Anthropic Walked at the 90th Minute of a $6B Deal. Only One Reported Reason Was the Technology

The reported facts of the Decart deal set against the unconfirmed explanations — on the left a valuation path from roughly $4 billion in May to a reported $6 billion offer to no deal at all, a roughly 50% premium withdrawn after due diligence, for a three-year-old company of around 100 people building chip efficiency software and world models with revenue in the tens of millions, and on the right the four speculated reasons, that the technology may not hold up at scale, that the buyer could build it, whose money sits on the buyer's cap table, and whether the founder would stay — Everything on the right is unconfirmed. That is precisely why the shape of the list is worth reading — it is what experienced people thought plausible.

Most acquisition coverage tells you what a company sold for. This one is about a deal that did not close, which is rarer and, for founders, more useful.

Anthropic reportedly ended talks to acquire Decart for around $6 billion — what would have been its largest known acquisition — after going deep into due diligence. Israeli press described it as almost signed, walked away from in the 90th minute.

Neither company has said why. What we have instead is a list of things people close to it thought plausible, and that list is the interesting artefact.

1. What is actually reported

The Decart deal, as reported

The confirmed shape, before any explanation

Decart raised $300 million in May 2026 led by Radical Ventures, taking total funding above $450 million, at a valuation reported near $4 billion.
Anthropic's reported offer of around $6 billion was therefore roughly a 50% premium, months later.
The company is about three years old, has around 100 employees, and revenue described as several tens of millions annually.
Its products are the Decart Optimization Stack — chip-agnostic software for running models more efficiently across Nvidia, Google and Amazon silicon — plus real-time world models called Lucy and Oasis.
Talks ended after due diligence, and a person familiar told Bloomberg the two could still pursue other forms of collaboration.
Nvidia was reportedly an earlier suitor prepared to offer more, and is also a Decart investor from the May round.

Worth noting the timing: this lands weeks before Anthropic's anticipated IPO, with reporting pointing at mid-October marketing. A large, complicated acquisition immediately before a roadshow is a different proposition from the same deal a year earlier, and that is true regardless of anything about Decart.

2. The premium that stopped existing

The number I keep looking at is not the six billion. It is the gap.

Worth $4B in May. Offered $6B in September. Sold for nothing.

Decart is, by every indication, fine — a well-funded company with real revenue, real investors and a live product line. Nothing about the walk-away necessarily says otherwise. But an offer at a 50% premium arrived and then evaporated after the buyer looked closely, and that is a specific kind of event that founders rarely get to observe from outside.

The uncomfortable part is that a signed term sheet is not a price. It is a price conditional on what diligence finds, and diligence looks at considerably more than the thing you spent three years building.

3. The four explanations, none of them confirmed

Reporting floats several possibilities. I want to be exact here: these are unconfirmed, both companies declined to comment, and I have no independent knowledge of which if any is true.

Read the list, not the verdict

Even as pure speculation the set is informative, because it is what people who watch these deals considered plausible enough to print. Sort them by whether a founder could have done anything about them and the picture gets bleak fast: the technology performing at scale is the only one that sits squarely inside the company's control, and even the founder-commitment question is a judgement someone else makes about you.

One: the technology may not hold up. That it underperformed at scale, or on non-Nvidia chips. This is the only candidate that is genuinely about the product, and it is the one every founder prepares for.

Two: the buyer decided it could build it. The same calculation I wrote about when 32% of organisations cancelled a software purchase to build it themselves — except applied to a whole company rather than a subscription. A frontier lab with enormous infrastructure talent asking "or we could just do this" is build-versus-buy at its most expensive.

Three: geopolitics on the buyer's cap table. Questions reportedly touching Qatar's investment in Anthropic and the acquisition of an Israeli company. Whatever the merit, note the structure: a risk located entirely on the other side of the table, invisible in your data room, unfixable by you.

Four: whether the founder would stay. A reported concern that the founder was more interested in the transaction than in a long run inside the acquirer. Retention is a standard diligence question and it is answered by impression, not by documentation.

4. What diligence actually tests

Under someone else's conditions

The product

Not whether it works, but whether it works on their hardware, at their scale, against their benchmarks, run by their engineers. Chip-agnostic is a claim that gets tested on the chips the buyer actually owns. Know where your performance story is load-bearing and where it is extrapolated, before someone else finds out.

To buying you at all

The alternative

Every diligence process runs an implicit build case in parallel. The more your value is engineering the buyer already has, the more attractive that case looks after they have seen exactly how you did it. Diligence is also the most detailed technical briefing you will ever give a competitor.

Specifically, whether you stay

The people

For a company of about a hundred people, the acquirer is largely buying a team and its momentum. If the read is that the founder is optimising for the exit rather than the next four years, the asset being purchased partly evaporates on close. That read forms in meetings, not in documents.

Neither side controls

The context

Regulatory posture, investor composition on both sides, and what else the buyer has in flight. An acquirer weeks from an IPO has a different appetite than the same acquirer six months out. None of this is about you and all of it can end the process.

5. What a founder can and cannot control

The claims you can survive

Controllable

The single highest-leverage preparation is making sure every performance claim in your materials is one you would be happy to have re-run by a hostile engineer on unfamiliar hardware. Anything that only holds in your configuration should be described that way well before diligence, not discovered there.

What you reveal, and when

Partly controllable

Staging technical disclosure against deal certainty is ordinary practice and worth actually doing. The deepest internals belong late, after exclusivity and after the price is real, because a process that ends still leaves the other party knowing how you work.

Everything on their side

Uncontrollable

Their cap table, their IPO calendar, their internal build faction, their regulators. The only real defence is not needing the deal — a company with revenue and its own funding can absorb a walk-away, which is exactly the position Decart appears to be in.

That last point is doing quiet work. Decart raised $450 million and has tens of millions in revenue. A deal collapsing at this stage is a bad week, not an existential event. The same event at a company that had already spent its runway assuming the close would be a different story entirely.

6. What I would not conclude

This is not evidence that anything is wrong with Decart. Deals die for reasons that have nothing to do with the target's quality, and every explanation in circulation is unconfirmed. The company raised at a near-$4 billion valuation four months ago with Nvidia and Sequoia among its backers, and none of that has been retracted.

"Only one reason was the technology" is a claim about the reporting, not about reality. The actual reason might be the product, or something never mentioned publicly, or several things at once. What I can defend is that the plausible-looking list is mostly not about the product — which tells you what observers expect diligence to turn on.

Nobody involved is talking, and that is normal. Two companies declining to comment on a collapsed transaction is standard practice, not a signal. Treat all of this as one well-sourced report about a private process, and hold it loosely.

The honest summary

A three-year-old company was worth roughly $4 billion in May, was reportedly offered around $6 billion in September, and is now worth whatever it was worth before — which, to be clear, is a lot.

The founder-relevant part is what the near-miss exposes. Diligence is not a product review. It is a review of your product, your team's next four years, the buyer's alternative to buying you, and a pile of circumstances on their side that you will never see and cannot influence. Of the four explanations being floated for this one, exactly one lives inside the company being bought.

The practical response is not to prepare harder for the technical review, though you should. It is to build a company that does not need the transaction to close — because the reason it might not close is quite likely to be something you were never in a position to fix.

Sources: Bloomberg, "Anthropic Said to Walk Away From $6 Billion Decart Acquisition" · Calcalist, "What made Anthropic walk away from the $6 billion Decart deal?" · The Deep Dive on the terminated talks and Decart's funding history · PYMNTS on the scuttled acquisition · Deal figures, funding history and company details are as reported; every explanation for the walk-away is unconfirmed speculation, and the argument about what diligence tests is mine. For context on the other large AI infrastructure deal in play, see NVIDIA's Hugging Face acquisition.

NVIDIA Isn't Buying Hugging Face's Revenue. It's Buying the DefaultAugust 29, 2026

A Third of Your Buyers Have Already Cancelled a Purchase to Build It With AgentsSeptember 1, 2026

World Models in AI: The Next Frontier Beyond LLMs (2026 Guide)June 19, 2026

IdeaToMVP Academy

Want to build with AI — not just read about it?

4-week live cohort for founders. Learn to ship AI agents, scope MVPs, and automate your business — taught by the same team that writes these guides.

Explore the Academy →