Issue #003 — The Rest

The humanoid headline is always the same two names: Unitree and Agibot, shipping thousands of units, raising the largest rounds, getting the press. The story underneath is less reported.

I spent the last few weeks going through every humanoid company's public footprint I could verify — 90+ companies, looking at what they actually shipped, who bought it, and what the unit economics look like. Three things stand out, and they don't match the press.

1. The 17,000-unit number is mostly noise

Industry trade press in 2025 cited roughly 17,000 humanoid units shipped in China. That number is real, but the label "humanoid" is doing a lot of work. If you restrict to bipedal human-form robots only, the picture changes:

• Unitree and Agibot together: ~10,000 units, both around 5,000 each. Two different industry research firms have them in either order, both call it roughly 50/50. Together that's ~80% of the bipedal humanoid market.
• Leju (Kuavo + Roban + Aelos): high hundreds to low thousands depending on which product line. Their consumer Aelos robots are real shipments, but they're education robots, not factory-floor humanoids.
• Everyone else: a handful of companies at a few hundred units each — Zhongqing Robotics, Stardust Intelligence, LimX Dynamics, Fourier, Galbot — the second tier. Then a long tail where most companies have shipped under 100 units, or are still in prototype / undisclosed stage, or are software platforms with zero hardware shipped.

The "1,000+ unit" claims you see in press releases are often: industrial arms, cleaning robots, or earlier-generation quadrupeds repackaged as humanoids. Once you tighten the definition to bipedal human-form, the entire market compresses to ~17,000 units, 80% of which are two companies.

2. 96% of the units went somewhere that isn't a factory

This is the part that surprised me most. I expected the deployment data to show a heavy industrial use case. It doesn't. Of the 17,000 bipedal humanoids shipped in 2025, the destination breakdown looks roughly like this:

• Education and research institutions: ~42%
• Data collection (robot training, internal R&D): ~19%
• Customer-facing services (retail, hospitality, front desk): ~19%
• Entertainment and demonstrations: ~16%
• Industrial and logistics: ~4%

That last number is the one that matters. Four percent of humanoids shipped in 2025 actually went into a production environment — a factory, a warehouse, a logistics operation. The other 96% went to research labs, university engineering departments, data-collection rigs, and the front of malls.

This means the humanoid "boom" is currently a research-and-development market, not a deployment market. The customers are paying for future optionality (a robot that might be useful in 18 months) or for data collection (running the robot in a controlled environment to train the next model). Very few customers are buying because the robot is doing useful work today.

3. Unitree and Agibot are winning for two completely different reasons

I expected their moats to be similar. They aren't.

Unitree's edge is vertical cost control. From the prospectus: 82–86% of their bill of materials is in-house — motors, reducers, encoders, even dexterous hand components. Their M107 joint motor reportedly costs one-fifth to one-tenth of equivalent overseas units. The result: gross margins around 60% for a humanoid maker, which is unusual. A competitor trying to match Unitree on price isn't competing with a humanoid company; they're competing with an integrated motor-and-reducer manufacturer that also ships a humanoid.

Agibot's edge is ecosystem and deployment speed. Agibot built what they call an "A-chain" — a standardized parts supply chain — and went from 6 prototype units in year one to 15,000-unit production in three years. The 15,000th Gen-2 G2 humanoid rolled off the line and went the same day to a Longcheer (龙旗) tablet production line. They also have a unit in SAIC-GM's Ultium battery plant. This is "ship now, debug in the field" — the opposite of Unitree's cost-control strategy.

Both strategies are working right now. Both will likely face pressure within 12 months:

• Unitree faces a price war as more competitors reach unit-cost parity
• Agibot faces execution risk as field-deployed units fail and require service

The buyers are also the same on both sides: government and public procurement. A representative recent deal: China Mobile's 2025–2027 humanoid procurement, ¥124M total — Agibot took the full-size package (¥78M), Unitree took the small-size package (¥46M). The largest buyer group has consistently been universities, research institutes, and vocational schools — not factories.

What this means for the other 100+ companies

The companies between "shipped a few hundred units" and "still in prototype" are in a genuinely hard place. The tier-2 names I checked have real engineering, real customers in research settings, and a real path to maybe 500 units a year. But 500 units at typical humanoid ASP (~$20K–$50K) is $10M–$25M in revenue — not enough to cover a 100-person engineering team, sales, and supply chain overhead in a competitive Chinese labor market.

The break-even for a standalone humanoid company in China in 2025 looks like ~1,000–2,000 units a year (Unitree and Agibot are both at 5,000+). The companies at 100–500 units are not at break-even. They are subsidized by either venture capital or a parent company that has a strategic reason to keep them alive.

This is not a permanent state. Two of the second-tier names will likely consolidate into the top two. Another 10–15 will pivot to niche applications (medical rehab, agricultural inspection, specific industrial use cases where bipedal is actually a good form factor). The rest will quietly fold or pivot to software-only — selling training data, simulation environments, or model APIs to whoever does end up shipping robots.

The 17,000-unit headline is real. The "humanoid revolution" framing is not. The 2025 market for bipedal humanoids is, in the truest sense, a research-tools market. The next 18 months will be about which of these research tools turn into products that factories actually buy.

That's Issue #003. Next week (Thursday's slot — funding flow): a look at the new companies that raised in the last 30 days and the handful of disclosed valuations that didn't come from AI-generated tables.

→ [Browse the sourced funding table → mazeintelli.com/funding]
→ [Hit reply — especially if you have unit-shipment data for the second tier I'm missing.]