He decided the boring middle of robotics - the wiring, the arms, the supply chain - was the billion-dollar gap. Then he shipped it.
Mike Xia likes to describe today's robotics the way you'd describe a cruel joke. "Imagine building SaaS but being forced to cable up your own servers," he says. "That's where we are today." Every AI team wants to teach machines to fold laundry, sort parts, disassemble Legos. Almost none of them want to spend three to six months and up to $150,000 sourcing motors, mounting cameras and writing control loops before the first training run even begins. Xia counted the months other people were losing, and built a company to give them back.
Above: Xia teleoperating an OpenARM with a pair of VR controllers - the demo that turns skeptics into customers.
Anvil Robotics, the company Xia founded in July 2025, calls itself a composable platform for Physical AI. The plainer version: it sells ready-to-go robot devkits - arms, sensors, cameras, compute, controls - that arrive working out of the box. Press order, get a robot in two weeks, start collecting training data the same day.
The flagship is OpenARM, an open-source dual-arm system with seven degrees of freedom per limb, built to mimic the reach of a human arm. Pair it with a Meta Quest headset and an engineer can puppeteer the robot by waving their own hands, generating the imitation-learning data that hungry AI models need. The whole rig runs a control loop between 500 and 1,000 times a second and speaks ROS2, Foxglove and LeRobot the moment it powers on.
Anvil's customers read like a roll call of the Physical AI frontier: NVIDIA's GEAR lab, Figure AI, Path Robotics, Google, Stanford, Carnegie Mellon, plus a long tail of Y Combinator startups across the US and Southeast Asia. None were chased. Every one of the first fifty-odd came inbound.
What makes Anvil contrarian is what it refuses to do. Most robot startups sell a proprietary design and lock you into their hardware and software. Xia open-sourced the designs. Most outsource manufacturing; Anvil owns its own factory in Taiwan, which lets it air-freight units in as little as a day or two and lets customers pick the country of origin for individual parts - Taiwanese or Japanese components instead of defaulting to China-made.
The shorthand the press reaches for is "Legos for robots" or "the AWS for robotics." Both fit. Xia's bet is that the next decade of robotics won't be won by whoever has the smartest model, but by whoever makes the unglamorous hardware layer disappear.
Imagine building SaaS but being forced to cable up your own servers. That's where we are today.
- Mike Xia, Founder & CEO, Anvil Robotics
Before a single part was ordered, Xia and his co-founder did something unfashionable for a hardware startup: they talked, for six months, to more than fifty companies. The complaint was monotonously consistent. Teams - even well-funded ones - were spending half a year stitching together robot arms, cameras and open-source libraries just to reach a glued-together prototype. Using Universal Robots' UR5 as a baseline, the bill for getting to the starting line of model training could run to $150,000.
"This isn't a problem if you're Tesla," Xia notes. "But for many companies, even well-funded teams, standing up a robotic system is a huge challenge that costs both time and money." That gap - the months of yak-shaving between ambition and a working robot - became the entire thesis.
The co-founder, CTO Vijay Pradeep, came to the problem the same way Xia did. The two met not at a hackathon and not in a lab, but as entrepreneurs-in-residence at Matter Venture Partners. The firm later wrote Anvil's first check - a $1 million pre-seed in 2025 - then led the $6.5 million seed alongside Humba Ventures, with Supercharge.vc, Spacecadet and Position Ventures joining.
To prove the math, Anvil likes a small, almost cheeky demonstration: it trains an OpenARM to take apart Lego bricks for a total kit cost of about $8,600. The robot that builds toys, taken apart by a robot. The point lands.
Xia built Lumina, an AI camera brand, and scaled it to $5 million in annual recurring revenue - his first lesson in turning hardware into a real business.
He helped grow Voltage Park, a cloud-compute platform, to hundreds of millions in revenue - learning how a picks-and-shovels layer wins an AI gold rush.
An MIT education, plus stints linked to Menlo Research and Civitai, gave him the technical and product range he'd later fuse into a robotics company.
The through-line: cameras, then compute, now robots. Each time, sell the layer everyone else finds tedious.
Prices and specs per Anvil's own store. A $150,000, six-month build, reduced to a checkout button.
This isn't a problem if you're Tesla. But for everyone else, standing up a robot costs you both time and money.
- Mike Xia on why Anvil exists
Cloud platforms democratized compute by hiding the servers. Xia wants Anvil to do the same for robots - a default hardware-and-software layer any Physical AI team can build on without losing a season to assembly. The mission, in the founders' words, is to accelerate this new wave of builders "by democratizing access to key technologies through a coherent ecosystem of hardware, software, and data solutions." If he's right, the most ambitious robots of the next decade will arrive in a box, ready to learn.