Anthropic completed diligence on Decart's efficiency stack and then walked away from $6bn
What Anthropic wanted from Decart was cheaper serving, not a new product line, and the walk-away leaves the buy-versus-license question on compute efficiency sitting exactly where operators keep finding it.
The Product Desk
What happened
- Anthropic completed due diligence on Decart and then decided against a purchase first reported in August at around $6bn, Bloomberg reported, with representatives for both companies declining to comment.
- The attraction was Decart's optimisation stack, built to make chips work harder across training and inference so that Anthropic's existing compute could absorb more demand, rather than its world models.
- Decart raised $300m in May in a round led by Radical Ventures with NVIDIA among the new investors, valuing the company close to $4bn per the Wall Street Journal, up from $3.1bn the previous August.
- Anthropic is spending heavily on compute ahead of an initial public offering it expects to match or exceed SpaceX's record listing, according to people who described the preparations to Bloomberg.
- One person familiar with the talks told Bloomberg the two companies may still pursue other ways to work together.
Why it matters
- Who has to decide decision Any team weighing whether to buy the vendor that lowers its unit costs now has a public comparable in which the buyer got a diligence-grade look at an efficiency stack and still chose not to own it.
- What it limits constraint A listing window narrows which savings a company is allowed to buy, because a premium paid close to a prospectus has to be defended to incoming shareholders even where ownership beats a contract on the numbers.
- Who's exposed exposure Decart's backers, NVIDIA among them, now carry a valuation set four months before a $6bn bid that did not close, and the next round will be priced against a widely reported non-deal.
- What doesn't add up contradiction thenextweb.com treats the efficiency rationale as the substance of the deal, while Gizmodo reads the original leak as possibly serving one side of a negotiation, which changes how much weight the $6bn number deserves.
Next story · Invest · 5d ago
Amazon in talks to buy Decart for about $7 billion, maker of software that optimizes AI models across NVIDIA, Google and Amazon chipsConfirmed11 publishers
An efficiency layer is one of the few acquisitions where the buyer can check the arithmetic before signing. You know your compute bill and your utilisation rate, and a vendor promising to "squeeze every ounce of performance from every chip" [5] is making a claim that shows up in a line of your own accounts. Anthropic ran that check to the end of diligence and passed [1].
What ownership adds, on top of the saving, is exclusivity and control of the roadmap. A supply agreement gets the saving on its own, for a fraction of the money. When the acquired thing never touches a customer, the gap between those two prices is mostly what you are paying to stop someone else buying the same discount. Decart's stack sits under both training and inference and was meant to let Anthropic's existing compute absorb more demand [4], which per Bloomberg's reporting made this a deal about the cost of serving customers rather than entry into a new product category [6].
Anthropic's most notable acquisition to date was $400m for a team of fewer than ten people [12]. Against that, around $6bn [2] is roughly fifteen times its largest deal so far [18], and about $2bn over the valuation Decart's May round set [19], a mark-up of roughly half in four months [15]. A company close to a listing wants a cost base that looks efficient and an acquisition history that looks disciplined, and buying Decart would have improved the first while complicating the second [16]. Reporting does not say which of those considerations ended the talks, or whether the sticking point was price or something diligence turned up [8].
Gizmodo, which read the original Bloomberg story as the kind of anonymously sourced piece that might exist to help one side in a negotiation [17], supplies the right discount on the $6bn figure: someone with a use for it put it in circulation. Both accounts agree on the mechanism. The Decart Optimization Stack tunes the interaction between AI hardware and software to cut compute cost [10], and no user of the product ever sees it.
What separates ownership from a contract in a case like this comes down to whether most of the benefit is available by agreement, and whether a customer ever touches the thing. Decart's own two halves land on opposite sides of that line: Lucy takes live video of a person and generates high-resolution video of them wearing something they are not, with eBay as both investor and customer [9], while the optimisation stack is a cost line [10]. Anthropic wanted the invisible half [4]. The version of this decision that survives a board meeting starts with the licence quote and treats the gap up to the purchase price as the cost of exclusivity. Paying an equity multiple for a discount only makes sense if the buyer can name the rival who would otherwise have licensed the same stack.
What to watch
- Whether the mooted collaboration becomes a commercial agreement for the optimisation stack rather than ownership, which per thenextweb.com would be the cheaper version of the same idea and would not need explaining in a prospectus.
- Decart's next raise, and whether investors mark it above or below the close-to-$4bn set in May now that a $6bn bid is public and unclosed.
- The IPO timing Gizmodo reported as rumoured around Halloween, and what the filings say about how Anthropic plans to hold down serving costs instead.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence45
- Adoption18
- Hype gap+12
- Incentives72
- Confidence55
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Anthropic performed due diligence on Decart but ultimately walked away from a purchase, Bloomberg reported, citing people familiar with the matter who asked not to be identified.
- [2]
The Anthropic-Decart talks were first reported in August at around $6bn, with the caveat that nothing had been finalised and the negotiations could fall apart.
- [3]
One person familiar with the matter said Anthropic and Decart may still pursue other opportunities to collaborate.
- [4]
Decart's optimisation stack is designed to make chips work harder across both training and inference, and the idea, per Bloomberg, was to let Anthropic's existing compute absorb more demand.
- [5]
Decart's own website promises to "squeeze every ounce of performance from every chip".
- [6]
The deal was about the cost of serving customers, not about Anthropic entering a new product category.
- [7]
Anthropic is spending heavily on compute ahead of an initial public offering it expects to match or exceed SpaceX's record listing, according to people who have described the preparations to Bloomberg.
- [8]
Nobody involved has said which consideration ended the talks, and the reporting does not establish whether Anthropic balked at the price or at something diligence turned up.
- [9]
Decart's Lucy model takes live video of a person and generates high-resolution video of them wearing something they are not, the virtual try-on problem; eBay is both an investor in Decart and a customer.
- [10]
DOS, the Decart Optimization Stack, purportedly optimises the interactions between AI hardware and software, cutting down on compute cost.
- [11]
Representatives for both Anthropic and Decart declined to comment to Bloomberg.
- [12]
Anthropic rarely makes large acquisitions, and its most notable to date was $400m for a team of fewer than ten people.
- [13]
Decart raised $300m in May in a round led by Radical Ventures, with NVIDIA, Adobe Ventures, Valor Equity Partners and Atreides Management joining, alongside earlier backers Sequoia Capital, Benchmark and Zeev Ventures.
- [14]
The May round valued Decart at close to $4bn, per the Wall Street Journal, up from $3.1bn the previous August.
- [15]
A $6bn purchase would have been roughly a 50% mark-up on Decart's May valuation four months later.
- [16]
A company preparing to list has reason to want its cost base looking efficient and its acquisition history looking disciplined; buying Decart would have improved the first and complicated the second.
- [17]
Gizmodo had described Bloomberg's initial reporting on the talks as "one of those anonymously-sourced Bloomberg articles you can't help but suspect might exist solely to help out one or another side in an ongoing negotiation", adding that the deal could still crumble.
- [18]
At around $6bn, the Decart deal would have been roughly 15 times the $400m Anthropic paid in its most notable acquisition to date.
- [19]
A $6bn price is about $2bn above the close-to-$4bn valuation Decart's May round set.
- [20]
Gizmodo reported that Anthropic is planning what may soon be the world's largest IPO, rumoured to be coming around Halloween.
- [21]
Whether the collaboration the two sides are said to be still considering means a commercial agreement for the efficiency stack rather than ownership of it is the thing to watch; it would be the cheaper version of the same idea and would not need explaining in a prospectus.
Sources
2 independent publishers whose own reporting we read for this story.
- gizmodo.comNope, Anthropic Will Reportedly Not Buy That Un-Anthropic-Like Company
1 article · September 8, 2026
- thenextweb.comAnthropic has walked away from its $6bn Decart deal, Bloomberg reports
1 article · September 7, 2026
Topics and entities
Topics
- AI compute efficiency
- World Models
- AI Company IPOs
- AI mergers and acquisitions
Entities
- The Wall Street Journal
- Radical Ventures
- Bloomberg
- Nvidia
- Dean Leitersdorf
- eBay
- SpaceX
- Anthropic
- Decart Optimization Stack