Amazon's $7B Play for Decart: Buying What Anthropic Rejected

A $7 billion handshake that isn't one yet

Let's be clear about what actually happened this weekend: nobody signed anything. The Wall Street Journal reported on Sunday, citing people familiar with the matter, that Amazon is in talks to acquire Decart — an AI startup most people outside the industry have never heard of — in a deal valued at around $7 billion. The report, carried by Reuters, added the sentence that should make any careful reader pause: a deal could happen soon, if talks do not fall apart or another suitor emerges.

That hedge isn't boilerplate. It's the entire story. Because the last big AI lab to look seriously at Decart did exactly what that hedge warns about: it walked away. Anthropic explored buying Decart for roughly $6 billion in August, Bloomberg reported, ran due diligence — and in September decided not to pursue it. So the question hanging over Amazon's weekend isn't just whether the talks hold. It's what Amazon thinks it sees in this company that Anthropic, having opened the books, decided wasn't worth the price.

Amazon declined to comment outside regular business hours. Decart did not respond to requests for comment. Reuters said it could not independently verify the report. In other words: treat this as reported talks, not a done deal, and read everything below with that in mind.

What Decart actually sells (and why it matters to AWS)

Decart was founded in San Francisco in 2023 by three Israeli engineers: brothers Dean Leitersdorf, the chief executive, and Orian Leitersdorf, and Moshe Shalev, the chief product officer. The company has raised about $450 million in total, according to Calcalistech — most recently $300 million in May, in a round led by Radical Ventures that valued it at nearly $4 billion (PitchBook data, via the Journal), up from $3.1 billion the previous August. The investor list reads like a who's who of the AI economy: Nvidia joined as a new investor in May alongside Atreides Management, Valor Equity Partners and Adobe Ventures, with earlier money from Sequoia Capital, Benchmark and Zeev Ventures.

Decart makes two kinds of things, and the distinction matters. First, there's the unglamorous but lucrative bit: systems software that squeezes more performance out of AI chips and — this is the part Amazon cares about — lets AI workloads move between chips from different vendors with minimal re-engineering. The company's own pitch is that it helps developers "squeeze every ounce of performance from every chip." "Decart's innovation makes AI generation not only more efficient, but also more accessible for any type of user," Victor Lazarte, a general partner at Benchmark who led an earlier round, said in a statement.

Then there are the world models — the flashy stuff. Lucy, backed by Nvidia, edits live video in real time: point it at a live stream and it can generate high-resolution video that makes it look as if the person on screen is trying on a dress or a bag, a task Bloomberg noted has long bedevilled fashion e-commerce. eBay, an investor in Decart, is also a customer, and streamers like CodeMiko use the technology on Twitch, TikTok and YouTube. Oasis, the older sibling, generates simulated 3D environments for training and testing robotics and autonomous-driving systems. On stage with Bloomberg at the Raise AI conference in Paris in July, Dean Leitersdorf — speaking, as it happens, to the very newsroom that would later break the Anthropic walkaway — explained the company trains its AI on text and millions of hours of video to learn the physical properties of the real world. "I really can't imagine being in any other market," he said. "This market is moving so fast. We get to reinvent every different aspect of the economy in the next 18 to 24 months."

We pulled the Commons file record and the 1,280-pixel original of the photograph used in this piece ourselves before writing — an aerial shot of AWS's us-west-2 availability zone in Morrow County, Oregon, three data centres with a fourth under construction. It is a real, licensed photograph, not a stock render, and it's worth staring at: this is the physical footprint Amazon wants Decart's software running on.

Why Amazon — and why now

Here's the thing about Amazon's interest: it didn't come out of nowhere. The Journal reported that AWS gave Decart early access to Trainium3, Amazon's third-generation custom AI chip, in 2025 — after, in the paper's telling, being impressed by the startup's founders. That detail is the tell. Amazon doesn't hand early silicon access to startups for the fun of it; it does it to learn whether the software on top of its chips is worth owning outright.

Trainium is the centre of Amazon's entire AI strategy, and Decart is the missing layer on top of it. AWS unveiled Trainium3 at re:Invent last December, promising 4.4 times the compute of the previous generation and fourfold power-efficiency gains, and has claimed customers can save up to 50% on training costs versus GPUs. The chip works. The problem is everything around it: customers have spent a decade building on Nvidia's CUDA ecosystem, and moving workloads off it is expensive, fiddly and slow. Decart's portability software is purpose-built to erase that friction — which would turn Trainium from a cheaper chip into a genuinely easier switch. That's the difference between competing with Nvidia on price and competing on pain.

This is the same playbook Amazon has been running all year. The company that spent $6.5 billion on MGM in 2022 just announced an $11 billion deal for satellite firm Globalstar. AWS landed Synopsys' business earlier this month in a day that sent the chip-design stock soaring. And Nvidia, the company Amazon is really buying against, has spent the year defending its alliance strategy as hyperscalers build alternatives to its hardware. A $7 billion Decart deal — nearly double the startup's May valuation — says Amazon believes compute leverage, not models, is now the scarcest asset in AI.

The uncomfortable question: what did Anthropic see?

Now the awkward part. Before Amazon, it was Nvidia that reportedly came close to buying Decart, per Calcalistech. Then Anthropic emerged as the preferred bidder in August, with the deal expected to value the company at $6 billion to $7 billion — and then, after due diligence, walked away in September. The Israeli press even reported that SpaceX had been in talks; Elon Musk called that report "fake news" on X. So Decart has now been passed over by Nvidia, rejected by Anthropic, denied by Musk, and is suddenly worth $7 billion to Amazon. Something doesn't add up — or rather, something adds up differently for different buyers.

We should be honest about what we don't know. Nobody outside Anthropic's deal team knows what the due-diligence process turned up — whether it was the price, the technology, the overlap with Anthropic's own inference work, or simply a strategic judgement that a model lab had no business owning an infrastructure company. One detail from the August coverage is suggestive: had the Anthropic deal gone through, Decart's workforce would have joined Anthropic's inference and performance division. That's a talent-and-tooling acquisition, not a product one. For Amazon, though, the calculus is inverted. AWS doesn't need Decart's models; it needs Decart's portability layer sitting between its Trainium chips and every enterprise workload in the world. The buyer who needs it most is rarely the one who builds models — it's the one who sells the picks and shovels.

And yes — Nvidia's own bruising acquisition record is relevant context here. Nvidia wanted OpenRouter in July; Stripe closed the deal instead, and Nvidia has since gone on a nine-figure shopping spree across the AI stack. In this market, losing a target doesn't mean the target was bad. It often means the bidding was.

Our take: this is a declaration of war on the CUDA moat

Strip away the deal theatre and Amazon's logic is brutally simple. Nvidia's real moat was never its chips — it's the software ecosystem that makes leaving those chips painful. Every startup that dies on the CUDA lock-in hill proves it. Decart is one of the very few companies whose entire business is making that hill flat. If Amazon owns it, Trainium stops being the cheaper, harder option and becomes the cheaper, easier one — and every AWS customer with a GPU bill becomes a sales lead. That's worth a premium, even a colossal one.

The risk is equally simple, and it's the one the Journal's hedge is really about: $7 billion is a lot of money for a company whose last serious suitor walked away after looking under the bonnet. If these talks collapse the way Anthropic's did, Amazon will have spent its credibility on a company the market's smartest lab already passed on. But if they hold, this is the most aggressive move yet in the hyperscalers' quiet rebellion against Nvidia — and the clearest signal that the AI chip war's next phase will be fought in software, not silicon.

What happens next

Three things to watch. First, whether a signed deal materialises — and at what price; the $7 billion figure is a talks number, and talks numbers move. Second, whether another suitor emerges, which the Journal's sources explicitly left open; Nvidia and the other hyperscalers know exactly what Decart is worth to AWS. Third, scrutiny: a $7 billion AI-infrastructure acquisition by a company of Amazon's size will attract regulatory attention on both sides of the Atlantic, and the recent history of big-tech AI deals suggests nothing closes quietly anymore.

We'll update this story as it develops.

By Daniel Rao
Daniel covers AI infrastructure and the business of compute for TechPulse USA — chips, clouds, and the money moving between them.

Sources: The Wall Street Journal (via Reuters), Bloomberg (via The Hindu BusinessLine, Moneycontrol, Communications Today), Calcalistech, The Jerusalem Post (Walla! Money), Data Center Knowledge, Decart.

Update, Oct. 12: what the filings actually say about Decart

The deal is still just a negotiation — the Wall Street Journal's scoop remains single-sourced, Reuters "could not immediately verify," and Amazon and Decart have both declined to comment. But two days of re-syndication have filled in who exactly Amazon is circling. Decart was founded in 2023 by brothers Dean and Orian Leitersdorf and Moshe Shalev, Israeli engineers. PitchBook valued the startup at nearly $4 billion in May — so a $7 billion price tag would be roughly a 75% premium in five months. The models in question: Lucy, a real-time video-editing model backed by Nvidia, and Oasis, which builds simulated environments for robotics and autonomous-vehicle training. And the cap table is already stacked: Decart raised $300 million in May in a round led by Radical Ventures, with Nvidia joining as a new investor. If the talks hold, the expectation is it could happen soon — and Amazon would own the one video-generation shop Anthropic passed on.

Related: Nvidia's talks with Reflection AI — another Nvidia-adjacent startup deal in the same week — and Meta AI's distribution play for the competitive context.

One more detail from the Israeli press, not yet in this post: Decart's founders reportedly turned down an acquisition offer from Nvidia before Anthropic's bid, preferring Anthropic's all-stock proposal. Anthropic then walked away after due diligence — which is how Amazon ended up at the table.

Search This Blog

TechPulse USA