Dream Labs Raises $90M Seed to Sell Robot Brains by the Hour

A $90 million seed for a robot brain sold by the hour

Dream Labs is a San Francisco startup founded in 2026 by former Nvidia robotics researchers. It came out of stealth on October 19 with $90 million in seed funding and a public product that customers can use immediately.23 The product is Dream API. Developers send it a robot's camera feed and a plain-English instruction, and the company's cloud-hosted model, Dream-1, sends back motor commands. Dream Labs says this works even on hardware the model has never seen.27 The company calls it the first publicly available API for a robot foundation model.34

Investors in the round include Kindred Ventures, A.Capital, Unusual Ventures, SV Angel, Wing VC, Scribble VC and angel investor Adrian Aoun.19 The company has not disclosed a valuation.8 It says most of the money will go to compute, both to train Dream-1 on more tasks and robot types and to handle more API traffic.25

The money is only half the story. What matters more is the business model. Dream Labs does not build robots. It is betting that robot intelligence can be split off from robot hardware and sold as a metered cloud service, much as large language models are.

What Dream-1 actually does today

Dream-1 learns from video of people doing everyday tasks. It predicts what should happen next in a scene and works out the robot movements needed to get there.13 Forbes put it this way: Dream-1 treats robot control as a video-prediction problem and is trained on egocentric footage, meaning video shot from the viewpoint of the person doing the task.3 This approach comes directly from the founders' research. CEO Joel Jang led robotics research at Nvidia's GEAR Lab and worked on Project GR00T and related projects including LAPA, DreamGen, DreamDojo and DreamZero. The company credits him with coining the term "world action model" for this kind of system.7 Wing VC's investment note also presents Dream-1 as a continuation of the world action model research, which predicts video and robot actions together.2

The product's real scope is narrower than the "any robot" headline. Right now Dream-1 handles tabletop tasks such as picking up and placing objects and folding laundry.413 Forbes reports that the flagship model can operate two-armed robots with parallel grippers out of the box.3 Support for wheeled mobile robots and dexterous hands is still planned, not shipped.4 One outlet's fact-check concluded that the confirmed claim is cross-embodiment transfer for robots with grippers, and that this should count for more than the marketing language until real deployments produce independent data.1 That reading seems right to us.

The launch also included a fine-tuning service and a browser-based "Robot Playground," where users can send instructions to physical robots in the company's San Francisco office.45 Jang presented this openness as a deliberate choice: the company is giving people a working model they can test themselves, mistakes included, instead of edited demo videos.4 The company's own technical post adds a few details. It reports tests on 27 robot bodies in simulation that the model had never seen, with 300 rollouts per robot, plus real-world runs on arms including the 18-year-old Willow Garage PR2.8 Dream Labs has not released Dream-1's parameter count, inference latency, a full list of supported robots or reliability rates in real deployments.2

Pricing, and where the reports disagree

The coverage largely agrees on the flagship price: $3 per robot-hour for Dream-1.38 It splits on the cheaper tier. Forbes built its headline around a 75-cent option for a less capable model.3 Other outlets that cite the company's Business Wire release and launch post list Dream-1-flash at $2 an hour.81 One report went further and said Dream Labs had not published pricing in its official announcement at all.2 The most likely explanation is that the published flash price and figures from interviews or early pricing do not match. Buyers should check the live price list before taking either headline figure at face value.

The per-action math shows why the price matters. By the company's own figures, each API call returns 1.6 seconds of actions. An hour is therefore 2,250 calls, which works out to about $0.0013 per action on the flagship model.8 Dream Labs is targeting cheap hardware. Its more than 20 hardware partners make robots that generally cost $2,000 to $10,000.108 At $3 an hour, the model costs about the same as the electricity and wear on those machines, not the price of a specialist integrator spending months on setup. That is the core of Jang's "minutes, not months" pitch, which he illustrated with a rental-car comparison: a driver doesn't relearn to drive in each new car.4 The company has not named most of those partners.2

The reports also disagree on who led the round. Several describe B Capital as the sole lead with Kindred Ventures participating.12 Others describe B Capital and Kindred as co-leads.4613 This difference doesn't change the substance of the deal, but readers will see both versions.

Why the humanoid crowd should care

Dream Labs is not a humanoid company, and that is part of why it matters for the humanoid market. Vertically integrated players such as Figure, Tesla's Optimus program and Boston Dynamics tune their control software to their own hardware.2 Dream Labs is pitching a horizontal layer instead: a single reusable control model that robot makers could license rather than spending years building their own.1

Dexterous hands are where the two worlds meet. Jang told Forbes that dexterous hands already work with the model, but that a pair currently costs about $100,000. He expects capable hands to reach $5,000 to $10,000 by next year, and Dream Labs plans to expand support as they arrive.3 In effect, the company's path toward humanoid-style manipulation depends on falling hardware costs, not on a research breakthrough. If he is right about hand prices, a model-and-API vendor could serve humanoid and semi-humanoid platforms from smaller makers that cannot afford their own foundation-model teams.

A crowded, expensive field

The $90 million figure stands out for a company that is only months old, but it is not unprecedented. Genesis AI came out of stealth with a $105 million seed in July 2025. Dream Labs' better-funded rivals are on another scale. Physical Intelligence raised $600 million at a $5.6 billion valuation and has reportedly held acquisition talks with Anthropic. Skild AI raised $1.4 billion earlier this year at a valuation reported to be near $14 billion, also pitching a general-purpose robot brain.7

In that context, $90 million is a modest compute budget. One analysis noted that it is a fraction of what the largest AI labs spend on compute in a single quarter. Dream Labs therefore has to make Dream API the default choice before bigger players release similar horizontal products.2 The same analysis pointed to wider uncertainty: chip stocks have swung sharply in the second half of 2026 over whether AI compute demand justifies current valuations. A seed round spent mostly on compute is a bet that this climate will keep rewarding infrastructure-heavy AI.2

Our read

We see the round as a bet on the team that is also credible as a product launch. It is not proof that robot foundation models have solved the problem of working across different machines. Investors are clearly paying for pedigree before revenue, a pattern that runs through robotics funding in 2026.6 What sets Dream Labs apart is that it shipped something people can use on day one, with a price per hour, instead of a research preview.2 That turns the company's claims into something outside developers can test, which is more than most seed-stage robotics companies offer.

Unite.AI identified the key metric: how much custom data and engineering developers still need after connecting a new robot.5 The company's own product description admits that some cases need substantially more data. Three things will show whether the claim holds: how many API hours customers actually use, whether the 20-plus hardware partners stay, and how much fine-tuning each new robot body requires. If Dream-1 stays limited to tabletop gripper tasks, the round will look early. If it expands convincingly to mobile bases and affordable dexterous hands, Dream Labs will have shown that a robot's brain can be sold separately from its body.