpaceXAI. CEO Jensen Huang has previously described Vera as a "brand new $200 billion TAM" for the company, a claim he made in May.
Amazon is also adopting Nvidia's full physical-AI stack for its warehouse robotics fleet, including Omniverse for simulation and digital twins, Cosmos for world modeling, Isaac for robotics development, and Jetson for on-robot compute. On the enterprise side, AWS will serve Nvidia's Nemotron open models through Amazon Bedrock and Amazon SageMaker, giving Nvidia distribution inside AWS's managed AI platforms.
The striking part is that Amazon is expanding its Nvidia dependency at the same time it is trying to reduce it. AWS AI chief Peter DeSantis has said the company is in talks to sell its Trainium chips, a direct alternative to Nvidia's H100 and Blackwell parts, to outside customers. Amazon's Arm-based Graviton CPU targets the same server sockets as Intel and AMD. On its last earnings call, Amazon said its custom-chip business crossed a $25 billion annualized revenue run rate, backed by $225 billion in total commitments from labs including Anthropic and OpenAI.
Even with that momentum, AWS is buying more Nvidia silicon, not less. The economics behind the decision are the ones Huang laid out on the call.
Whether that translates into returns for AWS's customers is the open question. AI companies are pouring hundreds of billions of dollars into infrastructure on the bet that each additional GPU produces more profitable inference than it costs to run. Nvidia has committed $279 billion to secure supply and manufacturing capacity for current and future data-center projects, up sharply from $119 billion last quarter. That total includes $92 billion in projected spending for the rest of the fiscal year and another $87 billion for fiscal year 2028.
Nvidia's own Q2 report, which we covered earlier this week when the company guided to $108 billion for Q3, already showed data center revenue more than doubling year-over-year. The Amazon expansion suggests the pipeline through 2028 is being locked in now, not negotiated later.
The skeptics' case is that this is a circular trade. Hyperscalers commit tens of billions to Nvidia; Nvidia commits hundreds of billions to fabs and memory suppliers; AI labs commit hundreds of billions to hyperscalers. If token demand from paying customers slows before Rubin Ultra ships in 2028, the whole chain revalues at once. Amazon's parallel bet on Trainium is a hedge against exactly that scenario, though a hedge that so far is not slowing its Nvidia orders.
For Nvidia, the AWS expansion is the clearest signal yet that the largest cloud provider building its own accelerators is not, in practice, treating that program as a substitute for Nvidia at the frontier. For Amazon, tripling the order five months in is an admission that internal silicon cannot close the capacity gap fast enough to serve the workloads Anthropic and other tenants are running today. The 2027 and 2028 delivery windows mean the biggest test of whether AI compute converts to profit at this scale is still two years out — and both companies are pricing in that it will.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
