Anthropic and Amazon close a deal for up to 5 GW: the AI race is now also electric
Anthropic announced on April 20, 2026 a new agreement with Amazon to secure up to 5 gigawatts of capacity to train and operate Claude. The news is more than a cloud expansion: it shows that the competition for AI models has become a race for physical infrastructure, involving proprietary chips, data centers, energy, global regions and long-term contracts.
According to Anthropic, the agreement includes new capacity in Trainium2 in the first half of 2026 and almost 1 GW between Trainium2 and Trainium3 by the end of the year. The company also says it will commit more than $100 billion to AWS technologies over the next ten years.
Why 5 GW matters
Gigawatts often came up in discussions about power grids, power plants, and heavy industry. Now appear in AI ads. This detail reveals the scale of the problem: frontier models don't just depend on better algorithms; they depend on a continuous supply of computing.
The agreement deepens a relationship that began in 2023. Anthropic says that more than 100,000 customers already run Claude on Amazon Bedrock and that Project Rainier, created with Amazon, uses more than one million Trainium2 chips to train and serve Claude.
In practice, the objective is to reduce bottlenecks. When an AI grows among users, companies and developers, the limitation is not just the quality of the model. It's availability, latency, inference cost, and the ability to respond at peak times.
The role of Trainium
Trainium is Amazon's bet on silicon suitable for AI. For AWS, convincing Anthropic to use Trainium for a decade is an important validation of the custom chip strategy. For Anthropic, diversifying and expanding capacity is a way to sustain growth without relying on a single chain of GPUs.
The company says the commitment covers Graviton and Trainium2 through Trainium4, with an option to purchase future generations of chips from Amazon. There will also be expansion of inference in Asia and Europe, a critical point for reducing latency and serving international customers.
This type of agreement indicates that the computing layer is becoming as strategic as the model itself. Those who control capacity, cost per token and availability can deliver more stable and aggressive products.
Record demand and operational pressure
Anthropic claims that enterprise and developer demand accelerated in 2026, along with usage on consumer plans such as Free, Pro and Max. In the announcement, the company says that its annualized revenue exceeded US$30 billion, compared to around US$9 billion at the end of 2025.
This growth comes with predictable problems: usage spikes, reliability degradation and performance pressure. Anthropic itself recognizes that the growth in consumption impacted Free, Pro, Max and Team users at times of greatest demand.
This is where infrastructure becomes product experience. An excellent but unstable model loses value for companies that need to put AI into critical flows. The agreement with Amazon tries to attack exactly this point: more capacity in a few months and almost 1 GW before the end of 2026.
What does this change for companies
For corporate customers, the announcement reinforces that the choice of AI will be increasingly linked to the choice of cloud. Claude is available via Amazon Bedrock, and Anthropic says the Claude Platform will come directly to AWS, with the same account, controls and billing.
This reduces friction for companies already on AWS, but also increases the importance of multi-cloud architecture and vendor governance. Anthropic highlights that Claude remains available on the three largest cloud providers: AWS, Google Cloud and Microsoft Azure. Still, the agreement makes clear what the partnership's gravitational center is.
AI has entered the era of industrial contracts
The most interesting point of the ad is the change in language. We are no longer just talking about models, prompts or rankings. We are talking about gigawatts, a ten-year commitment, our own chips, inference regions and billion-dollar investment.
This suggests that the next phase of AI will be decided by two fronts at once: scientific advancement and industrial capacity. Labs that fail to transform demand into reliable infrastructure may fall behind even with good models.
For the end user, the expected effect is simple: Claude is more available, faster and more present in business products. For the market, the message is bigger: the AI race is no longer just about software.
Sources
- https://www.anthropic.com/news/anthropic-amazon-compute
- https://aws.amazon.com/bedrock/anthropic/