Why Amazon's $7 Billion Decart Deal Rewards Middleware

On September 8, Decart's founders learned their company was not for sale. Anthropic had spent months negotiating to buy the Israeli AI startup for roughly $6 billion, then walked away after completing due diligence. Thirty three days later, The Wall Street Journal reported that Amazon is in talks to acquire the same company for around $7 billion. No new funding round closed in between, and no revenue figure was disclosed either.

That 33 day gap is the most instructive thing in startup news this week, because it separates two questions that routinely get tangled together. One is how good a company is. The other is who needs it. The first barely moved between September and October. The second one changed the price by roughly a billion dollars.

Decart is not an unknown quantity. It was founded in late 2023 by Dean Leitersdorf, Orian Leitersdorf and Moshe Shalev, engineers who came out of Israel's Unit 8200. It has raised more than $450 million, including a $300 million round in May 2026 led by Radical Ventures at a reported valuation near $4 billion. Nvidia is an investor. Amazon was already a strategic customer, which is how Decart got early access to Amazon's Trainium chips.

The reported $7 billion price works out to roughly a 75 percent markup on that five month old valuation, for a company that has never published a revenue number. Set that against what Amazon is spending this year: about $220 billion in capital expenditure, most of it AI infrastructure. The entire acquisition is a rounding error against a single year of Amazon's own capex, which is the first clue that this is not a normal purchase.

What the $7 Billion Is Actually Buying

Decart runs two product lines, and only one of them justifies the number. The first is a set of generative models: Lucy for real time video, and Oasis, which generates interactive simulated environments. The second is infrastructure software, the layer Decart calls its optimization stack, which makes models run faster and, more importantly, lets developers move workloads between chips built by different manufacturers.

That second product is the one a hyperscaler actually wants. Today, building AI software means building it for a specific chip, and switching vendors later is expensive enough that most teams simply never do it. A portability layer turns that one time cost into a configuration choice, which matters enormously to anyone spending nine figures a year on compute. Every company at that scale has a strong opinion about lock in.

Amazon has more than an opinion. Its custom silicon business, anchored by Trainium, passed a $25 billion annualized run rate in the June quarter and is growing at triple digit percentages. Amazon sells compute as well as buying it, and every workload it moves onto its own chips is revenue that doesn't go to Nvidia. Owning the software that makes leaving Nvidia cheap is a distribution tool, closer to a sales weapon than a cost saving.

Why the Same Asset Is Worth More to Amazon

Anthropic's interest was never mysterious. It buys enormous amounts of compute, and it has said publicly that it is looking at building its own chips to work around shortages. A tool that squeezes more performance out of hardware the company already owns is worth real money there. Think of it as a savings account, a known quantity with a defensible payback, which is why Bloomberg's reporting on the $6 billion talks framed the whole thing as a compute efficiency story.

For Amazon, the same capability functions as a sales argument rather than a cost reduction. When a customer can move a training job from one vendor's silicon to another without rewriting the pipeline, the thing keeping them loyal to Nvidia weakens. Market share plays get priced differently from efficiency plays, and that gap is probably most of the $1 billion separating Anthropic's reported offer from Amazon's.

The wider pattern explains why AI valuations keep detaching from earnings. Prices in this market get set by what a specific buyer's strategy requires, not by what a discounted cash flow spreadsheet can support. That dynamic runs well beyond chips, and it's the same force behind the valuations reshaping smaller AI businesses right now. Founders who model their own worth on multiples from a different market are working with the wrong map.

A Dead Deal Is Not a Verdict

The Anthropic collapse got read as a verdict on Decart, which it never was. Reporting in the Israeli business press afterward pointed to an ordinary fight about where the combined company's people and operations would physically sit after close. That kind of disagreement ends large acquisitions all the time, and it has almost nothing to do with whether the technology works.

Any founder holding a term sheet should internalize that. Deals die on headcount, reporting lines, earnout structures and which country the engineering team relocates to. They rarely die because a buyer stopped believing in the product. When a deal collapses, founders tend to ask what's wrong with their company. The better question is which operational clause nobody could agree on, since that one has answers.

Decart's real advantage was never a single buyer. It had Nvidia as an investor, Amazon as a strategic customer and Anthropic as a suitor, which meant the second conversation didn't start from zero when the first one ended. Founders who build one buyer relationship and call it a pipeline are one bad diligence process away from a very quiet quarter. Two or three live commercial relationships is a different negotiating position entirely.

What Founders Should Take From This

The transferable lesson is about placement inside a customer's budget rather than about valuation. Decart sat on compute cost, the largest expense line inside every AI company and every hyperscaler on the planet. Products that touch a customer's biggest cost get budgeted differently from products that touch the innovation line, and they tend to survive the cost cutting rounds that kill everything else. If your product disappears when a CFO gets nervous, that's information about where you sit.

the switch is worth more than the thing being switched. Interoperability, migration tooling and anything that reduces a customer's dependence on one vendor tends to gain value when that customer's strategy shifts, because strategy shifts are what turn switching costs into urgent problems. The unglamorous half of Decart's product line carried the price. The impressive demos earned the headlines.

take the strategic customer early even when the pricing is unflattering. Amazon's early access arrangement with Decart looked like a favor to Amazon at the time, since the world's biggest cloud provider was handing a startup hardware before most competitors could touch it. Ten months later that relationship functions as an extended job interview for a $7 billion role. Deep strategic discounts are sometimes the cheapest marketing a founder ever buys.

None of this should be read as permission to skip revenue. Decart has a scarce strategic position and found a buyer with a $220 billion capex budget and a reason to overpay for it. Most founders sell to operators, and operators want evidence that growth can fund itself. Treating this deal as proof that revenue doesn't matter would be the most expensive misunderstanding available this week.

What the Gap Between Buyers Proves

Back to where this started. Decart's founders went from believing their exit had died to negotiating at a higher number in roughly the time it takes to fill a hiring cycle. The software didn't improve in that window. The buyer list did, and the value of the company turned out to track that list more closely than anything on a pitch deck.

Amazon's talks could still fall apart, another bidder could appear, and Decart could end up independent with a good story and no exit. What the last five weeks do demonstrate is narrower and more useful: leverage at the table comes from who is standing behind you and what you own that they can't get anywhere else. If you're building something that touches a large cost inside a bigger company, the work isn't looking more impressive. It's making yourself expensive to remove. What Amazon is buying is the switch, and there are very few switches for sale.