AWS and Nvidia said on Wednesday, August twenty sixth, 2026 that AWS will deploy two million additional Nvidia GPUs across 2027 and 2028, covering Blackwell Ultra, Rubin and Rubin Ultra parts. The commitment lands five months after AWS agreed at GTC 2026 to add more than a million GPUs starting this year, and Nvidia says demand has run past those expectations since. The announcement is broader than silicon volume: it also covers Vera CPUs, networking, robotics platforms and Nemotron open models delivered through Bedrock and SageMaker. Nvidia disclosed it alongside quarterly results showing eighty nine billion dollars of data centre revenue.
The short answer
AWS and Nvidia announced on August 26, 2026 that AWS will deploy two million additional Nvidia GPUs across 2027 and 2028, spanning Blackwell Ultra, Rubin and Rubin Ultra. That is roughly triple the commitment made at GTC 2026, which covered more than one million GPUs starting this year. Nvidia says demand has exceeded those expectations. The expansion also covers Vera CPUs, some integrated with Rubin and some standalone, networking, the Omniverse, Cosmos, Isaac and Jetson platforms, and Nemotron open models served through Amazon Bedrock and SageMaker. Financial terms were not disclosed.
Five months is a short time to triple a number that was already large. In March, AWS said it would add more than a million Nvidia GPUs. On Wednesday it said it would add two million more on top, and Nvidia explained the revision in one sentence: demand has exceeded those expectations.
What is actually in the commitment
The two million GPUs are scheduled for 2027 and 2028, across AWS global infrastructure including the sites Nvidia now calls AI factories. Three parts are named: Blackwell Ultra, Rubin, and Rubin Ultra.
That spread is the detail worth noticing. A commitment covering one product generation is a purchasing decision. A commitment covering a generational transition is a capacity plan, because it has to survive whatever the Rubin ramp actually looks like. Vera Rubin is the platform Nvidia has been building the rest of its rack scale story around, and it is the same silicon showing up in the decode accelerator Nvidia put into full production this week and in Anthropic's West Virginia lease.
Neither company disclosed what it costs. Outside estimates put it in the tens of billions of dollars, which is consistent with the scale but is not a number either party has confirmed.
The half of the announcement that is not GPUs
Read past the headline figure and this is a co engineering agreement rather than a supply contract.
Nvidia will ship an unspecified number of Vera CPUs, its Arm based server processor for the Rubin generation. Colette Kress, Nvidia's chief financial officer, said some will arrive integrated with Rubin and others standalone. The standalone case is the interesting one, because it puts Nvidia into general purpose server CPU territory inside a cloud that has spent a decade building Graviton for exactly that slot.
The rest spans networking, AI factory design, data processing, and robotics through the Omniverse, Cosmos, Isaac and Jetson platforms. Nvidia's Nemotron open models will be served through Amazon Bedrock and SageMaker. AWS is also expanding current Blackwell capacity, including RTX PRO 4500 Blackwell Server Edition GPUs for Amazon EC2 G7 instances.
That last item is the only part of the announcement with a near term effect on anything you can launch today.
Buying Nvidia hard while building Trainium hard
The obvious tension is that Amazon has spent years reducing its dependence on Nvidia, and has just committed to two million more of its GPUs.
Both things are true and neither is a retreat. Amazon's custom silicon business has crossed a twenty five billion dollar annualised revenue run rate, and Trainium keeps taking a share of training work, notably Anthropic's. Meanwhile AWS reports around two hundred and twenty five billion dollars in commitments from AI labs including Anthropic and OpenAI. When the forward book is that large, the question stops being which vendor you prefer and becomes how many watts you can energise and how much silicon you can physically obtain.
The same logic has played out across the industry all year. Microsoft running 2.2 million AI chips while short of the power to run them, Nvidia guaranteeing a hundred and five billion dollars for an OpenAI campus in Ohio, and three trillion dollars of AI commitments sitting off big tech balance sheets are the same story told from different sides. Capacity is the scarce good, and vendor loyalty is not really a variable.
The earnings call framing matters
Nvidia disclosed this during its quarterly results, which is a choice rather than a coincidence.
The quarter itself was large: ninety six point two billion dollars in total sales, eighty nine billion of that from the data centre segment, up one hundred and seventeen percent year on year, with guidance of roughly one hundred and eight billion for the next quarter. Announcing a named customer commitment in that setting turns an abstract demand narrative into a specific counterparty with a specific delivery window. It also sets a benchmark that other clouds now get measured against, which is generally the point.
What to do with this
If you buy or plan capacity, the practical read is about timing rather than volume. The two million GPUs land in 2027 and 2028, so they say nothing about the supply you are fighting for this quarter. What they do suggest is that the current squeeze is expected to persist long enough to justify committing two years out, which is not the shape of a market about to loosen.
If you rent, watch instance pricing on the G7 family and on whatever the first Rubin backed instances end up being called. Announcements move slowly into price lists, but price lists are where a capacity commitment becomes something you can act on.
And if you are choosing between CUDA and an accelerator with a smaller ecosystem, note that neither side of that decision got easier this week. AWS is committing heavily to both paths at once, which means it is not going to make the choice for you.
Sources and further reading
- AWS and NVIDIA to Deliver 2 Million Additional GPUs and Next-Generation Infrastructure for Agentic and Physical AI, NVIDIA Newsroom, August 26, 2026
- AWS and NVIDIA to deploy 2 million more GPUs for AI in 2027-2028, About Amazon, August 26, 2026
- Amazon just tripled its order of Nvidia chips over surging demand, TechCrunch, August 26, 2026
- AWS, NVIDIA Expand AI Infrastructure with 2 Million Additional GPUs, The Fast Mode, August 2026
Frequently asked questions
Is two million GPUs a new order or a restatement of the old one?
It is additional. At GTC 2026 in March, AWS said it would deploy more than one million Nvidia GPUs starting during 2026. This announcement adds two million on top of that, scheduled for 2027 and 2028, which is why the coverage described it as roughly tripling the order. The parts are also different. The first commitment was largely current generation Blackwell. This one names Blackwell Ultra, Rubin and Rubin Ultra, so it spans a generational transition rather than simply buying more of the same silicon. Neither company disclosed financial terms, and outside estimates put the value in the tens of billions of dollars.
Does this mean AWS is backing away from Trainium?
The numbers say the opposite. Amazon has said its custom chip business crossed a twenty five billion dollar annualised revenue run rate, and Trainium continues to take a share of internal and customer training work. What is happening is that both lines are growing at once, because the constraint is total capacity rather than vendor preference. AWS sells to customers who have written code against CUDA and will not port it, and to customers who care only about price per token and will run wherever it is cheapest. Serving both means buying Nvidia at scale while continuing to build silicon that reduces dependence on it. Those are not contradictory positions when demand exceeds everything you can rack.
What is a Vera CPU and why does it appear in a GPU announcement?
Vera is Nvidia's Arm based server CPU, designed as the host processor for the Rubin generation. Nvidia's chief financial officer Colette Kress said some will ship integrated with Rubin and others standalone. It matters because rack scale AI systems are no longer sold as accelerators you slot into someone else's server. The CPU, the interconnect and the GPU are co designed, and the host side increasingly does real work in agentic pipelines: orchestration, tool calls, retrieval, data preparation. A standalone Vera option also puts Nvidia into a part of the server market that has been contested between x86 and Graviton, which is a notable second front.
When does any of this actually reach an instance I can launch?
Not soon, for the two million. The deployment window is 2027 and 2028, and hyperscaler capacity announcements describe procurement rather than availability. The nearer term item in the same announcement is AWS expanding Blackwell capacity, including RTX PRO 4500 Blackwell Server Edition GPUs for Amazon EC2 G7 instances, which is a graphics and inference oriented tier rather than a training one. If you are planning workloads, treat the two million figure as a signal about where pricing and availability are heading over the next two years, not as capacity you can reserve now.
What does the Nvidia earnings context tell us?
It tells us the demand claim is not marketing. Nvidia reported total sales of ninety six point two billion dollars for the quarter, with data centre revenue of eighty nine billion, up one hundred and seventeen percent year on year, and guided to roughly one hundred and eight billion for the following quarter. Announcing a customer commitment during an earnings call is a deliberate act, because it converts a forward looking claim into a named counterparty. Amazon separately reports around two hundred and twenty five billion dollars in commitments from AI labs including Anthropic and OpenAI, which is the demand this capacity is meant to serve.