Generalist AI Raises $400 Million at a $2 Billion Valuation to Build Foundation Models for Robots, Backed by Nvidia and Bezos
The San Mateo startup, founded in 2024 by ex-DeepMind and Boston Dynamics researchers, has now raised more than half a billion dollars to scale its robot foundation models.
Overview
Generalist AI has raised $400 million in new funding at a valuation of approximately $2 billion, according to SiliconANGLE and Robotics & Automation News. The round was led by Radical Ventures and brings the company’s total funding to more than half a billion dollars, The Robot Report reported. Generalist AI, which builds general-purpose artificial intelligence for a range of robot form factors, says it will use the money to advance what it calls physical artificial general intelligence.
What We Know
The company was founded in 2024 and is based in San Mateo, California, according to The Robot Report. Its chief executive is Pete Florence, a former senior scientist at DeepMind who helped create the RT-2 and PaLM-E robotics models, per SiliconANGLE. The same outlet reports that chief technology officer Andrew Barry was formerly a roboticist at Boston Dynamics, and that Andy Zeng serves as chief scientist.
Radical Ventures led the round, The Robot Report reported, with new investors 8VC, Union Square Ventures, Hanabi Capital and Norwest. Existing backers that participated again included NVIDIA’s NVentures, Boldstart Ventures, Spark Capital, Bezos Expeditions and NFDG. The same report lists Eric Yuan, Bin Lin, Fei-Fei Li and Naval Ravikant among new angel investors. Generalist AI plans to use the round to accelerate its mission to build physical artificial general intelligence and make it useful to everyone, The Robot Report reported.
The raise follows the April release of GEN-1, the company’s latest robot foundation model. According to Robotics & Automation News, GEN-1 demonstrated 99% reliability on dexterous manipulation tasks and executed them up to three times faster than previous systems. In its own technical blog post, Generalist AI reports that GEN-1 reached a 99% average success rate on a set of demonstrated tasks where its predecessor, GEN-0, averaged 64% and a model trained from scratch averaged 19%. The company says GEN-1 can match GEN-0’s performance using 10 times less task-specific data, and that adapting the model to a new task requires only approximately one hour of robot data.
The model is trained on roughly half a million hours of real-world data collected using low-cost wearable devices worn by humans performing everyday activities, rather than data gathered solely on robots, according to the company’s blog post. The same post describes GEN-1 running across a series of repetitive demonstrations, including folding T-shirts 86 times in a row, packing blocks more than 1,800 times in a row, and packing phones more than 100 times in a row.
Generalist AI released GEN-0 in November, a model that Robotics & Automation News reports demonstrated scaling laws in robotics.
What We Don’t Know
The published reporting does not detail the specific equity terms of the round, how the $400 million will be split between compute, hiring and data collection, or a timeline for any commercial deployment of GEN-1. The figures for GEN-1’s success rate, dataset size and per-task data efficiency come from the company’s own materials and benchmarks rather than independent third-party evaluation.
Analysis
The round places Generalist AI among a cluster of well-funded startups racing to build foundation models for the physical world, a field where the central bottleneck is collecting enough diverse real-world interaction data to make robots reliable outside controlled demonstrations. Generalist AI’s pitch, reflected in its decision to gather training data through wearable devices on humans rather than expensive robot teleoperation, is that this data problem can be attacked at scale. With Nvidia’s venture arm and Bezos Expeditions among the backers, the financing underscores continued investor appetite for the bet that general-purpose robot intelligence, not single-purpose machines, will define the next phase of automation.