Physical AI Market Report: September 2026

On 3 September 2026, Figure announced a partnership with Nscale for up to 100,000 NVIDIA Vera Rubin GPUs, with an initial compute commitment of $3.5 billion and intent to scale past $6 billion. Three weeks earlier, Unitree's 19 August STAR Market listing gave the market the first audited look at humanoid economics from a public mainland issuer.

That contrast is the September story. Physical AI is contracting for training infrastructure at hyperscaler scale, while audited industrial revenue for humanoids remains a thin slice of the business. Warehouse AMRs and collaborative arms, by contrast, already clear cash at research-house scale measured in billions.

This is the September 2026 edition of Inside Deep Tech's Physical AI Market Report series, a catch-up launch covering the landscape as of early September 2026. Below: five takeaways, four snapshot tables with sources, segment deep-dives, named movers, honest limits, Inside Deep Tech's take, methodology, and what to watch in October.

Key Takeaways

• Figure's 3 Sep 2026 Nscale deal commits $3.5B of Vera Rubin compute (up to 100,000 GPUs), while its Series C still stands at $39B from Sep 2025.
• Unitree listed on 19 Aug 2026; its filing shows industrial application at 9.01% of humanoid revenue in 9M 2025, with guided tours half to seven tenths of that slice.
• Interact Analysis puts mobile robots on a 19% CAGR path from under $5B (2024) to $14B (2030), after a 12% forecast cut on tariffs, still far ahead of 2.4% fixed-automation growth.
• IFR World Robotics 2025 (25 Sep 2025): 542,000 industrial robots installed in 2024; operational stock 4.664 million (+9%). Interact (29 May 2026): cobot shipments ~57,000 in 2025 (+14.5%).
• Apptronik closed a $520M Series A-X on 11 Feb 2026 (Series A total >$935M). Boston Dynamics committed all 2026 Atlas product units to Hyundai RMAC and Google DeepMind.

September 2026 snapshot tables

Numbers below are dated and sourced. Where a company has not confirmed a competitor comparison, the report says so. Treat research-house forecasts as research-house views, not filings.

Table 1. Funding, listings, and compute commitments

Here's why that matters. Capital and compute are the scarce inputs private humanoid builders are bidding for in September 2026. Audited industrial revenue is still the scarce output.

Table 2. Humanoid and platform company comparison

Table 3. Mobile robot / AMR market size and structure

Table 4. Regional and segment install base (industrial + cobots)

For the audited Unitree breakdown Inside Deep Tech already published, start here:

What Unitree's filing actually says about humanoid robots

Unitree listed on 19 August 2026. Research and education took 73.60% of humanoid revenue in 9M 2025; industrial application took 9.01%.

Inside Deep Tech

Segment deep-dives

Let's start with the four buckets that actually move budgets in September 2026: humanoids, warehouse AMRs, industrial robots and cobots, and foundation models / software platforms.

Humanoids: capital ahead of industrial mix

Humanoids own the headlines. They do not yet own the cash-clearing industrial revenue mix, at least not in the one audited public breakout available.

Unitree Robotics listed on the Shanghai STAR Market on 19 August 2026 at 150.80 yuan a share (ticker 688836), issuing 40,446,434 new shares. TechNode, citing filings, puts 2025 revenue at about 1.699 billion yuan and net profit attributable at about 278 million yuan. CNBC separately cites 867.8 million yuan of 2025 humanoid sales.

The diligence numbers: more than 5,500 humanoid shipments in 2025, and a 9M 2025 humanoid revenue mix of 73.60% research and education, 17.39% commercial and consumer, and 9.01% industrial application. Corporate guided tours account for roughly 50% to 70% of that industrial slice, which implies something under 5% of humanoid revenue from robots doing industrial work. Average selling prices fell from 593,400 yuan (2023, five units) to 167,600 yuan (9M 2025, 3,551 units) while humanoid gross margins held above 60% in 1H and Q3 2025.

On the private side, Figure's Series C (16 September 2025) exceeded $1 billion in committed capital at a $39 billion post-money valuation. On 25 August 2026, Figure launched Index with 264,000 downloads across 108 countries, more than 44,000 weekly active users, over 16 million videos, 30 minutes of video processed every second, $15 million paid to Creators, and a commitment to spend over $1 billion on data and compute in the next 12 months. Nine days later came the Nscale deal: up to 100,000 Vera Rubin GPUs, H2 2027 Barstow target, $3.5 billion initial compute commitment, intent past $6 billion.

Factory evidence remains narrow and company-sourced. Figure's 30 June 2026 post says Figure 03 is at BMW Plant Spartanburg Hall 52, and that Figure 02 contributed to assembly of 30,000 cars in the prior year, now moving toward sequencing. Boston Dynamics (5 January 2026) says all 2026 product Atlas units go to Hyundai RMAC and Google DeepMind, with additional customers early 2027 and Georgia Metaplant sequencing targeted for 2028. Specs published with Atlas: 56 DoF, 2.3 m reach, lifts up to 50 kg. Hyundai cites a U.S. robotics factory capable of 30,000 robots per year.

Apptronik's 11 February 2026 Series A-X added $520 million after a $415 million Series A in 2025, taking Series A capital past $935 million and total raised near $1 billion. Bloomberg reported a valuation above $5.5 billion. Named partners include Mercedes-Benz, GXO Logistics, Jabil, and a Google DeepMind collaboration on Gemini Robotics.

See also Inside Deep Tech's Figure AI deep dive.

Warehouse AMR: where the dollars already move

Here's the problem with reading Physical AI only through humanoid demos. The mobile-robot ledger is already a multi-billion-dollar market with named research-house forecasts and visible order books.

Interact Analysis (9 January 2026) forecasts mobile robot revenue from just under $5 billion in 2024 to $14 billion in 2030 at a 19% average annual growth rate. That is after a 12% downward revision from its May forecast on tariff uncertainty. Even cut, mobile robots still outpace fixed automation's 2.4% predicted CAGR.

Structure matters as much as the headline CAGR. AGV revenue is expected to fall from about 33% of mobile robot revenue in 2024 to about 20% by 2030 as AMRs take share. China's unit share eases from 58% (2024) to 46% (2030); its revenue share falls from 36% to 27% on lower average prices. Order fulfillment robots are forecast at about half of shipments by 2030. Mobile forklifts are the high-ASP wedge: about one-third of revenue by 2030 on only about 14% of shipments, per Interact.

Warehouse automation order intake rose 7% year on year in 2025 (Interact, 23 February 2026), despite a weak macro backdrop. Project values were inflated by rising steel and labor costs, and a handful of large retailers (Amazon, Walmart, Tesco, and peers) moved the needle. Interact's global warehouse automation CAGR for 2025–2030 is about 6%, with EMEA (~7%) ahead of the Americas (~6%) and APAC (~5%). Interact expects 2026 demand to broaden beyond mega-customers. That is a testable claim for vendor commentary next month.

Industrial robots and cobots: the installed base Physical AI must plug into

Of course, most factory floors still run on classical industrial robots and a growing cobot layer. That install base is the distribution channel for simulation, virtual commissioning, and foundation-model control stacks.

IFR's World Robotics 2025 industrial report (press dated 25 September 2025) recorded 542,000 industrial robot installations in 2024, the second-highest year on record and the fourth straight year above 500,000 units. Operational stock reached 4,664,000 units, up 9% year on year. Asia took 74% of new deployments, Europe 16%, and the Americas 9%.

China alone installed 295,000 units (54% of the global total), with domestic suppliers taking 57% of the home market and operational stock crossing 2 million. Europe fell 8% to 85,000 installs (EU 67,800). The Americas installed 50,100 (−10%), with the United States at 34,200 (−9%). IFR's outlook: global installations up 6% to 575,000 in 2025, and past 700,000 by 2028.

Cobots are smaller in absolute dollars but faster in shipment growth. Interact Analysis (29 May 2026) puts 2025 collaborative robot shipments at approximately 57,000 units (+14.5%), with revenue rising from around $1.1 billion to more than $1.2 billion. From 2025 to 2030, Interact forecasts a 17.3% shipment CAGR to 128,918 units and a 13.6% revenue CAGR to more than $2.3 billion. China's shipment share rises from 54.7% (2025) to 61.4% (2030), while the Americas and EMEA lose share.

That means… operators evaluating Physical AI software should ask which stack attaches to the 4.66 million industrial robots already in service, not only which humanoid walks in a keynote.

Foundation models and software platforms

On the flip side, the software layer is where September's private capital is concentrating: data throughput, simulation, and GPU runway.

NVIDIA's 16 March 2026 GTC release said FANUC, ABB Robotics, YASKAWA, and KUKA (global install base exceeding 2 million robots) are integrating Omniverse libraries and Isaac simulation into virtual commissioning. GR00T N1.7 entered early access with commercial licensing. GR00T N2, previewed on DreamZero research, was described as succeeding at new tasks in new environments more than twice as often as leading vision-language-action models, with availability targeted by year-end. Cosmos 3 was announced in the same wave. On 31 May 2026, NVIDIA announced an Isaac GR00T Reference Humanoid pairing a Unitree H2 Plus body with Jetson Thor and Sharpa hands, slated from Unitree in late 2026.

Figure's Index launch is the clearest company-published data-engine metric set this month: 373 unique tasks, 1,146 objects, and 116 environments per 1,000 hours of Creator video, alongside the throughput and Creator-payout figures above. That is not a third-party benchmark. It is a primary company disclosure about how one private builder intends to feed models.

Skild AI appears in NVIDIA's partner list with ABB, Universal Robots, and Foxconn references at GTC. SoftBank-led funding around $500 million at roughly $4–4.7 billion was widely reported in 2025. Later secondary marks above $14 billion circulate in aggregators. This edition treats those later marks as reported, not company-confirmed primary valuations.

Inside Deep Tech's related primers: NVIDIA's robotics platform and robot foundation models.

Named movers of the month

At first glance, September looks like a Figure week. Zoom out, and the month is a stack of capital, listing, and allocation decisions that reset diligence baselines.

Figure: Index data, then a $3.5B compute runway

Index (25 August 2026) and Nscale (3 September 2026) together say the scarce asset Figure is buying is training throughput. The Series C $39 billion mark from September 2025 is unchanged as a primary round valuation in this edition. The Nscale deal is a compute partnership with a strategic investment from Nscale, not a replacement Series D mark.

BMW remains the clearest named factory reference, with the caveats in the honest-limits section. Sequencing is the stated next use case after sheet-metal contribution.

Unitree: first mainland humanoid IPO with legal consequences attached

Public listing converts marketing claims into prospectus liability. That is why Unitree's 9.01% industrial mix and tour-heavy industrial slice matter more than another walking demo. ASP compression with retained gross margins above 60% in 1H and Q3 2025 is the other filing detail operators should not ignore: volume can rise while industrial utilization stays thin.

Apptronik: Series A crosses $935M

The A-X extension on 11 February 2026 puts Apptronik in the nearly-$1B raised club with strategic LPs spanning autos (Mercedes-Benz), logistics (GXO), manufacturing (Jabil), and hyperscale AI (Google / DeepMind). Bloomberg's >$5.5B valuation is press reporting on the round, useful context, not a substitute for a public filing.

Boston Dynamics Atlas: 2026 supply already spoken for

Allocation is the story. If all 2026 product Atlas units are committed to Hyundai RMAC and Google DeepMind, independent buyers are in an early-2027 queue. The 30,000 robots/year factory citation and 2028 Georgia sequencing target set the calendar. Do not confuse CES hardware with 2026 mass factory labor replacement.

NVIDIA: GR00T, Cosmos, and a 2M+ industrial attach surface

The durable NVIDIA claim for Physical AI diligence is not a single model score. It is the stated integration path into FANUC, ABB, YASKAWA, and KUKA environments covering more than 2 million installed robots, plus commercial licensing on GR00T N1.7 and a year-end target for N2.

Honest limits: what this data does not show

• Unitree's industrial percentage is the best audited humanoid revenue mix available. It is still one issuer in one jurisdiction. It does not describe Figure, Apptronik, Agility, 1X, or Tesla.
• Figure's BMW "30,000 cars" claim is a company statement about contribution, not a disclosed revenue, utilization, or cost-per-station figure.
• Tesla Optimus production and deployment counts remain management commentary without an audited unit table in this edition. They are omitted rather than guessed.
• Interact Analysis mobile-robot, cobot, and warehouse-automation dollars are market research forecasts, not company filings. Treat CAGRs as dated research-house views.
• IFR industrial-robot totals are the authoritative install census for classical robots. They are not a humanoid shipment census.
• Secondary quotes on private names (for example Figure share prints on Forge, or aggregator marks on Skild) are not new primary valuations.
• Warehouse automation's +7% order intake in 2025 includes price inflation from steel and labor. Nominal order growth is not identical to unit demand growth.

This report is written for operators, investors, and technical leaders who need citable numbers. It is not a buy list, a valuation model, or a substitute for primary diligence.

Inside Deep Tech's take

The team at Inside Deep Tech will keep this monthly report boring on purpose: tables first, opinions labeled, and every metric cut if it cannot be sourced.

Methodology and sources

Primary preference for this edition: company newsrooms and filings, IFR World Robotics 2025 industrial press materials (25 September 2025), Interact Analysis press notes (9 January, 23 February, and 29 May 2026), and NVIDIA newsroom (16 March and 31 May 2026). August and early September 2026 sources are prioritized when available. Older figures appear only when they remain the best public authoritative number and are date-stamped inline.

Every quantitative claim in the tables and body above maps to a named source and date in Inside Deep Tech's research log for this edition. Unsourced metrics were cut. Contested or unaudited competitor shipment comparisons attributed only in Unitree's own filing commentary are not restated here as fact.

Internal context used for cross-links (HTTP 200 as of 6 September 2026): Unitree filing breakdown, Figure deep dive, NVIDIA robotics platform primer, robot foundation models primer, 25 fast-growing Physical AI companies, and Physical AI startups to watch in 2026.

FAQ

What to watch in October

October is for evidence, not another keynote reel.

• Any second public issuer that is forced to break out humanoid industrial revenue the way Unitree did.
• Customer-published KPIs from BMW Spartanburg or Hyundai RMAC that survive outside company blogs.
• Ship dates and licensing terms for NVIDIA GR00T N2 after the March preview.
• Whether Interact's hoped-for broadening of warehouse automation demand beyond mega-customers shows up in vendor commentary.
• Follow-through on Figure Index throughput and the Nscale/Vera Rubin build timeline into 2027.
• Cobot ASP and China share commentary that either confirms or challenges Interact's May 2026 trajectory toward 61.4% China shipment share by 2030.

For the wider startup map, see Physical AI startups to watch in 2026. The next Physical AI Market Report lands in early October.