Feather sells a $30K modular humanoid you bring your own AI to

Feather Robotics came out of stealth this week with a $7.6 million pre-seed led by Gradient Ventures and a product that reads like a deliberate rejection of how everyone else sells humanoids: a modular robot at $30,000 that ships without an AI model, so developers bring their own. For a small operator, the price is interesting. The unbundling is the actual story.

What actually happened

Per TechCrunch, Feather was founded in 2025 by Hoa Mai, formerly of 1X, and Parsa Bakhtiari, a former Tesla Model 3 engineer. The hardware is modular — arm lengths and other components swap for the job — and the company says it has passed $1 million in revenue from early sales. Units are already working as cooks in restaurants and cleaning science labs.

The $30,000 price is roughly half of Unitree's H2 Edu. The model layer is open: developers can run policies from Nvidia, Skild, or Physical Intelligence on the same chassis.

Mai's framing of the gap is the cleanest sentence in the coverage: "You can't buy a Tesla robot today and develop on top of it." Gradient's Darian Shirazi claims Feather is the only US startup building a modular humanoid platform.

That last claim is an investor's, about his own portfolio company, so weight it accordingly. Treat "$1 million in revenue" the same way — it is company-reported and unaudited, and at $30,000 a unit it describes roughly thirty-odd robots, not a production line.

Why unbundled robotics matters for your business

Bundled hardware and model is the oldest lock-in there is. Buy a robot whose intelligence ships welded to the chassis and you have bought both a depreciation schedule and a vendor's roadmap. When their model stalls, your capital sits there being slow. We argue this constantly about API layers and model aliases; it is a sharper argument when the asset costs $30,000 and lives on your floor for five years.

Thirty grand puts physical automation inside an SMB capex conversation. That is a used delivery van. It is not a research budget. It does not mean you should buy one — it means the question moved from "someday" to "what's the payback period on one narrow task," which is a question you can actually answer with a spreadsheet.

Narrow and boring is how this works, same as software. Restaurant prep and lab cleaning are not general intelligence demos. They are repetitive, bounded, low-variance tasks with a clear failure mode. That is exactly the shape of work we tell clients to automate first in software, and the rule does not change because the actuator is physical. If you cannot describe the task in one sentence and measure it in one number, it is not ready.

Sources: TechCrunch.