Amazon Is Circling Decart for Up to $7B — the Real Prize Is Chip Portability

Amazon is in talks to acquire artificial-intelligence startup Decart in a deal worth as much as $7 billion, the Wall Street Journal reported on October 11, via Reuters. Three-year-old Decart makes software that lets developers run AI workloads across different AI chips — Nvidia GPUs, Google TPUs, and Amazon's own Trainium. Reuters said it could not independently verify the negotiations, so treat this as talks, not a done deal.

The subplot is almost as interesting as the deal: Anthropic reportedly considered acquiring Decart for around $6 billion last month and walked away. Either Amazon sees strategic value Anthropic didn't, or it's willing to pay up for talent and technology. Possibly both.

Why the Amazon Decart acquisition talk is about portability, not chips

Decart doesn't make chips. It makes chips interchangeable — optimization software that lets developers use whichever AI hardware is available or cheapest, instead of being locked to one vendor's silicon. In a market defined by GPU scarcity and everyone racing to secure compute, portability is leverage. The company that controls the abstraction layer between models and hardware gets a say in every infrastructure decision its customers make.

That's why Amazon wants it, and why the price is $7 billion for a three-year-old company. AWS already works with Decart, and Amazon has been pushing its Trainium chips hard — the company claims its Trainium3 implementation delivers four times faster real-time video inference at half the cost of GPUs, though the announcement offered limited comparison details. Owning Decart would let Amazon offer customers a graceful path: run on whatever you have today, migrate to Trainium tomorrow, with the portability layer smoothing the transition. It's the oldest enterprise software playbook in the book — own the migration path, own the customer.

The competitive picture

The acquisition talks land in a week when the AI infrastructure trade is wobbling. CoreWeave's debt load has investors questioning whether neocloud operators can turn profitable fast enough to service their borrowing, Oracle shares slid on the same concerns, and analysts are openly debating whether the $650 billion hyperscalers plan to spend on AI infrastructure in 2026 is financeable. Against that backdrop, Amazon spending up to $7 billion to tighten its grip on the chip-software stack reads as a incumbency move: when the building boom gets questioned, buy the toll roads.

There's a regulatory shadow here too. A cloud provider with its own chips acquiring the leading independent chip-portability software would put cross-vendor support commitments under scrutiny — customers evaluating hardware choices would need documented guarantees that Decart keeps optimizing for Nvidia and Google hardware after an Amazon acquisition. Watch for that language if a deal is announced. Its absence would tell you everything.

What happens next

Three scenarios. One: Amazon closes the deal and Decart becomes the portability layer for Trainium's enterprise push — good for AWS, nervy for everyone else. Two: the talks leak, a bidding war starts, and the price goes higher — Anthropic already looked, and Google has its own TPU incentive to keep Decart independent. Three: nothing happens, and Decart keeps raising as the Switzerland of AI chips, which might be the most valuable position of all.

For everyone else in the stack, the signal is clear: the abstraction layers are where the strategic value is accumulating. Models get cheaper and more numerous every week — this week's open-weight flood proved that again. Hardware gets more varied. The software that lets you move between both without rewriting everything is worth $7 billion. Build accordingly.