Amazon Triples Its Nvidia GPU Order to 2 Million Chips for AWS
Amazon ordered another 2 million Nvidia GPUs for AWS, tripling a deal made just 5 months ago. Chips include Blackwell Ultra, Rubin, and Rubin Ultra, arriving in 2027-2028.
Amazon and Nvidia announced on August 27, 2026 that Amazon will add 2 million more Nvidia GPU chips to AWS data centers in 2027 and 2028. This triples a deal struck just five months ago, when Amazon committed to deploying over 1 million Nvidia GPUs across AWS this year. Nvidia says demand since that original agreement "has exceeded those expectations." The chips span three generations: Blackwell Ultra, Rubin, and Rubin Ultra, and the total deal is estimated to be worth tens of billions of dollars, though neither company disclosed exact terms.
What happened
| Detail | Fact |
|---|---|
| New GPU order | 2 million additional Nvidia GPUs for AWS |
| Previous order (5 months ago) | More than 1 million Nvidia GPUs |
| Chip generations included | Blackwell Ultra, Rubin, Rubin Ultra |
| Delivery timeline | 2027 and 2028 |
| Estimated deal value | Tens of billions of dollars |
| Nvidia Q2 revenue | $96.2 billion (beat analyst estimates) |
| Nvidia data center revenue Q2 | $89 billion, up 117% year over year |
| Nvidia Q3 revenue guidance | $108 billion expected |
| Nvidia supply commitments | $279 billion (up from $119 billion last quarter) |
Amazon and Nvidia announced the expanded deal during Nvidia’s quarterly earnings call. The companies cited “surging demand” from startups, enterprises, AI labs, and governments as the driver. The partnership is not limited to GPU procurement. Nvidia’s networking hardware (the interconnects that link thousands of GPUs into a single system), its open AI models, CPUs, data processing software, and robotics platform will all be integrated across AWS services.
Nvidia CFO Colette Kress confirmed that an unspecified number of Vera CPUs will also go to AWS, “some integrated with Rubin, others standalone.” CEO Jensen Huang framed the Vera CPU opportunity as a “brand new $200 billion TAM” (total addressable market) when he first discussed it in May. Kress added that Vera is expected to ship to “every major hyperscaler, neocloud, AI lab, and system OEM,” with Oracle and SpaceXAI named as early partners already receiving shipments.
The robotics and enterprise side of the deal
Amazon plans to adopt Nvidia’s full physical AI stack for its warehouse robots. That stack includes four platforms:
- Omniverse: Nvidia’s simulation and digital twin environment
- Cosmos: its world model platform
- Isaac: its robotics development toolkit
- Jetson: edge computing hardware for robots and on-device AI (a new entry-level version was announced this week)
On the cloud software side, AWS will serve Nvidia’s Nemotron family of open models through Amazon Bedrock (its managed foundation model service) and SageMaker (its managed cloud ML platform).
Does Amazon’s own chip strategy contradict this deal?
Not exactly, but the tension is real. Amazon has been building its own AI accelerators specifically to reduce dependence on Nvidia. Its Trainium chips target the same deep learning workloads as Nvidia’s H100 and Blackwell chips, and Amazon’s AI chief Peter DeSantis has confirmed the company is in talks to sell Trainium to other businesses for use in external data centers. Amazon’s Arm-based Graviton CPU is also positioned as a challenger to chips from Intel and AMD.
Amazon noted on its last earnings call that its custom chip business crossed a $25 billion annualized revenue run rate, supported by $225 billion in total commitments from AI customers including Anthropic and OpenAI. That is a healthy in-house chip business. But the scale of Wednesday’s Nvidia order signals that internal supply, at least for now, cannot keep up with total AWS demand.
Why it matters
For Nvidia, this is confirmation that Rubin-generation demand is real before most of those chips have even shipped. Investors had been watching for early Q3 Rubin sales as a signal that the upgrade cycle would hold. The Amazon deal and Nvidia’s $108 billion Q3 guidance answer that question clearly. Nvidia has also committed $279 billion to lock in supply and manufacturing capacity, up from $119 billion just last quarter, including $92 billion for the rest of this fiscal year and another $87 billion in fiscal 2028.
For businesses using AWS, this means more GPU capacity will be available through services like Bedrock and SageMaker, and Nvidia’s software stack (including robotics and simulation tools) will be more deeply embedded in the platform they may already be using. As we covered in our piece on OpenAI’s data center leadership changes, the infrastructure race is intensifying at every layer of the stack.
For anyone budgeting AI workloads, the supply picture is improving, but costs are not likely to fall fast. The companies driving this demand are spending hundreds of billions of dollars, and those costs flow through to API pricing.
Our take
Jensen Huang’s quote from the call is worth sitting with: “AI is generating profitable tokens. If we had more compute, we could generate more profitable tokens.” That is a straightforward economic argument for infinite GPU spending, and right now the market is accepting it. But “more compute equals more profit” is only true if inference demand keeps growing at the same rate as capacity. That is the bet every hyperscaler is making simultaneously.
The dual strategy Amazon is running (buying Nvidia at scale while building its own Trainium chips) is not contradictory. It is rational hedging. The risk is that Amazon’s internal chips never hit the adoption or performance curve needed to matter at scale, leaving Nvidia’s pricing power unchallenged. For businesses running serious AI workloads on AWS, understanding which chip type your workload runs on, and at what cost, is worth more attention than most teams give it. If you are evaluating where to build or move an AI-driven product, our AI integration work often starts exactly there.
What to do about it
- Check which AWS compute instances your AI workloads run on today. Graviton, Trainium, and Nvidia-backed instances carry different price and performance profiles.
- Watch for new Bedrock model options. Nvidia’s Nemotron models coming to Bedrock may offer a cheaper or faster option for certain inference tasks than your current setup.
- If you are evaluating robotics or edge AI, Nvidia’s Jetson stack will now be more tightly supported on AWS, which changes the build-vs-buy calculus for warehouse or industrial automation projects.
- Track Rubin GPU availability on AWS. New chip generations typically arrive in AWS regions months after the announcement. Knowing the timeline helps with capacity planning.
The bottom line: Nvidia remains the dominant AI chip supplier, and AWS is betting its infrastructure future on that staying true through at least 2028. Plan your cloud AI spend accordingly.
Frequently asked questions
How many Nvidia chips is Amazon buying for AWS?
Amazon is adding 2 million more Nvidia GPU chips to AWS data centers, on top of the 1 million-plus GPUs it agreed to deploy five months earlier. The new chips include Blackwell Ultra, Rubin, and Rubin Ultra GPUs, and will arrive in 2027 and 2028.
How much is the Amazon Nvidia chip deal worth?
Neither company disclosed the financial terms, but based on GPU unit costs, the deal is estimated to be worth tens of billions of dollars.
What did Nvidia report for Q2 2026 earnings?
Nvidia reported Q2 2026 revenue of $96.2 billion, beating analyst estimates. Data center revenue was $89 billion, up 117% from the same period a year ago. Nvidia guided for $108 billion in Q3 revenue.
Does Amazon have its own AI chips that compete with Nvidia?
Yes. Amazon's Trainium chips target the same deep learning workloads as Nvidia's H100 and Blackwell GPUs, and its Graviton CPU competes with server chips from Intel and AMD. Amazon's custom chip business recently crossed $25 billion in annualized revenue, though the company is still buying Nvidia chips at scale to meet total AWS demand.
