36,000 Humanoids Sold in H1 2026 — But Only 3% Went to Work

Frost & Sullivan’s 2026 Global Humanoid Robot Market Report, released on 28–29 September, has one headline that travelled fast: in the first half of 2026 the world bought 36,000 humanoid robots, more than in all of 2025, and every one of the top five vendors by revenue is Chinese.

That headline is accurate. It is also, on its own, close to useless for anyone actually specifying or buying a humanoid platform. The interesting numbers are one layer down: how the revenue splits by machine class, how much of it comes from robots doing paid work rather than waving at a trade show audience, and why a second research house puts the same half-year at 25,000 units instead of 36,000.

The reported figures

Top five by revenue in H1 2026: UBTech RMB 620M (16.5%), AgiBot RMB 610M (16.1%), Unitree RMB 580M (15.3%), Galaxea RMB 210M (5.5%), Leju RMB 150M (4.0%) — 57.4% combined. Frost & Sullivan states that intra-group and related-party transactions were stripped out to align with audited revenue.

The average selling price went up. That is a mix effect.

Divide revenue by units and the ASP rises from about US$14,300 in 2025 to US$15,000 in H1 2026 — up roughly 5%, while every trade headline of the past six months has been about prices collapsing.

Both are true, because the average is being pulled by composition. Split the market into full-size embodied humanoids and everything else:

• Full-size machines: ASP rises from about US$49,100 to US$55,200 (+12%), and their share of units grows from 15% to 16%.
• Everything else (desktop arms, education kits, entertainment shells, small bipeds): ASP falls from about US$8,100 to US$7,300 (−10%).

The −10% is the price war, and it matches what is happening at retail: Unitree’s R1 dropped from RMB 39,900 to RMB 29,900, Songyan’s “Xiaobumi” launched at RMB 9,998, Qiyuan’s Q1 and T1 start at RMB 19,999. The +5% at the top of the page is a weighting artefact from selling a few hundred more full-size units.

Why this matters for procurement: a “market price” benchmark is meaningless in this category. A US$15,000 average describes a market that does not exist — nobody sells a US$15,000 humanoid. You either buy a sub-US$10,000 unit, or you buy a US$50,000+ full-size platform, and the bill of materials behind those two products has almost nothing in common.

Only 3.3% of units do production work — but they carry 14.4% of revenue

Frost & Sullivan splits demand into four scenes: entertainment, research and education, data collection, and production work. Entertainment is flat, education slipped, data collection is roughly unchanged — it exists to feed embodied model training rather than to earn revenue from a customer’s process.

Production work is the only scene compounding. It grew from 700 units and US$47M in 2025 to 1,200 units and US$78M in H1 2026, +65.9% in half a year against a full-year base.

Run the shares:

Two things follow. First, the revenue per unit is roughly four times the market average. A production-work deployment is not a robot sale; it is a robot plus integration, safety validation, process engineering and service. If you are budgeting from a list price, you will be wrong by a multiple, not a margin.

Second, the scene that everyone cites in a pitch deck is still 3% of the volume. That is the honest state of the industry in mid-2026: thousands of units proving the technology, and about a thousand units proving the business case.

The leaderboard is flattening

UBTech held first place in 2025 with 24.8% of revenue. In H1 2026 it is still first, at 16.5% — a fall of more than eight points in six months, while AgiBot and Unitree closed to within 1.2 and 1.9 points.

UBTech’s own volume tells the story better than the share does: 16,123 units, about 45% of everything sold worldwide, for 16.5% of the revenue. That works out to roughly RMB 38,500 per unit, well under the market average. AgiBot, at over 8,600 units and RMB 610M, sits near RMB 71,000 per unit.

Read that as a strategic split rather than a ranking. One vendor is buying volume in entertainment, education and data collection; another is concentrating on fewer, dearer machines. Both are rational. Neither tells you who will win production work, because production work is currently decided on hours of uptime, not units shipped.

The overseas picture is consistent with that: Tesla’s Optimus, Figure and Agility all have credible hardware and embodied-model work, but Frost & Sullivan places them in prototype iteration and small internal pilots. Optimus Gen-3 did enter volume production in Q2 2026 — into Tesla’s own lines, which the report’s related-party exclusion means is largely outside the counted revenue.

Two research houses disagree by 44%

IDC puts H1 2026 global humanoid shipments at close to 25,000 units, +432% year on year, with China at about 19,000 and 77.9% of the world. Frost & Sullivan says 36,000. That is a 44% gap on the same six months, and neither number is wrong.

The difference is scope. IDC counts shipments of what it classifies as humanoid robots; Frost & Sullivan’s 36,000 includes entertainment units, education kits and data-collection machines, which is where the long tail of small vendors lives. When you see a humanoid market figure, the first question is not “how big” but “counting what”.

Practical rule: if a number is used to justify a market-entry decision, demand the definition. If it is used to justify a component order, ignore it and count your customer’s purchase orders.

What this means if you are sourcing

1. Rank by revenue, buy on uptime. The report ranks vendors by revenue with related-party deals removed — a good proxy for external commercial traction, a poor proxy for whether a specific arm will hold ±0.05 mm for 20,000 cycles.
2. Ask which stage the reference is at. Frost & Sullivan’s five-stage model — technical demo, customer pilot, paid deployment, routine operation, scale replication — is a usable scoring rubric. A logo wall of pilots is stage two. Ask for paid deployment with a named process and a monthly unit count.
3. Budget the integration, not the robot. At roughly US$65,000 revenue per production unit against a sub-US$50,000 machine price, integration and service are the majority of the cheque. Get that scope quoted separately.
4. Treat the price war as a component opportunity. Sub-US$10,000 bipeds are only possible with cheaper harmonic drives, frameless torque motors and tactile sensors. If you buy those parts, the same pressure is working for you — but verify that the cost-down came from design, not from tolerance.
5. Expect the volume leader and the technology leader to diverge. With the top five within 1.2 points of each other at the top and the leader down eight points in half a year, supplier concentration risk is lower than it was — and so is the value of picking “the market leader”.

Sources and method

Figures are from Frost & Sullivan’s 2026 Global Humanoid Robot Market Report (released 28–29 September 2026), as reported by China Securities Journal, Securities Times and Economic Daily. Shipment comparison from IDC’s H1 2026 humanoid robot report. Retail prices from vendor announcements; Goldman Sachs’ physical-AI report is the source for the ten-year average-price forecast (US$41,800 in 2025 to US$21,300 in 2035). Percentage shares, per-unit revenue and ASP splits marked as implied are computed by Latest Robot News from the reported figures and are not part of the original report.

Related reading

• Embodied AI funding, September 2026: big rounds cool, money moves upstream — where the capital behind these volumes is actually going.
• Robot component sourcing: from prototype to production — how to qualify a supplier once a platform reaches volume.
• Robot tactile sensors compared: flexible arrays, magnetic Hall, vision-based — the sensing bill of materials behind a dexterous hand.