The $7 Billion Exit

The Wall Street Journal reports that Amazon is in talks to pay $7 billionReportedWSJ, the 11th of October 2026 for a company that builds neither a model nor a chip. It builds the ability to leave one, and that may be the industry's most valuable product.

Key · companies amber

The Journal's report, published over the weekend, says Amazon is in advanced talks to buy Decart for around seven billion dollars. The piece is paywalled; I read it through four other outlets. It says a deal could arrive within days, or fall apart.

Seven billion is a strange price for a company this size: $450 millionConfirmedDecart's own funding page, May 2026 raised in total, valued near $4 billion in May, about a hundred people, no chatbot and no phone. What it sells sits underneath everything else: the ability to run the same model on a different chip.

What Decart actually built

Its core product is DOS, the Decart Optimization Stack. In most AI workloads a model is tuned to one maker's silicon; moving it is a project: re-porting has taken months of engineering; DOS compresses it to about two weeksReportedSiliconANGLE, the 11th of October 2026. Some estimates reach eight times faster; that figure depends on the workload. Treat the eight as a claim.

Customers don't care what chip we use.

That was Decart's chief executive, Dean Leitersdorf, in December 2025. Most software is built for one maker, and every rival, Amazon included, sells alternatives with one unspoken surcharge: switching will cost you. Decart sells the removal of it, across Nvidia, Google and Amazon silicon.

The price has a shape

Why Amazon wants the drawbridge

Amazon has built its own chips since 2015, when it bought Annapurna Labs, an Israeli designer, for around $350 millionReportedReuters, the 22nd of January 2015, on undisclosed terms. That became Graviton, Trainium and Inferentia. Decart is already a Trainium customer: its video model ran four times fasterReportedCTech, the 12th of December 2025 on the chip than on Nvidia's GPU.

A challenger chip does not beat a champion inside its fortress; it wins only if the drawbridge is open. Lock-in protects the leader; portability arms the challenger. The seven billion is for the drawbridge, and the sales pitch writes itself: try our chip, because you can leave.

The same weekend, the Financial Times reported Nvidia in early talks over Reflection AI, after it bought Hugging Face for $12.93 billionConfirmedSiliconANGLE, the 3rd of September 2026 and licensed Groq's technology for a reported $20 billionReportedSiliconANGLE, the 24th of December 2025. Everyone is buying the layer beneath the model.

Move the plug yourself

The same wall sits under my own desk: a 27-billion-parameterFirst-handmy own machine, not a test of DOS model runs on one consumer card, and the day I change that card, the move is mine. Multiply that by a million chips and you have Decart's market.

The honest counterpoint

The objection: the software that moves a workload off any chip also decides where it lands. An exit owned by one vendor is a corridor; today's lock-breaker can become tomorrow's lock, priced in AWS minutes. Any buyer would deny it; the structure would say it.

One hole predates the deal. In September, Anthropic reached the final stage of due diligence on buying Decart for about $6 billionReportedBloomberg via CTech, the 8th of September 2026, then walked away. CTech's own analysis listed the candidates, gains that shrink at scale, engineering deeper than the demo, and left it open. That is the fact that could make this story wobble.

What would settle it: independent benchmarks of DOS on Trainium and on Nvidia silicon, and confirmation from Amazon. What would sink it: the talks collapsing, or the porting gains failing outside the company's demos.

Every lock has a key, and the key is worth what the lock protects. For a decade, the industry's complaint about Nvidia was that leaving was impossible. The exit is now an asset, valued at seven billion dollars of reported intention, and its buyer sells locks too. The model was never the product. The way out was.

Moving your own workloads? I do that for a living: email brandon@kreostudio.co.uk.