Skip to content

every ai lab is building physical ai – except openai and anthropic

pulse Illustration of a man clutching a glowing purple disc while facing a robotic arm in a graffiti-lined industrial corridor

today alibaba dropped the qwen-robot suite – three foundation models, a full stack for embodied intelligence:

• qwen-robotnav – mobility brain. one model, 5 navigation tasks, agentic tool interface
• qwen-robotmanip – manipulation brain. unified state-action space across robots, trained on 38,100+ hours of open data
• qwen-robotworld – world model. 20+ embodiments, natural-language actions, physically grounded futures

each independently useful. together, the toolkit for agents that don't just see the world but act in it

and qwen isn't alone. nvidia, google, xiaomi, and tencent have all shipped physical ai models this year (see below). musk's empire is in too, via tesla

us. china. open-source. closed. everyone's in. except openai and anthropic

why?

- openai had a robotics team.
shut it down in 2021. spent the gap years investing in figure, 1x, and physical intelligence, then supplying language brains via partnerships. that broke too – figure fired them in feb 2025 over technical disagreements. they finally restarted internal robotics on may 31, 2026. rebuilding from scratch while everyone else has a 3-year head start. no released model yet

- anthropic never started. their bet is different: claude as a general reasoning layer that controls robots through code and tool use, not a trained-on-robot-trajectories vla. project fetch (nov 2025) showed claude can program a robot dog to autonomously fetch a ball – but that's uplift research, not a foundation model. their contribution to the physical stack is mcp, the protocol others use to wire claude into ros

the structural reason both are behind: vision-language-action models (the brains running robots) need robot trajectory data. millions of hours of it. the labs winning here either own factories, ship robots, or sit on years of egocentric video. openai and anthropic have text and images. wrong data flywheel

physical ai is becoming the next axis of ai competition. and on this axis, the language model leaders are spectators

Table comparing physical AI models from Nvidia, Google, Tesla, Alibaba, Xiaomi, and Tencent with release dates in 2026

Stay in the loop

Get the latest AI news delivered to your inbox weekly

Thanks for subscribing!