Nvidia and Amazon announced an expanded partnership on August 26, with AWS placing a new order for 2 million Nvidia GPUs. The previous agreement between the two companies, signed five months ago, covered just over 1 million chips — meaning this increase effectively doubles the original commitment. Both companies cited demand exceeding earlier forecasts, coming from startups, enterprise customers, AI labs and government agencies.
In-house silicon hasn't displaced Nvidia
Amazon's own accelerator chips are now running at a $25 billion annualized revenue run rate, the highest among any cloud provider's homegrown AI silicon. The conventional assumption is that stronger in-house chips mean less external purchasing. This order pushes back against that assumption: both tracks are expanding at the same time.
The reason lies in the customer mix. AWS's largest AI workloads don't all originate from Amazon itself — external labs like Anthropic and OpenAI have already committed to roughly $225 billion in Nvidia-side purchases, and most of their engineering stacks are tied to CUDA, where migration costs are high and timelines are tight. What a cloud provider can do is make sure the chips are on hand. Amazon's in-house silicon absorbs the portion of workloads where Amazon itself controls scheduling — the two aren't substitutes for each other.
Beyond the chip order
The scope of this partnership goes well beyond “buying GPUs.” Alongside the chips, AWS will adopt Nvidia's Vera CPU, bring in the physical-AI stack built for robotics, offer the Nemotron family of open models through Amazon Bedrock and SageMaker, and integrate networking hardware into AWS infrastructure. Nvidia CFO Colette Kress described the Vera rollout as "some integrated with Rubin, others standalone."
The Vera CPU line deserves attention on its own. Jensen Huang put its addressable market at $200 billion back in May, going up against x86, which has held its ground in data centers for two decades. Landing a customer the size of AWS amounts to the product line's first large-scale endorsement.
Both sides of the supply-demand equation are leveraging up
Nvidia's most recent quarterly data center revenue reached $89 billion, up 117% year over year, with guidance for the next quarter at $108 billion. To keep up with these orders, Nvidia itself is betting heavily upstream — committing roughly $279 billion to supply chain and manufacturing capacity, with $92 billion in spending planned for the rest of this fiscal year and $87 billion projected for fiscal 2028.
"The thing that matters for the industry is that AI is now doing productive and useful work. AI is generating profitable tokens."
Huang left that line after the earnings call, anchoring the case for continued capital spending back to revenue. Rough math on the scale involved: 2 million GPUs, priced at current-generation public price ranges, already represents hundreds of billions of dollars in hardware alone; add data center facilities, power and networking, and the rental revenue AWS needs to recoup from this capacity over the coming years is similarly measured in the hundreds of billions (an estimate — neither company has disclosed contract values). Cloud providers only sign at this scale when customer-side commitments are already stretched across the same time horizon.
For downstream users, delivery pace matters more than the total figure. The previous 1-million-chip agreement was signed five months ago, and its delivery timeline still hasn't been disclosed. Now that another 2 million chips have been stacked on top, the real measure of this expansion will be when AWS's GPU instances stop requiring a wait.
Sources: CocoLoop, TechCrunch, Nvidia earnings call; GPU counts, data center revenue and capex commitments verified against Nvidia's disclosures — contract value was not disclosed by either company.