An SEC filing shows 32 investors joined the $208.2M offering; Axios reports 8VC led at a valuation above $2B.
By Ryan Merket · Published
Primary source: Axios
Why it matters #
Generalist has raised nearly $600M in less than three months, showing how quickly investors are funding the data and compute race behind general-purpose robot models.
Generalist, the robot-model developer founded by Pete Florence (@peteflorence), Andrew Barry, and Andy Zeng, has sold $198.2 million of a new equity offering less than three months after closing a $400 million financing.
A Form D filed with the Securities and Exchange Commission on August 24th lists a total offering of $208.2 million, with about $10 million remaining to be sold. Generalist reported 32 investors and identified August 7th as the date of the first sale. Florence signed the filing as CEO.
Axios reported that 8VC led the financing alongside several existing investors. Axios also reported that Generalist's valuation exceeded the $2 billion mark set by its June round, without giving a more precise figure.
The pace is the point. Generalist announced $400 million on June 4th, bringing its total capital raised at the time to more than $500 million. Adding the amount sold in the latest filing puts Generalist's disclosed total above $700 million, with nearly $600 million arriving in less than three months.
An existing backer returns as lead
8VC participated in the June financing before taking the lead in the latest offering, according to Axios. That earlier round was led by Radical Ventures and included Union Square Ventures, Hanabi Capital and Norwest. Nvidia, Boldstart Ventures, Spark Capital, Bezos Expeditions and NFDG returned as existing investors. Angel investors included Fei-Fei Li and Naval Ravikant.
Generalist's transaction counsel, Fenwick, said the June financing valued Generalist at $2 billion. The latest filing does not identify individual investors, leaving Axios' report as the basis for 8VC's role and the higher valuation.
The financing puts more capital behind a founding team assembled from two of the most consequential robotics programs of the past decade. Florence was a senior scientist on Google DeepMind's robotics team and worked on PaLM-E and RT-2, models designed to connect language and visual inputs with physical actions. Zeng was a research scientist and technical lead at Google DeepMind, where his work included large-scale robot foundation models and robot data collection. Barry previously worked on Spot's robotic arm at Boston Dynamics before moving into machine learning research at the Broad Institute.
Generalist was incorporated in Delaware in 2024, according to the SEC filing. Generalist operates from the San Francisco Bay Area and Boston, building embodied foundation models focused initially on dexterity. Its pitch is an intelligence layer that can control different forms of robotic hardware instead of tying the model to a single humanoid or industrial machine.
GEN-1.5 gives investors a new technical milestone
The financing began before Generalist publicly introduced GEN-1.5 on August 19th. Generalist says the model can learn short, dexterous tasks from a single demonstration lasting between three and 12 seconds, without fine-tuning. In tests across 10 tasks, Generalist reported an average success rate of 59%, rising to 83% after 10 gradient steps using five minutes of task data.
Those figures remain company-reported, and Generalist describes the tasks as simple and short-horizon. A 59% completion rate is also far below the reliability required for most production work. The result matters because it reduces the amount of task-specific training needed before a robot can attempt a new action, one of the main costs slowing deployments outside controlled demonstrations.
Wired observed Generalist's robot arms stacking cups, handling purses and improvising with unfamiliar tools at its Cambridge, Massachusetts, office. Wired also reported seeing hundreds of Generalist's handheld grippers being prepared for data collectors. Generalist uses the devices to capture human physical actions, producing training data that can be applied across different robot designs.
That operation explains part of the appetite for capital. Generalist said in June that it planned to expand its physical data engine, compute infrastructure, model training and work with prospective industrial users. Each of those functions carries costs that software-only AI developers can avoid, even though Generalist does not manufacture the robots its models control.
The new offering gives Florence, Barry and Zeng substantially more room to train larger models and push them into commercial settings. Their immediate technical gap is visible in Generalist's own numbers: GEN-1.5 can attempt unfamiliar work after seconds of instruction, but reliable physical labor still requires a large jump from 59% success to production-grade performance.