Amazon Wants to Sell Its AI Chips to Everyone. Nvidia Should Worry.

For three years, the AI hardware story has had one protagonist: Nvidia. Every frontier model trained, every data centre built, every hyperscaler's capex announcement ultimately flowed back to Nvidia's GPUs. That monopoly is now facing its most serious structural challenge — from inside the cloud business that has been Nvidia's largest customer.

What Amazon Just Confirmed

Amazon is exploring selling its custom Trainium AI chips to third-party data centres, according to Bloomberg and TechCrunch. AWS AI chief Peter DeSantis told Bloomberg that AWS is in talks to allow other organisations to use Trainium outside of AWS, adding that there is so much underconsumption in AI that external sales will not hurt cloud revenue.

Read that carefully. Until now, Amazon's Trainium chips were available only through AWS — you rented compute, you did not buy silicon. DeSantis is signalling a fundamental shift: Amazon may begin selling the physical chips to data centre operators who are not AWS customers.

That puts Amazon in direct competition with Nvidia not just as a cloud provider, but as a chip vendor. The discussions are described as early-stage. But the strategic direction is unambiguous.

The Scale That Makes This Credible

This is not a side project. Amazon's custom silicon business — including Trainium AI chips — has surpassed a $20 billion annual run rate, with a nearly 40% quarter-over-quarter increase in Q1 2026.

Amazon CEO Andy Jassy has been explicit about the latent scale. Jassy wrote in his April shareholder letter that a standalone chip business selling externally could reach an annual run rate of around $50 billion.
The demand signal is already overwhelming. Trainium 2 chips offer a 30% price-performance advantage over comparable GPUs and are largely sold out. Trainium 3, which began shipping in early 2026, provides a further 30-40% improvement and is nearly fully subscribed, with substantial reservations already placed for Trainium 4, expected in about 18 months.

A chip business running at $20 billion, growing at triple digits, with its current generation sold out and the next generation already reserved — that is not a challenger experiment. That is a second credible AI silicon supplier emerging at scale.

The Customers Already Committed

The most striking detail is who is already building on Trainium. Anthropic has signed on for up to 5 gigawatts of current and future Trainium capacity. OpenAI has agreed to around 2 gigawatts of Trainium capacity through AWS.

Both of the world's leading independent frontier AI labs — fierce competitors with each other — are running on Amazon's silicon. Anthropic and OpenAI have made significant, multi-year commitments for Trainium, alongside growing interest from companies such as Uber, with commitments now totalling over $225 billion.

The adoption extends across the enterprise. Amazon Bedrock, used by over 125,000 customers, leverages Trainium for most of its inference tasks, with nearly 80% of Fortune 100 companies using Bedrock. Meta has committed to using tens of millions of Graviton cores for CPU-intensive agentic AI workloads.

Why This Is a Nvidia Problem

Nvidia's dominance has rested on two pillars: superior performance and the CUDA software ecosystem that locks developers in. Amazon's Trainium strategy attacks the first pillar directly — 30-40% better price-performance — and the external sales move attacks the second by giving data centre operators an alternative they can own rather than rent.

There is a notable irony in the timing. The same Andy Jassy whose phone call to the Treasury Secretary triggered the Anthropic export shutdown last week — as we reported — is also the CEO whose chips power Anthropic's models, OpenAI's models, and 80% of the Fortune 100. Amazon is positioning itself as the infrastructure layer beneath the entire AI industry, regardless of which model company wins. The chips are the same whether you are OpenAI or Anthropic. The leverage that creates is enormous.

The European Dimension — A Gap With No Filler

Here is the part of this story that should concern European policymakers most. The emerging AI silicon duopoly — Nvidia and now Amazon — is entirely American. There is no European equivalent. There is no European Trainium. There is no European hyperscaler with a custom AI chip business running at $20 billion.

Europe regulates AI deployment through the EU AI Act. Europe funds AI model development through Mistral, Aleph Alpha and the AI Factories programme. But the silicon layer — the physical chips on which all AI ultimately runs — is a domain where Europe has essentially no presence. When Amazon and Nvidia divide the AI chip market between them, European AI infrastructure will run on American silicon, manufactured in Taiwan, governed by American export controls.
The Anthropic shutdown last week demonstrated what American control over the model layer means for European users. American control over the silicon layer is a deeper and more permanent dependency — and one that no amount of European model funding can address, because a European model still needs chips to run on.

The European Perspective