What robotics companies look like, and what they’re betting on

A study by Binh Pham (flower and bolt, https://garden.binhph.am) of 652 robot companies in 44 countries: their robots, brands and words, the money behind them and the AI models they release. Data as of 2026-11-19. Read it in the interactive study, or download companies.csv, models.csv, rounds.csv and db.json.

Findings

2% → 42% · Robot money moved to humanoids and robot brains.

*In short:* Humanoid makers and robot AI model companies took 42% of the venture money raised in 2024–26, up from 2%.

Ten years ago, almost nobody funded robots shaped like people, or the AI that runs robots. In 2015–20 they took 2% of the venture money we traced. In 2024–26 they took 42%.

And there was far more money to take. $60B came in during 2024–26 alone, 74% of everything we traced. Most of what humanoids and robot AI didn't take went to defense.

Note: From 912 public funding rounds across 411 venture-backed companies. Robot AI includes Wayve, which builds software that drives cars; without it, the 2024–26 share is 40%. Undisclosed rounds, mostly Chinese, and companies that failed are missing.

72% · 64% · Seed money bets on humanoids and robot AI. Late money goes to defense.

*In short:* Humanoid makers and robot AI companies took 72% of the seed dollars we traced, while defense took 64% of the money in Series D and later rounds.

A startup's first cheques and its last ones go to very different companies. At seed, the first real round, humanoid makers and robot AI companies drew 72% of the money we traced. At Series D and later, defense drew 64%.

Robot AI led at Series A, the round after seed. The biggest late cheques went to companies like Shield AI, which makes military drones, and Saronic, which builds boats that steer themselves.

Note: Disclosed rounds only, added up by stage. Chinese humanoid makers often keep their rounds private, so their totals are a floor.

9 of 13 · Most of China's humanoid unicorns deliver. Most of the West's don't.

*In short:* Of the humanoid makers valued at $1B or more, 9 of the 13 Chinese ones have robots with paying customers, while only 1 of the 7 Western ones does.

Unitree and Figure are both worth more than $1B. Unitree sells its G1 humanoid to customers. Figure says it has built over 350 robots, all for its own teams.

That split runs through the whole group. In China, 9 of the 13 humanoid makers worth $1B or more sell or lease robots to paying customers. In the West, only Agility does, and it leases Digit as a service. The 10 without paying customers are worth $67B of the group's $104B.

Note: GigaAI, DexForce, Simplexity and Walden, all worth $1B since 2026, are still in trials.

26 of 104 · Raising $100M used to take years. Now it can take two.

*In short:* None of the 160 venture-backed robot companies founded in 2010–17 raised a round of $100M or more within two years, compared with 26 of the 104 founded in 2022–24.

Picture a robot startup on its second birthday. Of the companies founded in 2010–17, none of 160 had raised a single round of $100M by then. Of those founded in 2022–24, 26 of 104 had.

Figure AI, Sharpa, Rhoda AI and Generalist AI are in the new group. Money that once took most of a decade now arrives in two years.

Note: One public round of at least $100M by the second calendar year after founding, loans left out. Founding years are approximate, and only companies that survived are counted.

$962M vs $76B · Humanoids and robot brains sold $962M last year. Investors value them at $76B.

*In short:* The 9 makers of humanoids, legged robots and robot AI with known sales sold $962M last year and are valued at $76B; the typical one is valued at 51 times its sales, against 9 times for 33 other robot makers.

Unitree sold $236M of robots last year. Fanuc, the old giant of factory arms, sold $5.7B, about 24 times as much. Investors value Unitree at $37B, about 157 times its sales, and Fanuc at $30B, about 5 times.

That gap runs through the group. The 9 makers of humanoids, legged robots and robot AI we could check sold $962M between them last year and are valued at $76B. Skild AI, which sells software that controls robots, sold $30M and is valued at $14B.

Put that next to companies outside robotics. Apple is worth about 11 times what it sells, Nvidia 9 and Tesla 14. Figma and CoreWeave, both listed in 2025, are at 6 and 2. The typical humanoid maker is priced like Palantir, at 49. Only the AI labs go higher: OpenAI at about 65 times its 2025 sales, and Anthropic, which went public this month, at about 327.

Note: The typical humanoid, legged robot or robot AI maker is valued at 51 times its yearly sales, other robot makers at 9 (listed companies alone: 94 against 9). The groups are small. Listed companies outside robotics: market value on 20 November 2026 over the last 12 months of sales. OpenAI: its March 2026 round over 2025 sales. Anthropic: its IPO valuation over 2025 sales; those sales grew twelvefold that year, and on its October run rate it is valued at about 19 times.

19 of 384 · 19 companies raised half the money.

*In short:* Of the $88B raised by 384 venture-backed robot companies, the top 19 took half, and the bottom half of companies took 4.5%.

Line up 384 venture-backed robot companies, biggest fundraiser first. Together they raised $88B. Count to 19 and you have passed half of it. The 192 at the back of the line share 4.5%.

The front of the line isn't humanoids. Anduril, which makes military drones, raised $11B on its own. Zipline's delivery drones and the defense makers Shield AI and Saronic come next. The first humanoid maker, Agile Robots, is number 9.

Note: Equity only, no loans. For each company, the larger of its reported total and the rounds we traced. Undisclosed rounds and failed companies are missing.

78 vs 30 · China now publishes most of the world's robot AI models.

*In short:* In 2026, 28 Chinese companies released 78 robot AI models with a public write-up, compared with 30 models from 15 American companies, reversing the previous year’s US lead of 36 to 33.

A robot AI model is the software that turns what a robot sees and hears into what it does. In 2025, American companies published more of them, 36 to 33. In 2026, Chinese companies published 78 and American companies 30.

More Chinese companies joined in, too: 28 published a model in 2026, against 15 in the US. Robbyant's LingBot-VLA 2.0 is one of them. Many others came out of papers written with universities.

Note: Trained models with a public paper, report or blog post, from 67 companies. Toolkits and simulators are left out. A model counts whether or not anyone used it.

59% vs 20% · China opens its robot models. America is closing its own.

*In short:* 59% of the robot AI models Chinese companies released in 2026 come with open weights, compared with 20% of American releases, down from 33% in 2025.

Say you want to run one of those models on your own robot. A paper tells you how it was made. To run it, you need the weights, the files that hold what the model learned.

In 2026, 46 of 78 Chinese releases shared those files. 6 of 30 American releases did. A year earlier the gap was small: 45% against 33%.

Note: Open weights means the model files can be downloaded (p < 0.001). Training code, data and licence terms can still be closed.

18 to 1 · The biggest robot models learn from people, not robots.

*In short:* In the 10 largest training sets that state their sources, 2.1M hours of human data outweigh 120k hours of robot data by 18 to 1.

Imagine teaching a robot by filming your own hands. That is how the biggest robot models now start. Dyna-2 began with over a million hours of video shot from people's point of view, and no robot data at all. Generalist's GEN-1 used half a million hours from people wearing recording gear.

Across the 10 largest training sets, people outweigh robots 18 to 1: 2.1M hours against 120k. In 2024, the biggest set, for Physical Intelligence's π0, was 10,000 hours of robot data. Two years later, the biggest sets are fifty to a hundred times larger, and mostly human.

Note: Training sets that publish a breakdown of their sources. Robots often practise each task afterwards, sometimes for an hour or so, and some teams leave those hours out.

27% vs 8% · Chinese brands say they lead. American brands say what they build.

*In short:* 27% of Chinese robot companies call themselves leading, first or a pioneer on their homepage, compared with 8% of American companies, which more often talk about what they build.

Open the English homepage of a robot maker and it will usually tell you one of two things: that it leads, or what it makes. HAI Robotics calls itself a leader in warehouse robots. Hello Robot tells you what its Stretch robot helps people do.

Chinese companies lean one way, American companies the other. 27% of Chinese pages say leading, first or pioneer, against 8% of American pages. Build, make or help appear on 30% of American pages and 11% of Chinese ones.

Note: English homepage titles, descriptions and taglines (p < 0.001). 25 of 47 Chinese makers founded since 2023 open their homepage in Chinese, so this is how companies introduce themselves to English readers.

1.2k vs 6.2k · China's humanoid makers are nearly absent from X.

*In short:* The typical Chinese humanoid maker with a recorded count has 1.2k followers on X, compared with 6.2k for American makers, and we found no account for 15 of 44 Chinese makers.

If X were your only window onto humanoids, America would look far ahead. Figure has 227k followers and Apptronik 24k. AGIBOT, worth $7.5B, has about 6.6k. RobotEra, worth $1.4B, has under 1k. Yet Chinese makers ship more humanoids.

The typical Chinese humanoid maker has 1.2k followers on X. The typical American one has 6.2k. We couldn't find an X account at all for 15 of the 44 Chinese makers.

Note: Medians over makers with a recorded count (p = 0.012). Weibo, Douyin and Bilibili aren't counted, and some Chinese makers are big elsewhere: Galbot has over 500k subscribers on YouTube.

4 of 106 · Humanoid makers sell intelligence, not autonomy.

*In short:* Only 4 of 106 humanoid makers use autonomous or autonomy on their homepage, against 22% of other robot companies.

Picture a Starship delivery robot rolling up to someone's door with nobody steering. That is what autonomous means, and 63% of delivery robot makers say it on their homepage. Among humanoid makers, only 4 of 106 do, against 22% of all other robot companies.

Humanoid makers talk instead about what their robots could learn. Intelligence appears at 28 of them, embodied, meaning intelligence in a physical body, at 14, and general-purpose at 10. The few that do say autonomous are General Autonomy, Hexagon Robotics, Muks Robotics and Tesla.

Note: Homepage titles, descriptions and taglines (p < 0.001). The words alone don't tell us how often a person has to step in.

44% vs 6% · Robots promise intelligence. Almost none promise safety.

*In short:* 44% of robot makers lead with intelligence and 6% mention safety, including 6 of 184 makers whose robots share a room with people.

Imagine a LOVOT rolling up to you at home, or a Digit walking past you in a warehouse aisle. What does the maker tell you first? Across 646 robot companies, 44% lead with intelligence: AI, smart, intelligent. 6% mention safety.

The gap is widest where you would expect it to be narrowest. Of the 184 makers of humanoids, home robots and companions, the robots built to share a room with people, 94 lead with intelligence and 6 mention safety.

Note: The first lines of each homepage: title, description and tagline. Safety pages, certifications and manuals deeper in a site aren't counted.

81% · Robots don’t wear the brand.

*In short:* 81% of robots are white, grey, black or bare metal, about the same share as for new cars, and only 63 of 578 companies use their own brand colour on the robot.

Picture a row of robots next to a row of parked cars. The colours would look the same. Of the 1,095 robots we measured, 81% are white, grey, black or bare metal, about the same share as new cars.

The brand colour stays on the website. Unitree's site is bright blue, but its G1 is grey and black. Only 63 of 578 companies put their own brand colour on their robot.

Note: Body colour measured from robot photos. Brand colour is the main accent of each company's palette, compared with the first robot pictured. Photos show the finish a maker chose to show; other colours may be on offer.

Source: New cars, including silver: 81%. Axalta 2025 (https://www.specialchem.com/coatings/news/axalta-unveils-2025-global-automotive-color-popularity-report)

66% · Chinese robots get codes. American robots get names.

*In short:* 66% of Chinese humanoid robot names contain a number, as in G1 or Walker S2, compared with 37% of American names, which include Atlas, Optimus, Apollo and Digit.

Put Unitree's G1 next to Boston Dynamics' Atlas. One sounds like a part number, the other like a character.

That is the pattern. 66% of Chinese humanoid names contain a number, like G1 or Walker S2. 37% of American names do; most are words like Atlas, Optimus, Apollo and Digit. Korean makers often go further and give their robots people's names.

Note: Distinct robot names per country with notes in brackets removed, counted as coded if they contain a digit. Walker S2 counts as coded, even though Walker could be a name.

Method

· *Sample:* 652 companies in 44 countries with a real robot and public imagery, as of November 2026, found by research agents across venture portfolios, university spin-outs, funding news and industry sweeps. Failed companies are missing, which flatters every comparison of leaders.

· *Who counts:* A company is in if it designs and builds its own robots or develops its own robot AI. Integrators, distributors and resellers are left out. For borderline cases, such as companies that run or adapt other makers' robots, three things decide: a brand of its own, a business that makes money, and credible funding. It must also show its robot, or its robot AI, in public: a photo, a demo or a release. Young companies join once they do, so the most recent founding years are undercounted.

· *Measured:* We measured body tone from pixels in robot cutouts and homepage darkness from screenshots. We read fonts from the browser and recorded each brand’s colour palette. We checked robot photos by eye.

· *Judged:* One research agent per company classified the face, shape, type style, logo style and brand persona. These classifications are a first pass and have no panel ratings.

· *Outcomes:* Funding from public reporting. Audience from public follower counts on X, YouTube, Instagram, LinkedIn and TikTok. Corporate divisions are excluded from startup comparisons.

· *Data:* All 652 companies, one page each; companies.csv for a table, db.json for every field, SCHEMA.md for what each field means, and llms.txt for agents.

· *Images:* All logos and images belong to their owners, are shown for research, and record their source.

Cite

Pham, B. & Claude. (2026). What robotics companies look like, and what they’re betting on. https://pham-tuan-binh.github.io/robotics-design-study/ (accessed 2026-11-19)