Skip to content

OpenAI is betting robotics is now a job for its general models. The biggest robotics checks are going to companies that own what those models cannot download.

0:00 / 0:00
OpenAI Merged Its Robot Team. Investors Are Paying for the Factory Floor.
OpenAI has stopped treating robotics as a field of its own. Fortune reported on November 13 that it has merged its standalone robotics team into its main research and applied groups, betting that its general models can run robots. Investors are reading the same field the other way. In the past three weeks, the biggest closed robotics rounds went to companies whose robots work on factory floors, in warehouses, or at building sites.

OpenAI Folded Robotics Into Its Core Teams, and Says Its Models Are Enough

Under the reorganization, OpenAI's robotics research moved into the main research team run by Chief Research Officer Mark Chen. The applied robotics and hardware teams moved to Applied, which reports to President Greg Brockman. OpenAI says the change is meant to speed robotics up, not drop it. OpenAI has said it is building its own humanoid robot, and CEO Sam Altman is targeting a first public demo in 2027.
OpenAI's robotics team since 2020, and where it went in November 2026. Graphic: EO.
This is the second time OpenAI has changed course on robots. It shut down its first robotics team in 2020, after building a robotic hand that solved a Rubik's Cube. Co-founder Wojciech Zaremba said in 2021 that progress came whenever the team had access to data. OpenAI formed a new standalone team in 2024.
This time the case rests on performance. In an Nvidia study, an agent built on OpenAI's GPT-6 Astra completed 74.8% of tasks on a simulated manipulation benchmark. The best robot-specific models on the same leaderboard completed 51.0%. The general model beat the specialists, in simulation.
An agent built on OpenAI's GPT-6 Astra against the best robot-specific model, in Nvidia's simulated test. Graphic: EO.
Guanya Shi
Assistant Professor at Carnegie Mellon University's Robotics Institute
Guanya Shi, an assistant professor at Carnegie Mellon University's Robotics Institute, told Fortune he reads the merger as a sign that robotics is now an application of general AI, much as frontier models now run coding tools. He still expects small, specialized models to be needed on robots that cannot rely on a constant internet connection. The companies raising money for robots have a different idea of what is scarce.

Investors Are Paying for What a General Model Cannot Download

In three weeks, Path Robotics, which makes AI-guided welding robots, raised $300 million; Agile Robots, a German robotics company, raised $500 million; and Bedrock Robotics, which retrofits construction equipment to drive itself, raised $600 million.
Tutor Intelligence, a startup that builds warehouse robots, raised a $75 million Series B led by Andreessen Horowitz on November 2. It puts its robots to work at customer sites. When one gets stuck, a human "tutor" steps in, and the correction becomes training data. A model trained in a lab has never seen that warehouse.
A Tutor Intelligence robot picks items from warehouse shelving.Source: Tutor Intelligence.
MIT robotics professor Russ Tedrake told Fortune that robotics companies keep an edge because their robots collect data from real deployments, such as factory floors. Better foundation models make that data more valuable, he says, because robots can learn from it faster. Josh Gruenstein, co-founder and CEO of Tutor Intelligence, argues that hardware, intelligence and customers are one design problem, and that a robot company has to own every layer.
One company in this boom sells only the intelligence, and its valuation in talks is higher than any of theirs.

A $20 Billion Valuation Sits on the Layer OpenAI Calls General

Physical Intelligence, a San Francisco company that builds AI models for robots, is in talks to raise at least $1 billion at a valuation of about $20 billion.
Partner robots running Physical Intelligence's models; center, Chef Robotics assembling a burger.Source: Physical Intelligence.
If OpenAI is right, a company that sells only a robot model has the least to fall back on. It would not own the lab's model, and it would not own the floor's data. Two things point that way. Specialist models scored 51.0% in Nvidia's test, and Shi sees robotics as an application of general AI.
Lightspeed partner Guru Chahal wrote on November 14 that robotics may scale through orchestrated specialists, with a reasoning model planning and a specialized policy moving, before any single universal robot brain arrives. A simulated score is also not a robot on a factory floor. Pete Florence, co-founder and CEO of Generalist, told Fortune that models can start from different places.
Pete Florence
Co-founder and CEO of Generalist
What will settle it is OpenAI's own robot. The company is targeting a first public demo in 2027. If a general model runs that robot as well as it ran the simulation, the robot-model companies will be selling something the labs already have. If it does not, robot-specific models will still have a job, and so will the companies that hold the floor data.

Made it to the end? Get the next one weekly.

We're building this with you.

Got 30 seconds? Share your thoughts and make EO better.

Send feedback

Next read

Lightyear Residency: 30 Days. Zero Distractions. No Excuses.

Jun 17, 2026
Inside Hacker HousesMore like this

A look inside the first ever Lightyear batch demoday

Explore more

Editor's Pick

Everyone in Robotics Is Fighting the Same Demo-to-Real Gap

Everyone in Robotics Is Fighting the Same Demo-to-Real Gap

Inside the launch of Industrial Next's Real2Sim2Real Arena

OCT 09, 2026
9 MIN READ

Editor's Pick

Nominal Raises $240 Million at $3 Billion to Tell Engineers Whether Their Machines Work

Nominal Raises $240 Million at $3 Billion to Tell Engineers Whether Their Machines Work

The defense boom drove its growth. Its next market is every company building machines that run on AI.

NOV 10, 2026
6 MIN READ

Try something new

 The Founder Mistake I Saw From Google to Dropbox

The Founder Mistake I Saw From Google to Dropbox

An interview with Dennis Woodside, CEO of Freshworks

OCT 04, 2026
16 MIN READ