Summary: Rhoda AI emerged from stealth on March 10, 2026, with a $450 million Series A reported by Reuters and Bloomberg at a $1.7 billion valuation and with FutureVision, a foundation model for robot intelligence. FutureVision uses Direct Video Action to pretrain on hundreds of millions of public videos, post-train on small robot datasets, and close an observe-predict-act loop every few hundred milliseconds. Its public proof point is a manufacturing evaluation, while broader licensing to other platforms remains a future plan rather than a released general product.
After eighteen months in stealth, Rhoda AI — a Palo Alto company — stepped into public view on March 10, 2026, carrying two pieces of news at the same time: a $450 million Series A (Reuters and Bloomberg report the valuation at $1.7 billion) and FutureVision, a foundation model for robot intelligence. The institutions writing checks are Capricorn Investment Group, Khosla Ventures, Leitmotif, Matter Venture Partners, Mayfield, Premji Invest, Prelude Ventures, Temasek, and Xora, with John Doerr participating as an individual investor.
The date deserves a second look. The money arrived one day after Nscale claimed the record for the largest funding round in European history (see Nscale’s $2B Series C), yet the two wagers point in opposite directions. One builds compute infrastructure; the other puts a foundation model inside robots — the intelligence layer. Capital markets, it seems, judge both roads worth backing at scale.
What Makes a $450M Series A Possible
The roster carries part of the argument. Co-founder and CEO Jagdeep Singh has started deep-tech companies before; Chief Science Officer Eric Ryan Chan researched computer vision and generative modeling at Stanford and previously served as a generative model architect at WorldLabs; co-founder Gordon Wetzstein leads Stanford’s Computational Imaging Lab.
The task itself is what stretches the round to $450 million: instead of control software for one machine, the team aims at a robot foundation model that transfers across embodiments. Premji Invest’s Patnam put the logic in plain words: a model of this kind can “kick-start a powerful data flywheel, creating a compounding advantage in capturing the long tail of real-world edge cases.”
FutureVision: From Watching Videos to Acting
Underneath FutureVision sits a proprietary architecture Rhoda names Direct Video Action (DVA). Training passes through two stages. The first pre-trains on internet-scale video — hundreds of millions of public clips — so the model picks up motion and physics priors; the second post-trains on much smaller robot datasets, teaching it to map video predictions onto actions and to tune behaviors specific to each embodiment.
At runtime the system closes its own loop: observe, predict upcoming states as video, commit an action, observe again — one cycle every few hundred milliseconds. The design descends straight from the world-model school: understand how the scene will evolve first, then choose the move. Standing up a new task takes roughly ten hours of teleoperation data, an order-of-magnitude drop from the per-task custom engineering traditional robotics expects.
Onto the Factory Floor: Manufacturing and Logistics First
The go-to-market pick is pragmatic: manufacturing and logistics. As one example, the press release points to a high-volume manufacturing evaluation where a component-processing task kept every cycle under two minutes without any human involvement and cleared the customer’s KPIs. From the buyer’s seat, Leitmotif’s Wiese offered the caveat: “The real challenge isn’t solving it once, it’s delivering consistent, reliable output under real-world production conditions.”
Commercially, Rhoda is deploying together with industrial partners at this stage, and over the longer run plans to license FutureVision to other hardware and software platforms — the robot brain is the product, not the robot itself. The parallel with how the LLM industry turned models into APIs is hard to overlook.
Follow the Flywheel: What the Money Is Actually Buying
Put three facts next to each other and the real object of the investment shows itself: the pre-training corpus is public internet video, which costs almost nothing to collect; post-training gets by with teleoperation data on the scale of ten hours; and every factory where the model goes to work keeps generating proprietary robot data that outsiders cannot reach. Every extra turn of the flywheel pushes down the entry cost of the task after that, while competitors’ data shortfall grows.
This playbook is shared across the 2026 robotics race: pre-training on video provides the “common sense,” and the data streaming back from deployments provides the “expertise.” Whoever scales deployment first converts the physical world’s long tail into a moat.
Three Takeaways for Robotics Developers
For practitioners, three points deserve attention. One: the robot foundation model is turning into a component you can procure and license — teams focused on a specific embodiment or vertical will compete over “whose intelligence layer,” a question that displaces “build it or not.” Two: video pre-training paired with a small amount of teleoperation has passed market validation more than once now, so a new team’s differentiation should live in its data and its domain, not in rebuilding pre-training from zero. Three: closed-loop control is acutely latency-sensitive (cycles of a few hundred milliseconds), which makes edge deployment and inference cost hard engineering constraints — before signing with a vendor, weigh its edge strategy as part of the selection.