Amazon triples its Nvidia GPU order and signals that AI capacity is set to get pricier

Amazon and Nvidia announced on Wednesday (26) an expansion of their partnership that includes adding another 2 million Nvidia GPUs to Amazon Web Services (AWS) data centers, according to TechCrunch. The announcement, made during Nvidia's quarterly earnings call, delivers chips from the Blackwell Ultra, Rubin, and Rubin Ultra lines to AWS facilities in 2027 and 2028.

The chips are designed to handle the heavy computing load of training and running AI models. Neither company disclosed financial terms, but based on the unit cost of a GPU, TechCrunch estimates the deal's value at tens of billions of dollars.

What came before, and why it tripled

The point that gives the move its name is speed. The announcement came five months after Amazon agreed to deploy more than 1 million Nvidia GPUs across AWS infrastructure starting this year. In a statement, Nvidia said that since then, "demand has outpaced those expectations."

The companies cited "surging demand" from startups, enterprises, AI labs, and even governments as the reason. In other words: the order didn't triple because of advance planning, but because appetite for training and inference capacity grew faster than expected half a year ago.

The deal goes beyond buying more chips. Nvidia said it will also integrate into AWS its networking hardware (which connects thousands of GPUs into a single system), its open models, CPUs, data processing software, and robotics platform.

Amazon buys from Nvidia even while making its own chips

The strategic detail is that the expansion happens even as Amazon itself invests heavily in its own silicon to reduce dependence on Nvidia, and even compete with it:

• Trainium: an AI chip positioned as a direct alternative to Nvidia's H100 and Blackwell GPUs for deep learning workloads. According to AWS's head of AI, Peter DeSantis, Amazon is in talks to sell Trainium to other companies.
• Graviton: an Arm-based CPU seen as a challenger to traditional server chips from Intel and AMD.

Amazon said its custom chip business has crossed $25 billion in annualized revenue, driven by $225 billion in total commitments from AI labs like Anthropic and OpenAI. Even so, it bought another 2 million GPUs from Nvidia. One possible reading: not even the company building the alternative can do without the dominant supplier when demand surges.

Along with the GPUs, Nvidia will also ship an unspecified number of Vera CPUs, "some integrated with Rubin, others standalone," according to Nvidia CFO Colette Kress. CEO Jensen Huang went as far as to say in May that Vera represents a "new $200 billion addressable market."

The Nvidia numbers behind the move

The partnership was announced on the same day Nvidia released results that help gauge the scale of the race:

| Metric | Value | Note | |---|---|---| | Q2 revenue | $96.2 billion | Above analyst estimates | | Data center revenue | $89 billion | +117% year-over-year | | Q3 projection | $108 billion | Partly driven by new Rubin chips | | Supply/manufacturing commitment | $279 billion | Up from $119 billion the previous quarter |

That jump in supply commitments (from $119 billion to $279 billion in a single quarter) includes $92 billion projected for the rest of the fiscal year and another $87 billion in fiscal year 2028, as Nvidia works to secure memory and manufacturing capacity.

The thing that matters for the industry is that AI is now doing productive, useful work. AI is generating profitable tokens. If we had more computing capacity, we could generate more profitable tokens, which results in more profit for every service.

>

-- Jensen Huang, CEO of Nvidia

TechCrunch itself points out the open question: investors will be watching to see whether additional computing actually translates into additional profit, as AI companies pour hundreds of billions into infrastructure.

What this changes for developers building software in Brazil

On the robotics and enterprise side, Amazon plans to adopt Nvidia's full physical AI stack (Omniverse, Cosmos, Isaac, and Jetson) for its warehouse robot fleet. And, more relevant for developers, AWS will serve Nvidia's Nemotron open model family through Amazon Bedrock (a managed foundation model platform) and SageMaker.

For Brazilian engineering teams, the practical takeaway is about cost dynamics. GPU capacity remains the expensive bottleneck of the AI equation, and the fight over it involves supply commitments in the hundreds of billions of dollars. When that capacity is scarce and contested, the price per hour of accelerated instances and the cost per inference token are unlikely to fall in the short term, which shows up in the bill for anyone consuming these services via API or cloud instances, billed in dollars and still subject to exchange-rate swings in Brazil.

Some takeaways worth factoring into planning for anyone running AI here:

• Open models on Bedrock/SageMaker: Nemotron's arrival as a managed option expands the menu for anyone who wants to run an open model without building their own GPU infrastructure, a relevant path for those seeking cost predictability.
• Trainium as a pricing alternative: with Amazon pushing Trainium as a competitor to Nvidia GPUs, it's worth tracking whether instances based on this chip offer a more competitive cost per training run or inference than Nvidia-accelerated ones.
• What's still unclear: there are no public financial terms for the deal, nor clarity on when (or whether) this extra capacity will ease prices for end customers. For now, the signal is that demand still outstrips supply, the opposite of a price-drop scenario.

While the industry bets that "more computing turns into more profit," Brazilian developers would do well to treat AI cost as a strategic architecture variable, not just an infrastructure detail.