Generalist AI raises $400m at $2bn for embodied foundation models, as its GEN-1 hits 99% task success

Summary

Generalist AI, the San Francisco embodied-AI startup founded by former DeepMind senior scientist Pete Florence, raised $400m on 4 June 2026 at a $2bn post-money valuation. Radical Ventures led; 8VC, Union Square Ventures, Hanabi Capital and Norwest joined as new investors alongside returning backers NVIDIA's NVentures, Boldstart Ventures, Spark Capital, Bezos Expeditions and NFDG. Florence previously co-created RT-2 and PaLM-E, the two foundational papers that established the academic programme in embodied AI. Generalist's first public model, GEN-1, was released in April 2026 and improved task success from 64% to 99% on standard dexterous-manipulation benchmarks while completing tasks approximately 3x faster than prior approaches. The company builds general-purpose foundation models for physical robots, aiming to provide the same kind of pre-trained backbone for hardware that GPT-class models provide for software applications. The $400m raise is the largest Series A in robotics AI this year, though it is a fraction of Skild AI's $1.4bn at $14bn, reflecting a different bet: Generalist is benchmarking on model quality rather than manufacturing partnerships. Nvidia's NVentures arm has now backed most of the leading embodied-AI companies, creating a structural alignment between GPU supply and equity stakes in the models trained on that supply.

The split

US robotics and enterprise AI press treat GEN-1's performance numbers (99% task success, 3x speed improvement) as the central claim, while noting they come from Generalist's own benchmarks rather than independent reproducibility. Silicon Republic's European angle focuses on the Nvidia-Fei-Fei Li coalition as a concentration risk: if a single compute supplier holds equity in the top models, it shapes which approaches survive on technical merit versus which survive because they get preferential GPU access. The Bezos Expeditions participation is noted without comment, consistent with his broad embodied-AI thesis but not exclusive to Generalist.

By the numbers

• $400m, Series A round size.
• $2bn, post-money valuation.
• 99%, GEN-1 task-success rate on dexterous-manipulation benchmark (vs. 64% for prior models).
• 3x, GEN-1 speed improvement vs. prior approaches.
• April 2026, GEN-1 public model release date.
• Founded: 2024.
• Founder: Pete Florence, former DeepMind senior scientist (RT-2, PaLM-E).

Why it matters

Generalist AI is the clearest test of whether the general-purpose embodied foundation-model thesis, one model backbone for all robots rather than a bespoke system per hardware platform, is technically viable. GEN-1's 99% success rate and 3x speed improvement are the strongest public benchmark numbers in the sector. If the model generalises across robot form factors beyond those tested in Generalist's own lab, it validates the analogy to software foundation models. Nvidia's continued co-investment in both the compute and the model equity suggests the company believes the embodied-AI stack will become as GPU-intensive as LLM inference, and is positioning to capture both sides.

What to watch

• Independent replication of GEN-1's benchmark numbers by academic labs and competing robot vendors.
• Whether Generalist licenses GEN-1 to hardware partners or tries to own the full stack.
• How Skild AI, the competing general-purpose robot brain backed by SoftBank at $14bn, responds on capability benchmarks.
• Whether the Nvidia NVentures coalition creates preferential GPU access that advantages Generalist over compute-independent competitors.

---