Matter Venture Partners just raised a $450 million fund to invest in physical AI. We sit down with ounding partner Haomiao Huang about where all that cash will be invested.
AI is eating the world, to paraphrase Marc Andreessen’s famous quote on software. And if so, hardware and physical AI are its teeth, says Haomiao Huang, founding partner of Matter Venture Partners, which just closed a $450 million second fund.
We can all see that agentic AI is finally doing real work: tell an agent to go build something, research something or fix something, and it comes back when it’s done. It might not be perfect, but it’s often serviceable if not excellent. The physical world, however, is not yet like that. There, we still have to do our own work ourselves.
Changing that is exactly what Matter Venture Partners is betting on.
The Silicon Valley firm closed its second fund at $450 million last month, up from its $300 million first fund. The goal, Huang told me in a recent Humanoid Daily podcast, is to back early-stage hard tech startups in semiconductors, robotics and physical AI, quantum computing, advanced manufacturing and energy. Its investors and limited partners read like a map of the global chip supply chain: ASML, TSMC, Quanta Computer, Resonac, Nitto Denko, Sojitz, the Development Bank of Japan and Kleiner Perkins.
“If AI is going to eat the world, hardware will be its teeth,” founding partner Haomiao Huang said in the fund announcement.
I asked him what that means.
“I think the amazing potential of AI is just that it can do almost anything and everything. But when you think about it, we live in a physical world, right?” Huang told me. “And when you think about where it can find its greatest expression in terms of uplifting our lives, making our health better, making our daily lives better, doing the things that we don’t want to do or can’t do, it’s going to be through robots. It’s going to be through physical AI.”
Unsurprisingly, hardware is also what AI runs on. Chips, data centers and fabs are the choke points, Huang says, so physical systems matter on both the supply side and the demand side.
The missing fab for everything
Huang, who earned his PhD at Stanford working on some of the earliest controllable quadcopter drones, has an interesting way of describing the gap between digital and physical.
A modern chip is designed almost entirely in software. It’s verified in simulation, then taped out once or twice.
“Can you imagine a space launch company that’s like, I’m going to design everything in software and then just launch a rocket once and it works?” he asks. “Or even building furniture, right?”
AI and world models are making it possible to design and simulate almost anything digitally. The bottleneck is turning that design into a real object, reliably and repeatedly.
“What’s the equivalent of a TSMC fab that can perfectly replicate in the physical world what you designed in the digital world? That doesn’t exist yet, but it can with physical AI and robotics.”
That’s a big idea. It’s also, essentially, the Matter investment thesis: physical AI is agentic in the world of stuff, matter, things. The vision is that robots will do for us in the “real” world what agents do for us in software. We’ll “prompt” them with questions or requests, and they’ll vacuum the house, clean the toilet, make an omelette (hopefully not in that order).
Or, in the world of industry and business, they’ll unload the truck, transfer the widgets, pack the shipping box, or whatever else they’re tasked with.
Why is physical AI so hot right now?
Robotics has been the next big thing for decades, so why is this time different? Huang points to three things arriving at once.
First, models. The core technology behind LLMs transfers surprisingly well to robot control. Precise motion was never the real problem, Huang says. The problem was edge cases: what happens when a couple of washers get stuck on a factory line? Or there was a small change in production process? AI models generalize, which means robots can increasingly handle the unexpected without a human reprogramming them.
Second, compute. You can now put what’s effectively a supercomputer on the robot itself. (It’s expensive, but doable. And it getting cheaper by the month.)
Third, demand. Onshoring manufacturing and building the data centers and power plants AI needs require physical capacity many richer countries no longer have. Workforces are aging – look at Japan – and replacement workers for manual jobs can be in short supply.
“That creates an infinite amount of demand for robots, which means capital is going to flow in,” Huang says. “And so when you have the models available, the compute available, and the demand there, that’s what creates this magical golden age.”
A humanoid is a state of mind
Then there are humanoid robots. I’m tracking almost 400 companies globally that are building them, and plenty are now ramping production: UBTech, AgiBot, XPeng, Tesla, Figure, Apptronik, 1X.
Huang’s definition of a humanoid is broader than most.
“Humanoid is a state of mind, not a specific type of robot,” he says.
What matters, he says, is capability: multiple arms for fine manipulation, some mobility, onboard compute and power, and a body that fits into a human-sized space. Legs, wheels and height can vary by job. To Huang, a humanoid is to robotics what the general-purpose CPU was to computing: one platform that software can repurpose for many tasks. A spider’s eight-legged form is just another potential iteration; one that could probably do much of our work very successfully.
But even the best of them aren’t ready for much of the work people want done today, he adds. Reliability is still a problem.
“If you have a robot, for example, putting away dishes for you, right? And it’s at 99% reliability, that means it probably on average drops a dish once a week,” Huang says. “Would you employ a house cleaner who’s breaking a dish in your house once a week?”
Robots are still borrowing parts
The bigger investment opportunity, and the reason a chip-heavy fund cares so much about robots, may be the components. (Andrew Kang, the CEO of the ETF RoboStrategy, has similar ideas, although Kang also funds the robot makers.)
Today’s robots are often built from parts designed for something else. Batteries come from lines optimized for laptops and cars, industrial components were engineered for machines that repeat one motion for 15 years. Huang has even heard of one top Chinese actuator maker repurposing a line originally built for automated mahjong tables.
He compares our current state in robot development to the path from mainframes through PCs to the iPhone. That path needed the iPod, which borrowed laptop hard drives, plus lithium-ion batteries and flash memory. Every one of those components scaled through multiple generations before the smartphone was possible.
Robotics hasn’t done that yet (although some manufacturers that are vertically integrating are getting close). But it’s coming. Today’s humanoids, Huang says, are the robotics equivalent of the MIT wearable computing club of decades past: people strapped into bulky gear that would eventually shrink into a phone. Tomorrow’s humanoids will be both smarter and better designed.
What’s next in general-purpose robots
In five years, Huang expects general-purpose robots at scale in factories and warehouses, looking more like Dexterity’s Mech than a two-legged humanoid. In homes, he sees something closer to a pet that’s useful at least part of the time. Frankly, he said, as a father of two, he’d be happy with a robot that put away 80% of the toys.
I’m a little more optimistic about the near term. Figure’s robots trained on its new gig-worker data program performed 6X better on household chores in a 30-home test. That’s just weeks into the program, and better hands are coming fast.
An 80% home robot could arrive sooner than Huang’s five years. But his bigger point holds: the robots, chips and components that make physical AI truly useful are still being invented.
And that’s what the $450 million is for.
This article was originally published on Forbes.com