Winner of robots

GPT-6 Astra and Claude Fable have dismantled the premise that robotics is a fresh start. General-purpose language models are acquiring capabilities that robotics specialists expected to develop themselves. In NVIDIA’s Astranomy study, Astra outperformed the two specialist robot models. The races to build language models and robot brains have merged.

The advantage passes to OpenAI and Anthropic. Their existing researchers can extend their work into robotics, while recursive self-improvement will make their models increasingly useful in developing successors. Established revenues give investors a demonstrated business against which to finance that research. Robotics startups must fund a comparable effort before their principal market has arrived. Google enters with the same institutional advantage.

Final commercialisation of robots require tuning language models, which strengthens the model owners’ position because tuning these models for robot control requires access to the weights. The effort required to build another Astra or Fable is unimaginable for a startup.

Skild AI and Physical Intelligence must now match the frontier laboratories merely to defend their original market. Young startups can pivot to specific applications. Companies financed to own the universal robot brain cannot do so without abandoning the investment thesis behind their valuations. For those that remain in the race, the endgame is acquisition or failure.

NVIDIA, meanwhile, will keep selling the shovels. Its investment in robotics software turns the industry’s growth into demand for its chips. Jensen Huang introduced Isaac in 2017, long before the current rush. NVIDIA’s next boom will owe much to the foresight of the man whose resignation its board has just accepted.