Amazon nears a $7.5B deal for Decart, the startup Anthropic passed on

Share
Amazon nears a $7.5B deal for Decart, the startup Anthropic passed on

Six weeks ago this was a $7 billion rumor. Now the founders have agreed to stay five years and the paperwork is close enough that a signature could land within days.

Amazon is in advanced talks to buy Decart, the Israeli startup whose software moves AI workloads between different chips, for more than $7 billion and potentially close to $7.5 billion, the Financial Times reported on November 21. The FT says a deal could come together as soon as next week. Calcalist, which has been ahead of this sale since August, reported on November 18 that the legal structure was reworked in recent weeks, that most outstanding issues were resolved, and that a signing could come as early as this weekend.

That is a long way from where this started. In August, Bloomberg reported Anthropic was in talks to buy Decart for about $6 billion — a move read at the time as the model lab buying an entry point into Israel ahead of its listing. Anthropic completed due diligence, exchanged drafts, and then walked away in early September. Decart tried to reopen talks with Nvidia, already an investor. Nvidia declined. Amazon, which had given the startup early access to its Trainium 3 chips in 2025, stepped in, and the price went up.

The price ladder is the story

Read the sequence as a market signal rather than deal trivia. Decart was valued at $4 billion in May, when it had raised more than $400 million from Sequoia, Benchmark and Radical Ventures. Anthropic's reported offer in August was roughly $6 billion, a 50% premium in three months. Amazon's number is north of $7 billion, and Calcalist puts the total near $7.5 billion once earnouts are counted.

So the company that the most AI-literate buyer in the market examined and declined is now being bought for roughly 25% more than the price it declined. Either Amazon sees something Anthropic's diligence missed, or Amazon is paying for a strategic asset whose value does not show up in a revenue line. Both readings are consistent with the deal's structure.

What Amazon is actually buying

Decart sells two things that look unrelated until you watch how AI compute is being bought. The first is porting software: code that lets a customer move a trained model from one vendor's accelerator to another in weeks rather than months. The second is world models — Lucy and Oasis — that generate interactive video in real time, which is a compute-hungry demo of the same underlying skill.

The porting layer is the one that matters to Amazon. AWS builds its own silicon, Trainium 3 included, and its pitch has always run into the same wall: Nvidia's advantage is not peak throughput, it is that everyone's software is already written for it. Every hour of engineering a customer must spend to move a workload is a reason to stay. A company whose product is removing that cost attacks the moat directly, and it does so for every chipmaker at once, which is why Decart's customer list is heavy with semiconductor companies.

Calcalist puts Decart's revenue in the low tens of millions of dollars, growing unevenly, mostly from projects for chip companies. Against a price of $7 billion-plus, that is a multiple well north of 100 times sales — the kind of number that only makes sense if you are buying a capability rather than a business.

The terms say what kind of deal this is

Founders Dean Leitersdorf, his brother Orian and Moshe Shalev have committed to stay with Amazon for five years, according to Calcalist's reporting on the terms. A significant portion of the consideration is paid only in the fifth year, and the three have agreed not to start a competing business.

That is an acquihire wearing an acquisition's clothes. You cannot diligence a compiler; you can only keep the people who wrote it, and the deferred payment is the mechanism for keeping them. It also explains the delay: the recent holdup was not price but the founders' employment terms and how they would work inside Amazon for half a decade.

There is a second, less comfortable reading of the same clause. If the software were the durable asset, Amazon would not need a five-year lock on three people to protect it. The lock exists because the knowledge walks.

The counter-case

None of this is signed, and the reported price is for a company whose revenue is estimated in the low tens of millions of dollars, in a market that just repriced everything AI-adjacent. The Nasdaq 100 fell 10.4% last week, its worst week in more than six years, and Anthropic — the buyer that walked — closed its second trading day at $88.91, down 8.5% and barely above its $83 IPO price. Amazon would be paying a record multiple into that drawdown.

The harder question is why Anthropic left. The reporting says the talks ran into difficulty after due diligence, not that the price was wrong, and Anthropic has not explained it publicly. A buyer with the deepest technical bench in the industry looked at the code and the contracts and concluded the asset was not worth $6 billion. Amazon is now paying more than that. One of the two is wrong, and the market will not tell us which for a while.

There is also the regulatory layer. A $7.5 billion acquisition of an Israeli company by a US hyperscaler will draw a US antitrust look, and the deferred, target-linked portion of the price is exactly the structure that reviewers scrutinize. Neither company has commented, and no filing has surfaced.

Why it matters beyond the deal

The middleware of AI compute is consolidating on both ends of the stack at once. AMD agreed to buy World Labs for $8.2 billion in September; Stripe bought the model-routing layer underneath OpenRouter; and now the porting layer goes to the company that most needs Nvidia's customers to feel mobile. Each of those purchases is a bet that the durable profit sits above the silicon and below the model — in the parts that decide which chip runs what.

For Amazon specifically, Decart is a hedge that costs nothing operationally and buys optionality. If Trainium 3 lands, the porting software makes switching cheap and AWS captures the workload. If it does not, Amazon owns a tool that makes every other vendor's silicon easier to swap, which is still a business. What it does not do is sell more GPUs, and Nvidia's fourth-place market position this week is a reminder that the layer above the chip is where the next round of margin gets fought.

What to watch: whether the deal is confirmed at the reported price or trimmed, how much of the consideration is contingent, and whether Decart's porting tooling shows up in AWS's own Trainium documentation before the year is out.

If you were Amazon, would you buy the porting layer or keep building silicon and let the software problem solve itself? Write to us via our contact page.

We first covered this sale when it was a $7 billion rumor in October — Amazon in talks to buy Decart for about $7 billion.

Get SiliconNoon in your inbox

One email a day with the AI stories that matter — what happened, the numbers, and why it matters. No spam, unsubscribe in one click.

Read more

Nvidia Rubin HPC: 50,000 CUDA cores for science, shipping 2028

Nvidia Rubin HPC: 50,000 CUDA cores for science, shipping 2028

Nvidia used SC26 to do something it has not done in years: talk about simulation instead of tokens. The GPU it showed off will not exist until 2028. The software it shipped alongside it exists today. Rubin HPC is a core-count play, and Nvidia is not shy about it. Announced at the SC26 supercomputing conference in Chicago, Rubin HPC is a new part in the Vera Rubin platform aimed at simulation and data processing rather than matrix multiplication — more than 50,000 CUDA cores on a single GPU, a…

Databento raises $100M at $1.1B: the market data five AI labs buy

Databento raises $100M at $1.1B: the market data five AI labs buy

The vendor that sells tick-level prices to quant funds turns out to be infrastructure for frontier AI — and it just got priced like it, with fewer than 40 employees. Databento raised $100 million at a $1.1 billion valuation and says it now holds enterprise contracts with five of the six largest frontier AI labs, the company announced on November 12. Fifth Down Capital, an investor since the company's earliest days, led the round.

Nebius Q3: revenue up 571%, but 2027 capex jumps to $40-45 billion

Nebius Q3: revenue up 571%, but 2027 capex jumps to $40-45 billion

Nebius just posted the best quarter of its life as a public company and gave back 10% of its share price for it. The numbers beat. The plan for 2027 is what the market refused to pay for. The quarter was a genuine beat Nebius reported third-quarter revenue of $980.8 million, up 571% from $146.1 million a year earlier, ahead of the $907.5 million LSEG consensus, according to the company's results release and CNBC's report on the print.

TSMC buys $500 million of Anthropic and OpenAI stock

TSMC buys $500 million of Anthropic and OpenAI stock

Two $500 million checks, both signed at the same board meeting, both for equity rather than capacity. The amounts are almost irrelevant. The direction of the money is not. TSMC's board approved up to $500 million of Anthropic stock and the same again in OpenAI on Tuesday, buying equity in the two AI labs whose accelerators it fabricates — a $1 billion loop that the foundry's own filing records in a few lines and explains not at all. The terms sit in TSMC's Form 6-K.

Applied Compute: $100M ARR in 18 months, $350M at $3.25B

Applied Compute: $100M ARR in 18 months, $350M at $3.25B

The funding cycle usually hands you a narrative. On Monday it handed over a number instead: Applied Compute says it crossed a $100 million annualized revenue run rate less than 18 months after it was founded, and it raised $350 million at a $3.25 billion valuation on the strength of it. Applied Compute closed $350 million at a $3.25 billion valuation, led by Elad Gil, with Nvidia, Blackstone, Kleiner Perkins, Lux Capital, Shopify Ventures and HubSpot Ventures among a long participant list…

Marvell buys Lumentum for $115B in all-stock AI optics deal

Marvell buys Lumentum for $115B in all-stock AI optics deal

Marvell has agreed to buy Lumentum, the laser and optical-component maker, in an all-stock deal with a $115 billion enterprise value. It is the largest bet yet that the next bottleneck in AI data centers is not the chip but the light path between chips. Marvell Technology will acquire Lumentum in an all-stock transaction that values the target at roughly $1,132 a share — a 23.9% premium over Lumentum's closing price on October 28, the last trading day before speculation about a deal began…

Alphabet Q3: free cash flow -$9.07B, Gemini at 1 billion users

Alphabet Q3: free cash flow -$9.07B, Gemini at 1 billion users

Alphabet's AI buildout finally showed up on the cash flow statement, and the market did not like the line item. Alphabet reported negative free cash flow of $9.07 billion for the third quarter, after spending $51.36 billion on property and equipment against $42.29 billion of cash from operations, according to the company's Q3 2026 earnings release. Revenue rose 25% to $127.58 billion but missed the $128.22 billion analyst consensus, and adjusted earnings of $2.67 a share missed $3.02.