Kalanick just hired the guy who made AI fast enough to run on your face to build foundation models for robots that cook, mine, and move things.
The Summary
- Vikas Chandra, Meta's AI lead for smart glasses, joins Travis Kalanick's Atoms as VP of AI to build "foundation models for the physical world"
- Atoms raised $1.7 billion from a16z this summer and absorbed Kalanick's CloudKitchens as it staffs up for robots in food, mining, and transportation
- The hire signals Atoms is going after the hard problem: getting AI to work in real time on physical robots, not just in the cloud
- Chandra's expertise: making models small and fast enough to run on wearables, exactly what you need when compute sits on a robot arm
The Signal
Atoms is assembling the team to do what most AI companies are avoiding. Foundation models for robots. Not chatbots. Not code generators. Vikas Chandra spent eight years at Meta making AI models run on smart glasses, which means he knows how to compress neural networks without breaking them. That skill translates directly to robotics, where you can't phone home to a datacenter every time a robot needs to decide whether to flip a burger or avoid a collision.
The job Chandra did at Meta is the job Atoms needs done at scale. Smart glasses require real-time inference on battery-powered chips with thermal constraints. Robots in kitchens, mines, and warehouses face the same constraints, plus higher stakes. A chatbot that hallucinates is embarrassing. A mining robot that hallucinates coordinates kills people.
"The physical world is a harder version of the same problem."
Atoms raised $1.7 billion this summer, one of the largest robotics rounds ever. Andreessen Horowitz led it. That money is hiring firepower. Former Uber CFO Gautam Gupta joined last month. Atoms acquired autonomous mining startup Pronto earlier this year. Now Chandra. The pattern is clear: Kalanick is buying talent that has shipped hard tech at scale, not researchers who publish papers.
CloudKitchens, Kalanick's ghost kitchen venture, is now a subsidiary of Atoms. That tells you something about the strategy. CloudKitchens was about digitizing food production and delivery infrastructure. Atoms is the bigger play: digitizing physical operations across industries. Food was the testbed. Mining, transportation, and manufacturing are the real targets.
Key elements of Atoms' strategy:
- Foundation models purpose-built for physical-world tasks, not repurposed LLMs
- Edge inference: models run on the robot, not in a cloud datacenter
- Vertical integration from model training to robot deployment across multiple industries
Foundation models for the physical world is the right framing. OpenAI, Anthropic, and Google are building foundation models for language and reasoning. Figure, Tesla, and now Atoms are building foundation models for manipulation, navigation, and coordination. The latter is harder. Language has discrete tokens and nearly infinite training data scraped from the internet. The physical world has continuous state spaces, safety constraints, and far less data. You can't just scale your way to a robot that works.
The Implication
Watch where Atoms deploys first. If Chandra's hire means anything, it means they are prioritizing speed and reliability over flexibility. They will go after environments where the same task repeats at high volume: commercial kitchens, mining operations, warehouse logistics. Places where you can amortize model development costs across thousands of identical robots doing identical work.
The bigger implication is about the agent stack. Software agents need APIs and databases. Physical agents need bodies, sensors, and models that run in milliseconds. Atoms is betting that the companies who win Web4 will be the ones who control the full stack from model to motor. If they are right, the next decade belongs to whoever can make intelligence cheap and fast enough to run on metal.