Amazon starts selling the AI chips it built to escape Nvidia

Information Technology 4 min 80 sources

Amazon is in talks to sell its custom Trainium chips to outside customers — turning a tool it made for itself into a product. Plus a Trump-claimed Intel–Apple deal, Meta's new data-center pact, and a US warning to ASML about China.

Key takeaways

Amazon is in talks to sell its custom Trainium AI chips to outside buyers, turning an in-house tool built to escape Nvidia into a product — and adding a third serious player to a market Nvidia dominates.
The catch is software: Nvidia's real lock-in is CUDA, the code developers already wrote, so Amazon must make switching cheap or Trainium stays a bargaining chip, not a replacement.
Across the day, the money kept flowing into AI's foundation — chips, data centers, and inference — not the apps on top, with Meta, Intel, ASML, and a $1.5B startup raise all in the mix.

For years, Amazon designed its own AI chips for one reason: to stop paying Nvidia so much. Now it wants to sell them to everyone else.

Amazon is in talks to offer its custom Trainium chips — the silicon that powers AI training and runs inside its cloud — directly to outside customers, according to reports from Bloomberg and Reuters [47][14][6]. Until now, Trainium was an in-house tool. You could rent it through Amazon Web Services, Amazon’s cloud arm, but you couldn’t buy the chip. The shift would put Amazon in direct competition with Nvidia, whose chips sit in roughly nine out of ten AI data centers today.

Why now

Two pressures are squeezing at once. Nvidia’s top AI chips are expensive and supply-constrained — companies wait months and pay premium prices. And the firms buying them, Amazon included, have spent the last two years learning that depending on a single supplier for the most important component in your business is a weak spot, not a convenience.

Amazon already builds Trainium for its own racks. Once you’ve designed a chip, made it work at scale, and proven it on real workloads, the extra cost of selling it to others is small — and the upside is a second revenue stream plus a dent in your rival. An Amazon executive said the company aims “to compete with Anthropic in coming year” on the model side, too [25] — a sign Amazon wants to own more of the AI stack, from chip to cloud to model.

What changes

If this lands, the AI-chip market stops being a two-name race. Today buyers choose mainly between Nvidia and AMD; Trainium-for-sale adds a third serious option from a company that already runs one of the world’s largest clouds. That matters for price. When a near-monopoly supplier gains a credible competitor, the pressure on margins is immediate — even before a single chip ships.

There’s a catch worth tracking. Nvidia’s real moat isn’t only the chip; it’s CUDA, the software layer developers have written their AI code against for years. Moving to a different chip often means rewriting that code. Amazon will have to make that switch cheap, or Trainium stays a price-negotiation lever rather than a true replacement.

If you build or buy AI compute: a third merchant chip is a number to watch on your cost sheet, even if you never touch one — its existence pulls Nvidia’s quotes down.

The chip-deal day around it

Amazon wasn’t the only chip story. Intel shares jumped about 9% after President Trump said Intel would partner with Apple on US chip design [12][13][52]. Intel and Apple have not confirmed the terms, and the details Trump described are unverified — treat it as a claim, not a signed deal. Intel’s stock has been trading less on its products than on belief that its foundry business — the part that manufactures chips for other companies — will eventually work [26].

Separately, the US told ASML, the Dutch company that makes the only machines capable of printing the most advanced chips, that it is concerned China may already have a top-tier chip-making tool, Bloomberg reported [20]. ASML’s machines are the chokepoint of the whole industry; whether China has cracked that barrier shapes years of export policy.

And Meta signed new AI computing deals with data-center firm Crusoe [14], part of a steady stream of big-tech spending on the physical buildings and power that AI needs. On that note, the US government just handed AI data centers a mandated fast lane to connect to the electrical grid [48] — a sign that compute, not code, is now the bottleneck everyone is racing to clear.

Money still pouring into the picks and shovels

The funding side stayed loud. Baseten, a startup that helps companies run AI models in production — the “inference” step, where a trained model actually answers requests — is reportedly raising $1.5 billion just months after its last large round [1]. General Intuition is in talks to raise $300 million at a roughly $2 billion valuation [4]. And Architect Labs raised $24 million to take on Broadcom and Marvell in the custom-chip business [11] — the same vertical-integration bet Amazon is making, at startup scale.

The thread tying the day together: almost no one is betting on a single AI app. They’re betting on the layer underneath — the chips, the data centers, the inference plumbing. When everyone sells shovels, the price of shovels is the story.

02 · Lesson · why it matters

When you build a tool to escape someone, you often become them

The cheapest way to stop depending on a supplier is to become one — and the same move that frees you quietly puts you in charge of the people standing where you used to stand.

A defensive move, pointed outward

Amazon did not set out to sell chips. It set out to stop buying so many of Nvidia’s.

The most expensive, most fought-over component in modern computing is the AI chip. Amazon needed mountains of them to run its cloud. Nvidia sold them dear and slow. So Amazon did what large, patient companies do when a supplier has too much power over them: it built its own. Trainium was a shield — a way to need Nvidia a little less each year.

Now that shield is becoming a product. The chip Amazon made to free itself is going on sale to everyone else. That turn — from defending against a supplier to becoming one — is the pattern worth seeing today. It is not unique to Amazon. It is what happens, again and again, whenever someone climbs out from under a dependency.

The economics that bend the move

Once you have built something hard, selling it costs almost nothing more.

This is the quiet engine underneath the story. The cost of a chip is mostly in the design and the first working version — the years of engineers, the failed prototypes, the deal for factory time. Amazon already paid all of that, for its own reasons. The marginal cost of letting a stranger buy the same chip is small. So a tool built for one purpose develops a second life, almost by gravity, because the hardest spending is already behind it.

You feel this in your own life. The person who taught themselves to fix their own car ends up fixing the neighbour’s. The company that built an internal tool to manage its own messy work eventually sells it as a product — Amazon’s entire cloud business began exactly this way, as plumbing built for Amazon’s own website. Self-reliance, done well, rarely stays private. The fixed cost is paid; the door to selling stands open; someone walks through it.

The thing you escape, you now sit on top of

Here is the part that should land harder than the cleverness of the move.

Amazon spent years resenting that Nvidia sat in the middle — that one company controlled the scarce thing everyone needed, and could set the terms. The fix was to claim some of that middle ground for itself. But notice what that actually means. Amazon does not become free of the middle. Amazon becomes the middle, for someone else. The startup that buys Trainium tomorrow is now where Amazon stood yesterday: dependent on a supplier that owns the scarce layer.

The position you fought to escape doesn’t disappear. It moves down a level and waits for the next person. There is always a layer that everyone above it needs, and whoever controls that layer holds a quiet power over everyone standing on it — power they often didn’t seek, that simply comes with the territory. Nvidia didn’t set out to rule the AI economy; it made a good chip and the dependency formed around it. Amazon is now stepping into a smaller version of the same seat.

No one is only a customer

We tend to read a market as two clean roles: sellers up here, buyers down there. The chip story shows it’s never that tidy.

Amazon is Nvidia’s customer and Nvidia’s rival in the same breath. It buys chips, builds chips, and sells chips, all at once. The data-center firms Meta just signed deals with are both suppliers to Meta and customers of the chipmakers above them. Everyone in this chain is buying from the layer above and selling to the layer below — a customer and a supplier at the same time, depending which direction you look.

That’s the truer shape of almost any system you’re inside. The dependency you feel — on a platform, an employer, a single tool you can’t switch away from — is real, but it isn’t a fixed rank. It is a position in a chain, and the people one rung above you are leaning on someone above them just as hard. The power that looks total from below is partial from inside.

What the whole looks like

Step back and the AI economy stops looking like a race between a few famous names and starts looking like a stack of dependencies, each layer needing the one beneath it. Chips need the machines that print them. Models need the chips. Apps need the models. You, reading this, sit somewhere near the top, depending on layers you’ll never see — the inference plumbing, the data centers, the rare machine in the Netherlands that prints the smallest circuits.

Amazon’s move to sell its chips is one company trying to climb one rung and, in doing so, becoming the thing it climbed away from. It’s a small, clear instance of a force that runs through everything: the middle layer holds the power, the power moves but never vanishes, and almost no one in the chain is only above or only below. We are all somebody’s supplier and somebody’s dependent — and the seat that looks like control, seen from the next rung up, is just another place that’s leaning on something it can’t see.

03 · Lab · your turn

The Supplier Trap

Rehearse the build-versus-buy choice against a scarce supplier, and feel how escaping a dependency turns you into one.

04 · Hope · carry this

Every monopoly that looked permanent has eventually met a rival built by someone who got tired of waiting — the scarce layer always tempts the next builder, and that restless impatience is how power keeps getting shared out.