September 17 briefing: the numbers confirming AI's shift to the physical world — 21,000 humanoids shipped in H1 (+272%), 93-97% made in China, and the world models race attracting $6B in a single quarter
🔎 AI is leaving the screen, and H1 2026 numbers prove it
For ten years, humanoid robotics lived on demo videos. That era is over: between 19,100 and 22,000 humanoids were shipped worldwide in the first half of 2026, growth of 272% to 300% year over year according to research firms. And more than 9 out of 10 units come out of Chinese factories.
The money follows exactly the same curve. More than $3B documented in world models in H1 2026, $900M raised by XPeng Robotics, Unitree going public in Shanghai. And above all, a first auditable industrial track record: at BMW, two Figure 02 robots helped produce more than 30,000 BMW X3 in eleven months.
This September 17 briefing dissects that very shift from demo to deployment. Sourced figures, counterpoints included — because at this pace, data quality makes all the difference between analysis and hype.
Key Takeaways
- 19,100 to 22,000 humanoids shipped in H1 2026 (+272% to +300% year-over-year); the midpoint, ~21,000 units, serves as the reference point for this article.
- More than 97% of shipments are Chinese (Smart Analytics Global); the global top 5 is 100% Chinese and accounts for 86% of the total (Counterpoint).
- AgiBot becomes the world's #1 (~8,400 to 9,700 units, ~44%) ahead of Unitree (~5,900 to 7,000, ~31%), UBTECH (5.2%) and Leju (4.9%).
- BMW + Figure: 2 Figure 02 units, 11 months, more than 90,000 parts loaded, more than 1,250 hours of operation, more than 30,000 BMW X3 produced.
- Figure now produces 1 Figure 03 per hour at its BotQ factory; more than 1,000 cumulative units as of July 23, 2026.
- XPeng Robotics raises more than $900M at a valuation of more than $6.3B — a Chinese record for embodied AI — to mass-produce the IRON in late 2026.
- Unitree goes public in Shanghai: ~$904M raised, ~$9B valuation (~219x 2025 earnings), retail oversubscription above 8,000x.
- World models: more than $3B in H1 2026, including $2.6B in Q1 alone; adding internal programs and compute infrastructure, the bill approaches $6B.
Recommended Tools
Physical AI can also be tracked with tools. Here are the ones that matter in September 2026, from open world models to hosting your news watch.
| Tool | Main use | Price (September 2026) | Best for |
|---|---|---|---|
| NVIDIA Cosmos | Open world foundation models, trained on 20,000 hours of real-world data | Free (open weights) | Training robot policies without a physical fleet |
| Marble — World Labs | Interactive 3D world generation | Pricing not disclosed (see the website) | Prototyping simulation environments |
| GWM-1 — Runway | Generative world models | Subscription (see runwayml.com) | Creators, physical scene previewing |
| BeyondMimic | Robot policies learned from motion demonstrations | Open source | Labs, makers, reproducibility |
| OM-1 — Reward AI | Policy trained solely on demonstrations | Open source (research) | Teams without access to expensive robots |
| GRID — General Robotics | Robotics platform where AI writes 80% of the code | On request | Robotics startups |
| Hostinger | Hosting your AI news site and your demos | From ~€3/month (Sept. 2026, check on hostinger.com) | Publishing analyses and portfolios |
19,100 or 22,000? The H1 2026 figures, decoded
Between 19,100 and 22,000 humanoids were shipped worldwide in the first half of 2026. The first figure comes from Smart Analytics Global, reported by Global Times on August 12 (+272% year-over-year); the second from Counterpoint Research, cited by The Standard on August 20 (+~300%). Keep ~21,000 in mind: that's the solid order of magnitude.
| Estimate | H1 2026 units | Annual growth | Chinese share |
|---|---|---|---|
| Smart Analytics Global (August 2026) | ~19,100 | +272% | >97% |
| Counterpoint Research (August 2026) | >22,000 | +~300% | Top 5 = 86% |
Put these figures in perspective: a year earlier, the market was counted in ~5,000 units. So we're talking about a market that triples in twelve months, from an already significant base — not a case of going from zero to almost nothing.
The official Chinese counterpoint announces more than 40,000 units in H1, in a report presented at the World Robot Conference on August 21 (China Economic Net). The counting perimeters are not comparable: complete units or modules, prototypes included or not.
My reading: the exact level matters less than the slope. Even at the bottom of the range, the market triples in a year. And Morgan Stanley raised its China 2026 forecast to 50,000 units (from 28,000 in January) and 2030 to 446,000. When a bank raises its target by 80% mid-year, it means the ground is moving faster than the models.
97% made in China: why China crushes the volume game
Because more than 97% of humanoids shipped in H1 2026 come from Chinese vendors, with a supply chain localized up to 90% at Unitree. China doesn't just dominate production: it controls the value chain, with more than 400 complete humanoid models — over half of the global total.
The top 4 for the semester is entirely Chinese:
| Player | H1 2026 market share | Estimated units |
|---|---|---|
| AgiBot | ~44% | ~8,400 (up to 9,700 according to Counterpoint) |
| Unitree | ~31% | ~5,900 (over 7,000 according to Counterpoint) |
| UBTECH | 5.2% | ~1,000 |
| Leju | 4.9% | ~940 |
Depending on the scope, the Chinese share therefore ranges from 86% (weight of the top 5 in Counterpoint's count) to 97% (all Chinese vendors per Smart Analytics Global). In every case, the conclusion is the same: the volume is Asian, massively so.
Unitree crosses a symbolic milestone: the first IPO of a humanoid "A-share" stock in Shanghai, with 40.45M shares at 150.8 yuan, ~$904M raised for a ~$9B valuation. Retail oversubscription exceeds 8,000x — Chinese retail investors are rushing into robotics.
Private money is following: 93.5 billion yuan invested in embodied AI in China in H1, across 322 deals (IT Juzi data). As I noted regarding Moonshot AI's $2B raise, China is accelerating across the entire AI stack — software, and now physical.
And this ramp-up isn't staying confined to factories: the appearance of Unitree G1s at Haneda Airport showed that Chinese robots are starting to deploy beyond China's borders, including in Japan. KraneShares sums up the stakes well: the race has moved from pilot to platform.
My take: China isn't winning because its robots are individually superior. It's winning because it applies the drone and EV playbook to humanoids — volumes, supply chain, rapid iteration.
BMW and Figure: finally a verifiable industrial track record
Eleven months of real production at BMW Spartanburg: two Figure 02 units, 10-hour shifts Monday through Friday, more than 90,000 sheet metal parts loaded in over 1,250 hours of operation, and more than 30,000 BMW X3s produced with their contribution. This is the first humanoid track record auditable by an industrial customer — not by a demo video.
On the production side, Figure went from 1 robot/day to 1 robot/hour in less than 120 days at its BotQ plant: more than 350 units by late June, a design capacity of 12,000 units/year for first-generation lines, and more than 1,000 cumulative units as of July 23 (Karmactive).
What's next at BMW: Figure 03 arrived in late June for logistics sequencing and assembly (~40 units deployed by late June), and a pilot is being prepared in Leipzig, with a "Center of Competence for Physical AI in Production" as part of the deal. The program is therefore expanding into Europe.
I explained in Figure 02 and the humanoid robot race why this model was the first to pass the industrial test. These BMW figures confirm it: eleven months without being pulled from the program is more than many of the "renewed 6-month" pilots seen elsewhere.
Let's keep a cool head, though: outside China, verifiable deployments remain in the hundreds, at best in the low thousands of units. The West has the quality track record; China has the volume. Both tell the same story, at different stages.
XPeng raises $900M: IRON mass production targeted for end of 2026
Over $900M raised at a post-money valuation of more than $6.3B: it's the largest private funding round for embodied AI ever completed in China, announced on August 24 (Reuters). And it funds a dated objective: mass production of the IRON humanoid as early as the end of 2026.
The structure of the round says a lot: IDG Capital in the lead, Gaorong Ventures, and above all Alibaba and Tencent as strategic investors. XPeng is injecting $200M, external investors $600M, and executives $100M. He Xiaopeng is personally driving the robotics business — when a carmaker's CEO puts his own name behind robotics, it's no longer an experiment.
Technically, IRON hits the mark: 76 degrees of freedom, including 21 per hand, and three in-house Turing chips for 2,250 TOPS of combined computing power. The funds will go toward hardware and software R&D, training of physical AI models, data generation, factories, and international expansion (XPeng press release).
What strikes me: vertical integration. In-house chips, factories, internal data generation, strategic capital from Chinese web giants — this is exactly the sequence that produced China's electric vehicle champions. Applied to humanoids, it becomes hard to counter.
World models: $3B documented, up to $6B committed — the real war is over simulation
More than $3B was committed to world models startups in H1 2026 (Forbes, June 30), including over $2.6B in Q1 alone. The stakes go beyond funding rounds: whoever controls the simulation infrastructure controls robot training — without needing a physical fleet of matching scale.
| Company | Round (2026) | Valuation | Notable investors |
|---|---|---|---|
| AMI Labs (Yann LeCun) | $1.03B (seed, Q1) | $3.5B pre-money | Bezos Expeditions, NVIDIA, Temasek, Samsung, Toyota Ventures |
| World Labs (Fei-Fei Li) | $1B (incl. $200M from Autodesk) | ~$5B in discussion | Autodesk |
| Runway | $315M | $5.3B | n/a |
| Odyssey | $310M (Series B) | $1.45B | Natural Capital, Amazon, AMD Ventures, GV, EQT, In-Q-Tel |
| ShengShu (Vidu) | $290M (RMB 2B) | n/a | Alibaba Cloud (lead) |
Why "up to $6B" in the headline? Because Forbes' $3B only counts documented rounds. Add the in-house programs — NVIDIA Cosmos, the Waymo World Model derived from Genie 3 — compute commitments, and rounds currently closing, and the bill doubles. Same lesson as with humanoid rollouts: it all comes down to what you count.
Look at the quality of the backers: Bezos, NVIDIA, Samsung, Toyota, Amazon, Autodesk, In-Q-Tel. When the automotive, chip, and cloud industries all invest in the same software layer, it's no longer a fad — it's strategic procurement.
On the open infrastructure side, NVIDIA Cosmos has surpassed 2M downloads and is being adopted by Figure AI, Agility Robotics, XPeng, Uber, and Waabi. A family of world foundation models trained on 20,000 hours of real-world data: that's the equivalent of a CUDA moment for robotics. And the software layer is already consolidating around these infrastructures, as shown by GRID, General Robotics' NVIDIA-backed platform where AI writes 80% of the code.
My bet: in three years, we'll count the physical AI players in two columns — those who rent the simulation infrastructure, and those who built it. The left column will be paying rent.
VLA: why robots are no longer programmed task by task
Vision-Language-Action models are replacing task-by-task coding: the robot executes a policy learned from demonstrations and in simulation, rather than a script hand-written for every single movement. It is this methodological shift that makes the shipping volumes — and all the figures above — possible.
The logic is simple: programming every degree of freedom of a humanoid with 40+ joints does not scale. By contrast, demonstrating the task, simulating it, then letting a VLA policy generalize — that is what can be industrialized. Open source research is converging fast, too: BeyondMimic, the Berkeley and Stanford framework that teaches humanoids to reproduce movements, and OM-1 from Reward AI, a robotic policy trained solely on demonstrations, prove that you can do away with task-by-task code.
Hindustan Times puts it well: what robots are "telling" us today is a change of method, not just of volume. The bottleneck has shifted from mechanics to data — hence the world models war from the previous section. The two topics of this digest are one and the same.
What these figures don't tell us
Three blind spots, and they need to be named. First, the scope of the counts: 19,100, 22,000, or 40,000 — these figures do not measure the same thing, and comparing them as-is is a methodological error before it is a market error.
Next, the quality of demand. Counterpoint projects that services and industrial uses will replace entertainment and data production as the drivers of demand within five years. In other words: some of the units shipped today are still being used to run demos and collect data, not to produce economic value.
Finally, valuations. Unitree trades at ~219x its 2025 earnings with retail oversubscription above 8,000x, and the 446,000 units projected for 2030 by Morgan Stanley remain a projection. AI's physical shift is real; its financial maturity is not yet. The H2 figures will tell whether productive demand follows the shipments.
The H2 2026 milestones to watch
Four events will help validate — or invalidate — this trajectory before the end of the year. Noting them down now will spare you from judging the market based solely on promo videos.
First milestone: the completion of Unitree's IPO, which will make public the financials of an industry leader — a first for calibrating the value of the entire market. Second: mass production of the IRON at XPeng, announced for late 2026, with the goal of catching up with AgiBot's and Unitree's volumes.
Third milestone: BotQ's production rate. Figure reports a design capacity of 12,000 units/year and a BMW pilot in Leipzig in preparation — if the ramp-up stalls, the Western narrative weakens. Fourth: the target of more than 50,000 units over the full year according to Counterpoint (2.1x), and the announced shift of use cases toward services and industry.
My advice: track operating hours and production volumes per customer, not capacity announcements. It's the only metric that separates a deployment from a showcase.
❌ Common Mistakes
Mistake 1: Mixing up counting scopes
Comparing the ~19,100 units from Smart Analytics Global to the 40,000 from the official Chinese report makes no sense: the methods and scopes differ. The solution: always cite the source, date, and methodology, and only compare on a like-for-like basis.
Mistake 2: Confusing shipments with productive deployments
A shipped robot may end up in a showroom or a data-collection factory. Demand the metrics that matter: operating hours, parts processed, production volumes — the benchmark being the REX BMW/Figure (90,000 parts, 30,000 X3s).
Mistake 3: Reducing world models to video models
A video model generates pixels; a world model predicts how an environment evolves under action, in order to train policies. Runway's pivot to GWM-1 illustrates precisely this difference. Confusing the two means missing the infrastructure stakes.
Mistake 4: Reading the Chinese market through a Western lens
93.5 billion yuan across 322 deals, over 400 models, an IPO: the Chinese dynamic is that of an industry, not a handful of labs. Follow supply chains and volumes, not just research announcements.
❓ Frequently Asked Questions
Why do two research firms give different figures (19,100 vs 22,000)?
Because they aren't counting exactly the same thing: types of units included, timing of the count, geographic scope. The gap (~15%) remains small compared with their shared growth (+272% to +300%). Remember the order of magnitude — ~21,000 units — rather than any single figure.
What exactly is AgiBot?
A Chinese player that became the world's No. 1 in humanoid shipments in H1 2026, with ~8,400 to 9,700 units (more than 43% of the market), ahead of Unitree. Its breakthrough illustrates how fast the sector turns over: the hierarchy can change within a few quarters, and volumes are now being contested among Chinese players.
Will humanoids replace traditional industrial robots?
No — in the short term, they complement them. Verified use cases (logistics sequencing, parts loading at BMW) target the loosely structured tasks that traditional automation handles poorly. Counterpoint, for its part, expects services and industry to become the main drivers of demand within five years, ahead of entertainment.
What is a world model, in one sentence?
A model that learns the dynamics of an environment — physical and causal — to predict how it evolves under actions, which makes it possible to train robots in simulation rather than on a physical fleet of matching scale. Examples: NVIDIA Cosmos, Marble by World Labs, GWM-1 by Runway.
How can you keep up with physical AI without drowning?
Three habits: research firm data (Counterpoint, Smart Analytics Global) every six months, stock market filings (Unitree's IPO finally brings public accounts), and daily monitoring through specialized media like humanoid.press. Automate the collection; keep the analysis for yourself.
✅ Conclusion
Physical AI is no longer a keynote promise but a measurable market — ~21,000 humanoids in H1, +272%, 93-97% in China, $3-6B on world models — one that China is industrializing while the West funds simulation infrastructure. What comes next will hinge on a single figure: the share of productive demand in H2 shipments. Publishing your own AI digest? A fast, affordable website is all you need — Hostinger does the job from just a few euros a month.