Anvil Robotics

Someone Has To Make The Arms

Here is a thing about the current robotics boom that people mostly don't say out loud: everyone wants to build the robot's brain, and almost nobody wants to build its body. The brain is glamorous. The brain gets the term "Physical AI" and the keynote slot and the foundation model with a name like GR00T. The body is a pile of CNC-milled aluminum, seven servo motors per arm, a Linux box, some cameras, and a shockingly large number of cables. The body is a supply-chain problem wearing a trench coat.

Anvil Robotics, a roughly 14-person company headquartered in Palo Alto (with a mailing address on Hillview Ave and manufacturing in Taiwan), has decided that the body is the business. Its founders, CEO Mike Xia and CTO Vijay Pradeep, met as entrepreneurs-in-residence at Matter Venture Partners - the fund that would later lead their seed round, which is either a nice bit of symmetry or a reminder that venture capital is a small world where the person evaluating your idea is sometimes also the person you got the idea next to.

The idea, roughly, is this. A team that wants to train a robot to do something useful - fold a shirt, pack a box, disassemble a Lego set - currently spends something like six months just assembling the physical robot before it can collect a single frame of training data. Xia's framing, which is good enough that I'll quote it directly:

If you have spent any time near cloud computing you can see where this goes. The entire premise of Amazon Web Services was that renting a server you don't have to rack yourself is worth paying for, because your comparative advantage is your software, not your ability to run a data center. Anvil's bet is that the same trade is about to happen for robots: your comparative advantage is your model and your data, not your ability to source servo motors from four continents. So they'll rack the robot. You bring the ambition.

Legos For Robots

The company describes what it sells as "Legos for robots," which is the kind of phrase that is either marketing or an actual architecture, and here it's mostly the latter. You buy modular, open-source hardware in configurations, not a single monolithic product. The flagship is OpenARM, an open-source bimanual - two-armed - manipulator with roughly seven degrees of freedom per arm, built from CNC parts and sheet metal because those are sturdy, cheap, and boringly manufacturable. It is, in the founders' own comparison, about the size of a middle-school child, which is a mental image I will not be recovering from.

The humanoid-ish form factor is not vanity. A two-armed robot shaped vaguely like a person is a robot a human can teleoperate with almost no learning curve - you strap on a Quest headset, the system reads your pose, high-speed inverse-kinematics solvers translate your arm motion into the robot's arm motion at 1 kHz, and you are now generating exactly the kind of demonstration data that models like ACT, GR00T, and Pi-0.6 eat for breakfast. The unglamorous body, in other words, is also the data-collection instrument. That's the neat part.

Prices run from about $1,900 at the bottom to roughly $10,000 for a fully kitted bimanual setup, with some teleop bundles listed higher. Critically, the designs are open source - the CAD files and URDFs live on the company's GitHub, which means a customer can, in principle, see exactly what they bought and modify it. Giving away the blueprints sounds like a strange move for a hardware company until you notice what it buys: trust, a community that files bug reports and builds accessories for you, and a distribution channel that runs on word of mouth rather than a sales team.

Follow The Money

In April 2026 Anvil announced a seed round. The company and its docs say $6.5 million; a couple of press outlets pegged the number at $5.5 million, which is the sort of discrepancy that happens when a pre-seed and a seed sit next to each other and everyone rounds differently. Either way, the round was led by Matter Venture Partners and Humba Ventures, with participation from Supercharge.vc, Spacecadet, Position Ventures, DNX Ventures, and Superhuman founder Vivek Sodera. There was a $1M pre-seed from Matter in 2025 before it.

The traction line is the one that makes investors lean forward. Since it started shipping in volume around September 2025, Anvil has moved hundreds of robots to more than 60 countries, built a base of 50-plus customers, and - this is the part - did it all inbound. No outbound sales machine, just people finding the shop and buying. When acquisition is entirely word of mouth, it usually means either the market is tiny and everyone already knows everyone, or the product solves a problem so acute that buyers go looking for it. In a field where teams were burning half a year on assembly, it's plausibly the second.

NVIDIA And A Chocolate Factory

The customer roster is where Anvil gets genuinely fun. On one end: Nvidia's GEAR lab, Google, and university labs at Georgia Tech, Columbia, and the University of Toronto - the exact institutions you'd expect to be prototyping embodied AI. Alongside them, a run of well-funded startups: Figure AI, Dexterity, Path Robotics, Formic. And then, on the other end of the barbell, a Portland-area chocolate factory.

That spread tells you something. A tool bought by both the most sophisticated robotics lab on earth and a regional confectioner is a tool that has crossed from "research curiosity" into "thing you buy because you have an actual job to do." The chocolate factory is not writing papers. It wants an arm that works.

In Other Words

Where This Could Wobble

The AWS analogy is load-bearing, and it's worth poking at, because analogies that are too good are how people talk themselves into bad bets. Cloud servers are fungible; a robot arm is not. A server you rent and never see; a robot arm you own, ship across an ocean, and occasionally have to repair. Hardware has real marginal cost, physical logistics, and tariffs - a detail Anvil itself flags in its own keywords, which mention prepaying robot duties, the least sexy line item in the history of robotics.

And "open source as a moat" is a fine strategy right up until a larger, better-capitalized competitor copies your open designs and undercuts you on price - which is the standing risk of every open-hardware company. Anvil's answer, implicitly, is vertical integration and speed: it controls its own Taiwan manufacturing and can air-freight a working robot to your lab in about two days. Speed and reliability are harder to fork than a CAD file. Whether that's a durable advantage or a temporary head start is the whole question, and it's genuinely unresolved. But as unresolved questions go, "can we stay faster than the copycats" is a much better one to be facing than "does anyone want this," which Anvil appears to have already answered.

How It Happened

• 2025
  Founded by Mike Xia and Vijay Pradeep; $1M pre-seed from Matter Venture Partners.
• SEP 2025
  Begins shipping robots in volume; passes 100+ units.
• APR 2026
  Announces $6.5M seed led by Matter Venture Partners & Humba Ventures.
• 2026
  Devkit lineup expands - OpenARM and OpenYAM configurations shipping in ~2 weeks; 60+ countries reached.