Amazon Nvidia GPU Deal Adds 2 Million Chips to AWS Through 2028

A purchase order that nearly triples itself in five months

Amazon Web Services and Nvidia announced on August 26 that AWS will add another 2 million Nvidia GPUs to its data centers. Nvidia disclosed the deal on its quarterly earnings call. Five months earlier, Amazon had agreed to deploy more than 1 million Nvidia GPUs across AWS infrastructure starting this year. Nvidia said demand had since gone beyond what either company expected.12 Most coverage puts the new chips in AWS facilities in 2027 and 2028. The order covers Blackwell Ultra, Rubin and Rubin Ultra parts, adding to the 1-million-plus commitment AWS made at Nvidia's GTC conference in March.120 Taken together, Amazon's announced Nvidia commitment is now above 3 million GPUs.217

The reports differ on timing. TechCrunch, citing Nvidia CFO Colette Kress, described the 2 million GPUs as arriving "starting in the third quarter."12 The joint announcement and most other outlets give a 2027–2028 window.11418 The likeliest reading is that some smaller items, such as Vera CPUs and the expanded Blackwell-based G7 instances, start shipping soon, and the large GPU volumes arrive over the following two years. In practice, the deal secures future supply. It does nothing for the compute shortage AWS customers face today.

The framing also varies. Several outlets called it a tripling of Amazon's Nvidia order.121516 That's accurate for the total commitment: 1 million becomes about 3 million. It does not mean Amazon's annual buying rate tripled, because the second tranche is spread across two calendar years.

More than GPUs: Vera, NVHBM and a shared rack

The chip count drew the headlines, but the other terms may matter more strategically. AWS will offer infrastructure built on Nvidia's Vera CPUs aimed at agentic AI workloads. The companies are also extending NVLink Fusion and adding custom Nvidia high-bandwidth memory, called NVHBM.1 Kress said some of the Vera CPUs will be paired with Rubin GPUs and others will ship on their own.12 Jensen Huang said on the call that AWS could end up buying millions of Vera CPUs as the product ramps.3

The memory work is the most unusual part. Amazon's chip-design unit, Annapurna Labs, will use NVHBM in future Trainium chips. The stated aim is to put Trainium accelerators and Nvidia GPUs in a common rack-scale architecture.120 One analysis says Annapurna is the first partner for NVHBM, which builds Nvidia's own memory controller into the HBM base die.6 Trainium4 had already been announced with support for Nvidia's NVLink Fusion interconnect. One review argued that this weakens the idea that AWS is building a "post-Nvidia" chip, since the next generation is designed to work with Nvidia's fabric rather than replace it.4

The deal also includes AI factories for the US government with 100,000 GPUs on secure AWS infrastructure for federal and national-security work.118 Nvidia's Nemotron open models will be available on Bedrock and SageMaker. Nvidia's cuDF and cuVS libraries are being integrated into Amazon EMR and OpenSearch.20 Amazon said the analytics upgrade runs about 3.7 times faster and delivers roughly 30% better price-performance than standard configurations. It also said vector index builds in OpenSearch finish about nine times faster on GPUs.18 Amazon Robotics is adopting Nvidia's physical-AI stack, including Omniverse, Cosmos, Isaac and Jetson.12

The Trainium paradox, and why it isn't one

The usual question is why Amazon would expand its Nvidia orders while building chips meant to compete with Nvidia. Amazon has said its custom silicon business has passed a $25 billion annualized revenue run rate, supported by $225 billion in commitments from AI labs including Anthropic and OpenAI.12 Trainium3 capacity has been described as almost fully booked.2 Earlier this year Amazon said it had deployed more than 2.1 million AI chips in the previous 12 months, and more than half were Trainium.6

The coverage mostly agrees on how to read this, and I think it's right. Value Add VC called it a hedge driven by demand that is outrunning what either supply chain can deliver alone.15 Tech-insider described an informal split: Nvidia GPUs and Vera CPUs for the hardest frontier training and agentic workloads, and Trainium as AWS's cheaper, vertically integrated option.3 Another analysis made the commercial point directly. If AWS pushed Trainium on customers who need CUDA or the newest GPUs, those customers could move to Azure or Google Cloud. So AWS keeps both and uses each as leverage in negotiations with the other.6

The customer list supports that view. Anthropic is the only frontier lab known to train on Trainium at large scale, with more than a million Trainium2 chips. OpenAI's original $38 billion AWS deal paid for Nvidia GB200 and GB300 systems, and its 2-gigawatt Trainium commitment had not been reported in production as of early October.4 Trainium is a serious product, but it is not displacing Nvidia. Nvidia remains the default for most of the market, and AWS's August order reflects that.

I'd add one point. The NVHBM arrangement makes Amazon a licensee of Nvidia technology inside its own competing chip, which ties the two companies' roadmaps closer together. Value Add VC noted that this kind of broad technical integration raises Amazon's cost of switching if it later wants to favor Trainium over Nvidia.15 Nvidia is selling GPUs to its biggest custom-silicon rival, and it is also getting its interconnect and memory designs into that rival's chips.

What it means for Nvidia

For Nvidia, the AWS deal was the headline item in a very strong quarter. The company reported $96.2 billion in revenue, including $89 billion from data centers, up 117% from a year earlier.12 Total revenue grew 106% year over year.3 Nvidia also more than doubled its supply and manufacturing commitments, from $119 billion to $279 billion, to secure memory and foundry capacity.12 Tech-insider, citing CNBC, said the stock rose nearly 9% the next trading day and added about $440 billion in market value.3

The Vera CPU deal could prove as important as the GPU numbers. Huang had said in May that Vera opened a "brand-new $200 billion TAM." Kress said she expects every major hyperscaler, neocloud, AI lab and system OEM to deploy it, with Oracle and SpaceXAI among the early partners.12 An AWS commitment gives that claim a large first customer and puts Nvidia in more direct competition with Amazon's own Graviton CPUs, as well as Intel and AMD. Tech-insider noted that AMD has not disclosed a hyperscaler GPU commitment of comparable size.3

The deal also differs from Nvidia's 10-gigawatt agreement with OpenAI. That agreement pairs compute commitments with up to $100 billion in potential Nvidia investment.1 Based on what the companies have disclosed, the AWS deal is an ordinary commercial purchase. That makes it a cleaner demand signal and less open to the vendor-financing criticism aimed at some recent AI deals.15

The capacity crunch is the real story

The deal's real significance shows up in AWS pricing. AWS planned to raise most reserved Nvidia GPU prices through EC2 Capacity Blocks by 15% or more on October 7. That would be the product's third increase this year.5 It followed a 20% increase in July.9 Andy Jassy has reportedly said Amazon's current capacity cannot meet demand, and that demand for 2028 already looks "remarkable." Amazon has also reportedly raised its 2026 capital spending forecast to $220 billion.2 Bloomberg, relaying a Financial Times report, said Amazon is exploring a way to move about $8 billion of Nvidia chips off its balance sheet.5 If that's accurate, the bill for this buildout is large enough that Amazon is looking for new ways to finance it.

The reading I'd commit to: this deal is mainly about guaranteed supply. Nvidia's output is booked years in advance, Trainium3 is close to sold out, and AWS is raising GPU prices every few months. Amazon is reserving 2027–2028 allocation now because waiting would mean giving that capacity to rivals. One pricing analysis noted that the deal doesn't ease 2026's supply-demand imbalance, because the chips arrive later.10

Risks worth watching

There are real counterarguments. Chips arriving in 2027 and 2028 are a long-term bet, and AI infrastructure demand could shift a lot before then.17 Some analysts warn that if AI spending levels off, AWS could be left with underused capacity and weaker margins.13 Huang's case is that more compute produces more "profitable tokens." TechCrunch noted that investors still need to see whether added compute turns into profit that neatly.12

Two things are worth tracking. One is the mix of Trainium and Nvidia hardware Amazon deploys over the next several quarters, which will show which side of the hedge Amazon expects to win.15 The other is AWS's AI infrastructure revenue once the new hardware starts arriving in 2027.17 For now, the evidence suggests both chip strategies are growing together, with Amazon and Nvidia increasingly tied to each other.