Bloomberg reported this morning, August 13, that Anthropic is in talks to buy Decart AI for around $6 billion. Talks, not a signed deal. That distinction matters and I will come back to it.
What caught my attention is not the number. It is where Decart came from.
The Minecraft thing
Decart got famous for Oasis: a playable Minecraft-looking world that no game engine was rendering. The model predicted every next frame based on what you pressed on the keyboard. 20 FPS, interactive, no scene graph, no collision system, no assets. Just a model hallucinating a consistent world fast enough that your hands believed it.
In late 2024 that read as an impressive demo with no obvious business behind it.
The company was founded in 2023. It has raised over $450M, was valued at $3.1B before this year's round, and its current research page describes three product lines:
- Oasis, a world model, now explicitly positioned for physical AI and robotics rather than gaming
- Lucy, a real-time video model running live at 30 FPS
- DOS, the Decart Optimization Stack: hardware-aware model design, custom kernels, proprietary compilers, inference optimization
The demo was the marketing. DOS is the engineering.
The reported reason is not robotics
Read the actual reporting carefully. Fortune says a deal would bring Decart's video-simulation and chip-efficiency technology into Anthropic's inference team. Bloomberg's sources point at the same thing: the chip efficiency work could help existing infrastructure absorb more demand.
So the sourced story is compute economics. Anthropic is compute constrained, spending enormously on capacity, and DOS is a margin lever that applies to every single Claude request on day one.
That is a boring, completely rational reason to spend $6B. It does not need a robotics narrative at all.
I still think the robotics reading is in there.
Why
Two things sit underneath.
First, Anthropic held acquisition talks with Physical Intelligence this spring. The Information reported it, the CEO denied the specific weekend rumor that blew up on X in July, but the talks happened. That is a robot foundation model company, valued around $11B, whose pi0.5 is one of the more widely used robot brains in research labs. You do not open that conversation by accident.
Second, and this is the constraint people skip: OpenAI is already a shareholder in Physical Intelligence. Founders Fund and Thrive are in there too. If a bidding contest for the robotics foundation model layer is happening, Anthropic may simply not be allowed to win that particular one.
If you cannot buy the robot brain, buy the world the brain gets trained in. And take the inference savings as immediate return while the option matures.
Three different bets on physical AI
This is the part I find genuinely useful, because "everyone is going into robotics" is a lazy summary. There are three distinct strategies running right now.
Vertical model (Mistral). On July 8 they shipped Robostral Navigate, an 8B navigation model. Single RGB camera, plain language instructions, hardware agnostic, trained entirely in simulation, 76.6% on the R2R-CE benchmark. They are hiring a robotics team and selling into Airbus and BMW. Narrow scope, deployed now, revenue this year.
World model (Decart). Learn the dynamics of the world from video, generate frames conditioned on actions. You are not building a robot. You are building the environment the robot learns in, plus the synthetic data to train on. Slower, more general, much bigger if it works.
General scaling (Anthropic). Do not build robotics at all. Measure whether the capability falls out of the general model on its own.
That third one sounds like a joke until you read Project Fetch Phase Two. In August 2025 their model could not connect to a robot dog's sensors. By June 2026 a newer model finished the whole task set in under ten minutes, roughly 20x faster than the best human-plus-Claude team from the year before. Anthropic states plainly that this did not come from any deliberate effort to improve robotics. It came from general scaling.
Read as strategy, that paper argues against founding a robotics division. It argues for more compute and better simulation. Which is exactly what Decart sells.
The distinction nobody is making
Here is the part I keep waiting to see in the coverage and do not.
In Project Fetch, Claude was not the control policy. Claude was the programmer. It wrote and debugged the code that drove the robot. The loop running at 50Hz on the hardware was ordinary software that a model happened to author.
Those are completely different capabilities:
- Model as control loop: perceive, decide, actuate, every tick, under latency budget
- Model as author of the control loop: read docs, write code, test, iterate, at human timescales
Almost every headline number about LLMs and robotics is measuring the second one and getting reported as the first. And Anthropic's own follow-up work, Claude plays robotics, is explicit that precise physical manipulation and closed-loop control remain a hard limit. The models are good at writing robot code. They are not yet good at being robot code.
This is precisely where a world model changes the picture. You cannot scale a control policy on internet text, because control policies need consequences. They need an environment that responds to actions, cheaply, millions of times. That is what Oasis is, structurally. A frame predictor conditioned on actions is a simulator you did not have to write physics for.
So the Minecraft demo and the robot training environment were never two different products. They were the same artifact with two different customers.
What I do not know
Being honest about the limits of this, since it is a story that is hours old:
- These are talks. They can fall apart. Nothing is signed.
- The Project Fetch numbers are self-reported and have not been independently replicated.
- The compute reading and the robotics reading are both fully consistent with the available reporting. I am arguing the second is present, not that the first is wrong. Most large acquisitions are over-determined.
- Disclosure: I use Claude Code daily and run a good chunk of my own tooling on Anthropic's models. I am not a neutral observer here.
The thing I will be watching is not whether the deal closes. It is whether Oasis stays a product with an API, or quietly turns into internal training infrastructure nobody outside gets to touch.
That would tell you which of the two readings was right.
Top comments (0)