From Punchline to Platform: Hard Things Round Two and the Real History of Robotics
San Francisco, CA: Figure AI is now valued at $39 billion. Humanoid robotics funding grew roughly 15x in three years, to $3.7 billion in 2025. Tracxn counts 121 humanoid companies globally, sitting on $6.89 billion of venture capital.
Six weeks ago, we wrote about the inaugural Hard Things — a Palo Alto living room with Veena Radhakrishna of Cartesian Kinetics, in the same week Jensen Huang disclosed a $1 trillion Nvidia order book at GTC. Veena gave us the operator’s framework: design for retrofit, choose the intelligence layer over the hardware layer, stay flexible long enough for the market to clarify.
Last week in San Francisco, Round Two gave us the other half. Veena gave us the “how”. Andra Keay of Silicon Valley Robotics gave us the “why now.”. She told us why a field unfundable for fifteen years has become the place every generalist VC is anxious about showing up late to.
Before It Was Cool
Andrea has run Silicon Valley Robotics, the world’s largest cluster of robotics and AI companies, since 2011. She also founded Women in Robotics and writes “Robotics and Startups” on Substack. If a robotics company in this Valley matters, Andrea probably saw it first.
Andrea arrived in Silicon Valley at the end of 2010 looking for a 20-year technology bet. She picked robotics. Then she had to find robotics. At her first SVR gathering, there were thirty companies. She mapped the rest of the local ecosystem and got to sixty. There was no venture industry for robotics in 2011. There were just a few hardcore investors drifting in from semis, a handful of founders trying to spin tech out of Willow Garage, and that was it.
For the next several years she crashed every venture event she could. She got laughed at. “Hardware is too hard” was the standing line. By 2015, she was tracking about $1 billion of global venture investment in robotics, and in in hindsight, probably double that. One of many mic drops of the evening: when Andrea went back and looked at the early investors who told her robotics was unfundable, the most common credential in their bios was a fine arts degree.
Today the bios read differently. Physics PhDs. Hardware operating experience. NVIDIA’s NVentures alone has roughly $2.1 billion of cumulative humanoid investment. Strategic capital from CATL, JD.com, Alibaba, and Amazon is outpacing traditional VCs. The category that used to be a punchline is now the belle of the ball.
What Actually Changed
We asked Andrea what the biggest unlock has been. Foundation models? Hardware costs? Capital? None of those. The biggest unlock has been time and wisdom. Founders learned a scientific approach, defining what they were trying to prove, what milestones would prove it, what would force a pivot. Investors learned to underwrite hardware risk. The ecosystem matured. The technology, in her telling, has always been moving. What was missing was the ability to build companies on top of it that could actually ship.
This is the historical proof point underneath Veena’s framework. The market always clarifies later than the models predict and faster than the incumbents are ready for. Andrea walked us through fifteen years of exactly that pattern: hype cycles running a year ahead of evidence, disillusionment burning a year after, real industries quietly compounding underneath both. The founders who survived built organizations flexible enough to keep learning while the market took its time.
It also vindicates the larger Round One thesis: “if AI is eating software, hardware and robots are at the top of the food chain.” The phrase is Andrea’s. Software gets easier to churn out of every quarter. Robotics gets “stickier” once it’s in, not because the metal is hard to rip out, but because the integration into operations and the digitization of three-dimensional process data creates a customer with nowhere to churn to. Every deployment is different. You cannot drop in a replacement.
The Bubble Isn’t What You Think
So inflection or bubble? Andrea’s answer: there’s a bubble, but it isn’t in robotics. It’s in capital.
Unitree, the world’s most prolific humanoid manufacturer, shipped about 5,500 units last year, and that may quietly include quadrupeds. Maybe ten companies sit at that production scale. There are two to three million industrial robots in the world today, after sixty years of industry development. The forecasts imagining ten billion humanoids by 2040 are, in Andrea’s word, “no way.” The realistic ceiling looks more like 500,000 units by 2035, and the market still hasn’t figured out what most of them will actually do for a living.
This is why the Figure-BMW pilot mattered, even at 50% uptime. Not a verdict on the technology, just a useful piece of discovery. Every humanoid deployment for the next five years should be read as discovery, not validation. Same lesson Veena pushed in Round One: let customer environments teach you what your roadmap can’t predict. Andrea gave us fifteen years of receipts.
The line that came up repeatedly: pick high-value use cases. Robotics fails chasing low-margin automation. It wins chasing labor that’s scarce, expensive, regulated, or impossible to attract. Think pharmaceuticals with chain-of-custody, blood-sample transport, cold-chain medication, industrial work nobody under thirty wants. The opportunity is high-value cargo where someone is “already” paying a fortune for a credentialed human to do the job badly.
And — Andrea channeling Tolstoy — staked the view that happy startups are all alike. Stop asking the founders who succeeded she says, and ask the ones who failed at exactly what you’re trying to do. She called it “institutional graveyard” wisdom. We are stealing the phrase.
The Hand Problem
One of the most arresting moments came from the audience. Tapan, Chief Scientist at Adverb, made a point most generalist humanoid investors do not yet appreciate. The current race is overwhelmingly focused on vision and language models trained on internet-scale data. But you cannot light a match with anesthetized fingers, no matter how many tokens you’ve seen. Without tactile feedback (think pressure, force, grip), robots cannot do the things humans do with their hands. And humans do most useful physical work with their hands.
Andrea agreed. The current AI advances still operate largely on abstractions of physical-world action. The missing layer is embodiment feedback. The most exciting near-term development she sees isn’t a humanoid company at all. It’s the wave of new sensor and tactile-grasp companies. “Hand as a service,” she joked. We aren’t sure she was joking.
Your First Household Robot Is Your House
Asked what non-obvious scenario the world is missing five to ten years out, Andrea said: your first household robot isn’t going to be Rosie. It’s more likely going to be your house.
The future, in her telling, isn’t a parade of humanoids. It’s intelligence quietly embedded in furniture, fixtures, vehicles, infrastructure, hospitals, fields. The average car is already a robot. Most of the robotics we live with in 2035 will be invisible. The flip side: judgment, presence, taste, relationships (the things most valuable about being human) become more economically valuable, not less. Robot chefs were supposed to obliterate restaurant work. People still pay more for food cooked by a person.
The Through-Line
Step back and the picture is consistent. Round One gave us the founder’s discipline: stay flexible, design for retrofit, choose intelligence over hardware. Round Two gave us 15 years of evidence that this is the only discipline that has ever worked in robotics — and the data showing the field is finally ready to reward the founders practicing it. Veena was the supply signal. Andra was the ecosystem signal. Physical AI is no longer a thesis. It is a capital allocation decision.
Practical Implications
If you are “building,” treat the two posts together as the operating manual. Pursue high-value use cases. Make pilot customers pay. Customers need to have skin in the game, even if not full price. Plan financials both ways, since only about 20% of any segment actually wants RaaS. Study founders who failed at your exact use case more than the ones who succeeded.
If you are “investing,” watch headcount growth, not press cycles. Watch order books, not pilot announcements. Underwrite founders who articulate what they expect to “learn” from each step. Most robotics failures are financial and operational, not technical.
If you are an “enterprise buyer,” treat early humanoid pilots as data-gathering. Focus on problems no one is solving today because the labor doesn’t exist or doesn’t want the job.
For help on the legal, capital, and strategic dimensions, Foley & Lardner and Mavka Capital are here. Reach out.
One More Thing
Hard Things Round Two did what we wanted. A small room. A speaker who has paid the price for her opinions. The kind of conversation that doesn’t happen on a panel. We are building, evening by evening, a serialized field guide for the next decade of Physical AI from operators who lived through the wilderness years and stayed. Round Three is coming.
Thank you to the Mavka Capital and Foley teams who make this series happen. Thank you to Tapan for the hand-problem moment we are still chewing on. And thank you to Andrea for reminding the room that the field everyone wants to fund right now is the same field almost nobody believed in 15 years ago. The fundamentals haven’t changed. The capital has finally caught up.
See you at the next one.
Further Reading
- “A $1 Trillion Order Book and a Palo Alto Living Room: What GTC 2026 and Hard Things Taught Us This Week About Physical AI (https://louislehotattorney.substack.com/p/a-1-trillion-order-book-and-a-palo)” — Hard Things Round One, with Veena Radhakrishna of Cartesian Kinetics
- “Robotics and Startups” by Andrea Keay on Substack
- “The Moat Is Dead. Long Live the Runner” by Vitaly Golomb
About the authors
Louis Lehot is a veteran Silicon Valley emerging growth technology company lawyer at Foley & Lardner LLP. Vitaly Golomb is the Managing Partner of Mavka Capital and Mavka Ventures, an early-stage embodied AI fund, and author of “Accelerated Startup.” Together, they co-host the Hard Things invite-only event series. For legal, fundraising, M&A, or strategic advisory on physical AI and robotics, reach out to Louis or Vitaly directly.
