Anthropic's $6B Decart deal is a robotics play disguised as a compute play

On August 13, Bloomberg reported that Anthropic is in talks to purchase Decart AI for approximately $6 billion. However, this figure represents the talks, not a confirmed deal. What stands out about Decart is its origin, stemming from a Minecraft-inspired world known as Oasis. This world is rendered through a model that predicts every subsequent frame based on user input, achieving 20 frames per second without any scene graph, collision system, or assets.

Initially, this demonstration held no apparent commercial value. Founded in 2023, Decart has garnered over $450 million in funding and was valued at $3.1 billion before the current round. The company's research page now outlines three product lines: Oasis, a world model positioned for physical AI and robotics; Lucy, a real-time video model running live at 30 FPS; and DOS, the Decart Optimization Stack, focusing on hardware-aware model design, custom kernels, and proprietary compilers.

Initially, the reported rationale for the deal was not robotics; rather, it was Anthropic's need to reduce compute costs. Anthropic faces immense computational demands, and DOS is touted as a potential lever to manage costs for every Claude request. While this explanation appears rational, the robotics narrative may still be present.

Anthropic had acquisition talks with Physical Intelligence, a robot foundation model company valued at around $11 billion. The founders of Physical Intelligence are part of OpenAI's shareholder group. If Anthropic can't acquire the robotics foundation layer, it might opt to purchase the world it operates in instead. This approach emphasizes three distinct strategies in physical AI: vertical models like Mistral's 8B navigation model, world models like Decart's, and a general scaling approach seen in Anthropic's Project Fetch.

In Project Fetch, Claude wasn't merely a control policy but actually wrote and debugged the robot's code, demonstrating a unique capability. This distinction between a model as a control loop and one as an author of control loops was not emphasized in the coverage. The coverage missed this crucial point: Anthropic's Project Fetch shows that while models are proficient at writing robot code, they are not yet adept at executing it.

This gap highlights the necessity of a world model like Oasis, which acts as a simulator without the need for physical physics, enabling more efficient training and potentially accelerating robotics.