Anthropic’s Reported $6B Decart Deal Is a Warning to Every AI Founder
If Anthropic is willing to pay about $6 billion for Decart, your AI startup is not being valued on revenue. It is being valued on who gets strangled without it.
A $6 billion price tag for a three-year-old AI startup sounds mad—until you realise the buyer may be trying to avoid a much bigger bill.
Anthropic is reportedly in talks to buy Decart for about $6 billion. If it happens, this will not be a cute acqui-hire or another Big Tech vanity purchase. It will be a blunt admission that, in AI, owning the bottleneck is worth more than owning the feature.
And that should make founders uncomfortable. Because most of the AI companies being funded today are building features. The serious money is moving toward the businesses that control compute, speed, distribution, proprietary workflow data, or the people who know how to make expensive models run cheaply.
The reported $6 billion bet on Decart
The reported deal puts Anthropic, maker of Claude, in discussions to acquire Decart at roughly $6 billion. The important word is reported: no completed transaction has been announced, and deals of this size can change, drag on, or die quietly in a boardroom.
But the number matters even before the paperwork is signed.
Decart has built itself around real-time generative AI. Its stack includes DOS, software designed to optimise training and inference, plus products including Lucy for real-time video generation and editing, and Oasis, a world model aimed at physical-AI use cases. The company said in May that it had raised $300 million in a round led by Radical Ventures, taking its total funding above $450 million.
That tells you what Anthropic would be buying. Not merely a video toy. Not merely a talented team. It would be buying a crack at making AI workloads faster and less costly at the exact moment model companies are finding that demand is wonderful right up until the electricity, chips and cloud invoices arrive.
Decart had already raised $100 million at a $3.1 billion valuation in 2025. A reported $6 billion sale price would therefore be a meaningful uplift in a short period, but it is not the wildest thing in this market. If a buyer believes Decart can materially improve inference economics across a massive installed base, the acquisition price is not really being measured against Decart’s last funding round. It is being measured against the cost of being slower, more dependent and less profitable for years.
That is the bit people miss when they look at a headline and mutter, “Another AI bubble.”
Maybe it is frothy. But strategic buyers do not pay billions because they enjoy lighting money on fire. They pay because the alternative can be worse.
AI has turned infrastructure into a hostage negotiation
For the past few years, AI companies have sold a simple story: make the model smarter, get more users, raise another monster round, repeat.
That story has a nasty practical problem. Intelligence is expensive.
Every prompt, image, video, agent task and automated workflow consumes compute. Training frontier models requires enormous capital. Serving them to millions of users creates an ongoing inference bill. And when your product becomes useful enough for enterprises to rely on it, they do not want occasional brilliance. They want speed, reliability, security and predictable costs.
That is why Decart’s position is interesting. Optimisation sounds boring beside a flashy chatbot. Good. Boring is often where the money is.
A model company can spend years improving benchmark scores and still have a lousy business if it cannot serve customers economically. Lower the cost per task, reduce latency, make more use of available chips, and improve the throughput of existing infrastructure: suddenly a company has more margin, more capacity and more control.
Those gains compound. If an optimisation layer lets Anthropic serve more requests with the same hardware footprint, it is not just saving dollars. It can price more aggressively, support more enterprise demand, reduce dependence on outside suppliers and reinvest the difference into product and compute.
That is a proper strategic asset.
I have seen versions of this in ordinary businesses. The company that owns the customer relationship looks powerful until the supplier raises prices. The company with the better sales pitch looks brilliant until fulfilment breaks. The sexy front end gets the headlines; the bottleneck gets the economics.
AI is no different. It is just wearing a hoodie and burning through power at industrial scale.
The second-order implication: AI M&A is becoming defensive
The overlooked point is that this sort of deal is not solely about growth. It is defensive.
Anthropic has been building an enormous business around Claude while competing against OpenAI, Google, Meta and a growing list of aggressive model and infrastructure players. In that market, depending entirely on partners for key parts of your technical stack can be dangerous. Partners can become competitors. Hardware capacity can tighten. Pricing can change. A rival can buy the company you were relying on.
So a reported Decart acquisition would make sense as a way to remove uncertainty from a critical part of the AI supply chain.
That should change how founders think about building companies in this cycle. The best acquisition targets are not always the businesses with the most obvious consumer brand. They are the ones that solve an expensive, recurring, ugly problem for a buyer that cannot afford to be exposed.
Ask yourself a harsh question: if your company disappeared next Tuesday, would a large customer be annoyed—or would their cost base, growth rate or strategic position get materially worse?
If the answer is “annoyed,” you have a feature.
If the answer is “materially worse,” you may have an asset.
There is a chasm between the two. Most founders spend years pretending it is a small step.
The contrarian view: $6 billion could still be cheap
Here is the sentence nobody wants to say while everyone is taking screenshots of valuations: a reported $6 billion deal could be cheap.
Not cheap in the normal human sense. Six billion dollars is a stupidly large amount of money. But cheap relative to the value of a durable infrastructure advantage inside a company operating at frontier-AI scale? Entirely possible.
The wrong way to judge this deal is to ask whether Decart has $6 billion of current revenue. That is how you assess a mature plumbing business or a retail chain. It is not how strategic technology acquisitions work when a buyer is trying to secure a capability that can influence future cost structure, speed of execution and competitive independence.
The right questions are uglier:
- How much compute spend could Decart’s technology save over five years? - How much faster could Anthropic deploy new products? - How much enterprise demand could it serve without building proportionally more infrastructure? - What would a rival pay to own the same capability? - What does it cost if Decart ends up inside somebody else’s walls?
That final question is usually the killer.
A strategic buyer often does not need a target to be worth the purchase price in isolation. It needs the target to be worth more inside a rival. That is why founders who only optimise for the next funding valuation can miss the real game. The prize is not appearing valuable on a pitch deck. The prize is becoming strategically intolerable to ignore.
Don’t confuse acquisition appetite with an exit plan
Now for the cold shower.
A handful of billion-dollar AI deals does not mean every startup with “agent” in its deck is suddenly saleable. It means the bar is becoming clearer.
Buyers will pay up for leverage. They will not pay up forever for interchangeable wrappers around the same underlying models.
If you are a founder, stop building your business around the fantasy that a giant will rescue you at a heroic multiple. Build something a buyer would hate to see in a competitor’s hands. That means real technical differentiation, deep integration into valuable workflows, measurable cost savings, difficult-to-recreate data, or a team with rare capability that actually ships.
And if you are an investor, be wary of companies claiming “strategic value” without naming the strategic pain they remove. A vague claim that AI is hot is not a moat. It is a fundraising adjective.
What this means for you
For founders: write down your customer’s three biggest cost lines, three biggest operational delays and three biggest risks. Then make your product attack one of them so directly that removing you would hurt. “Users like it” is not enough. You want, “Their unit economics get worse without us.”
For operators: treat AI cost and latency as product problems, not technical trivia. If a workflow is valuable but too slow or expensive to deploy broadly, it is not ready. Measure cost per completed task, not just model quality. The businesses that win will make AI useful at a price customers can live with.
For investors and savers: do not chase every company that says it has AI. Look for the toll roads: infrastructure, workflow ownership, distribution, proprietary data, and businesses reducing the cost of intelligence rather than merely decorating it.
And for everyone watching this reported Anthropic-Decart deal: remember the lesson. The next wave of AI wealth will not go only to the company with the cleverest demo. It will go to the businesses that control what everyone else cannot operate without.
That is where the real money always ends up.