Augment humanity with physical intelligence.
World Engine turns human experience into training data for humanoid robots. Founded by a USC professor and a Waymo foundation-model co-lead, it supplies eight frontier labs with the data, evaluation and infrastructure they need to build physical foundation models.
Thirteen chapters plus an appendix. Chapters 3 to 8 explain the company, chapters 9 to 13 cover the market, the competition and the business model, and the appendix defines every technical term. Two chapters are interactive.
World Engine sells the data, evaluation and infrastructure that humanoid labs need to train their models. This is the case for investing, the scale of the opportunity, and the numbers behind it.
No humanoid programme can build a general-purpose robot without a very large amount of human experience, and almost nobody has worked out how to collect it reliably at scale. World Engine does that job. In ten months it has become one of the few suppliers the best-funded robotics labs trust with the input their models depend on: roughly $100M of annualised run-rate, eight frontier-lab customers and 30 employees.
| Metric | Figure | Source |
|---|---|---|
| Run-rate, 2026 | ~$100M | Company-reported; echoed publicly by investor UpScaleX |
| Customers | 8 frontier labs | Company-reported; echoed publicly by UpScaleX |
| Employees | 30 | Company-reported |
| Hours processed to date | 200K | Company-reported |
| Signed deliveries | $15M within 2 months | Company-reported |
| 12-month data target | 12.5M hours | Company-reported |
| Research honours, 2026 | 3 CVPR awards (2 best paper, 1 runner-up) | Public (conference results, paper and code) |
| Known investors | ZWC Partners, UpScaleX | Public (PitchBook) |
An unusually fast ramp for infrastructure sold to the most demanding buyers in AI. Customer count and run-rate are also referenced publicly by investor UpScaleX.
Digital intelligence compounds on internet-scale data and billion-user deployment. Embodied models have neither: they train on 50 to 1,000 times fewer tokens than driving models and have almost no deployment surface to learn from.
The window opened in the last twelve to eighteen months and will stay open while labs compete on capability. Three forces did the work.
Vision-language-action models such as Ψ₀, NVIDIA GR00T and the π series now transfer across tasks, so data has moved from a research input to a production input. Labs will pay for volume and reliability once they know what to feed the model.
Around 13,000 to 16,000 humanoids shipped in 2025, with sector revenue below $1B. Goldman Sachs now models 6.5M units a year by 2035 and a market of roughly $138B, up from its earlier $38B. Every one of those programmes needs data before it can scale units.
About $50B went into physical AI in the first half of 2026 across roughly 500 deals (Crunchbase), and the median strategic-backed round has risen to about $100M. Buyers can fund multi-year data programmes today without waiting for robot revenue.
World Engine sits between robotics labs that need data and a global workforce wearing capture hardware. Labs send data specifications; operators return experience that has been engineered to speak humanoid.
A three-tier data pyramid trades volume for fidelity. World Engine owns the whole stack, from capture devices to post-processing to operations, so every hour is usable for training.
An autoregressive transformer predicts interleaved world and action tokens in one stream (language, vision, world state, action). Owning a reference model lets World Engine see which data moves performance, and an open model makes every developer on it a future data customer.
The company's position only makes sense once you see the pipeline. Click through each stage and tier below: these are the two graphics the whole investment case rests on.
One founder builds the science, the other has shipped a 140B-parameter robotic foundation model on a fleet of 1,000+ vehicles. That pairing is rare, and it is the main reason this team can sell to frontier labs.
Ψ₀ shows that a well-designed data curriculum beats brute-force scaling. It was trained on about 800 hours of human video plus 30 hours of robot data, roughly 8% of the data comparable methods use, and still outperformed strong humanoid baselines by more than 40% overall.
The highest-spec humanoid interaction dataset for real everyday tasks.
Open humanoid foundation model for long-horizon loco-manipulation, 40%+ above strong baselines.
Physical world models that the company says are a year ahead of embodied-AI startups.
Ten months in, World Engine reports roughly $100M of annualised run-rate from eight frontier-lab customers, plus $15M of deliveries signed within a two-month window. The figures below are the ones that matter as the business scales toward its 12.5M-hour plan.
| Metric | Figure | Source |
|---|---|---|
| Annualised run-rate, 2026 | ~$100M | Company; also cited publicly by UpScaleX |
| Enterprise customers | 8, all frontier labs | Company; also cited publicly by UpScaleX |
| Pipeline | 3 additional logos | Company |
| Signed near-term deliveries | $15M to deliver within 2 months | Company |
| Data processed to date | 200K hours (ego-centric, glove, teleop) | Company |
| Initial pilots | VinMotion (infrastructure), Becton Dickinson (deployment) | Company |
| Derived metric | Implied value | How we get there |
|---|---|---|
| Revenue per customer | ~$12.5M average | $100M across eight frontier labs |
| Delivery pace | ~$90M annualised | $15M delivered in two months, in line with the run-rate |
| Blended price per hour | ~$33 | $100M across roughly 3M hours of data delivered |
| Hours per dollar of revenue | ~30 hours per $1,000 | The same calculation, tracked as the mix moves toward ego-centric hours |
Thirty employees producing a $100M run-rate is roughly $3.3M per person, about ten times a good software company and higher than the most productive AI labs. Operators are clearly not on the payroll, so the labour cost sits in cost of goods sold rather than in headcount.
Two consequences follow. Headcount is not a measure of scale here, so "30 people" understates the business rather than describing a small one. And because the largest cost is bought rather than hired, margin is the line that improves as automation and tooling mature across the three tiers.
Eight customers at a $100M run-rate means one frontier lab is worth more than a dozen mid-market customers, and that a single onboarding is a material event. The economics of acquiring one, therefore, decide the growth model.
A low single-digit-million onboarding that pays back inside twelve months makes growth capital-light and repeatable, which is what puts the pipeline of three further logos within reach. Customer-level economics are the highest-leverage part of this story.
| Hours / year | $15 / hr | $25 / hr | $33 / hr |
|---|---|---|---|
| 3M hrs | 45 | 75 | 99 |
| 6M hrs | 90 | 150 | 198 |
| 12.5M hrs | 187.5 | 312.5 | 412.5 |
Nobody agrees on how big humanoids become, so we sized the layer World Engine actually sells to: the data, evaluation and infrastructure spend of the programmes building them. Three independent anchors land in the same range.
Total frontier-lab spend on humanoid data across all suppliers is plausibly $300M–$1B a year. A ~$100M run-rate would place World Engine among the leading suppliers with an estimated 10–30% share of that spend, before the market grows.
If the 12.5M-hour plan holds and the blended price drifts down to $20–$33 per hour, that plan alone implies $250M–$400M of revenue — several times today's run-rate, in a market we estimate at a few billion dollars by 2030.
Three independent methods converge on a data and infrastructure layer worth low single-digit billions annually within five to ten years. That is small next to the hardware market, but it is the earliest, most concentrated and most defensible part of the value chain, and it is being bought today.
The category is well funded and being priced aggressively. World Engine's advantage is breadth: three data tiers inside one pipeline, with a reference model that shows which data actually moves performance. The moat scores below are Tier One's, out of five.
| Player | Approach | Funding signal | Assessment vs. World Engine |
|---|---|---|---|
| World Engine | Ego-centric video, sensorised gloves and teleoperation; evaluation; an open humanoid model; infrastructure | ~$100M run-rate on 8 frontier labs (company-reported) | Widest data coverage plus the model feedback loop |
| XDOF | Teleoperation data and gripper hardware; ABC dataset with UC Berkeley | $70M Series A (Jun 2026, Thrive, a16z, Lux, Spark); reported in talks at ~$1.2B led by 8VC | Deep teleoperation bench, concentrated in the teleop and gripper tiers |
| Mecka AI | Human motion capture for humanoid training | Reported near a $500M valuation with a Sequoia-led round; a $60M round was reported led by Framework Ventures earlier in 2026 | Strong investor line-up; focused on human motion capture |
| Scale AI, Micro1 | Human-data platforms extending from language models into robotics | Large private incumbents | Enterprise distribution earned in the language-model market |
| Labs and OEMs | In-house capture and annotation teams | Funded from their own balance sheets | Early specification setters with the balance sheet to staff in-house teams |
This is not the first business built on selling human labour as training data. The outcomes were decided by two things: whether the supplier owned proprietary technology, and how concentrated its customers were.
| Precedent | What it built | How it ended | Lesson for World Engine |
|---|---|---|---|
| Appen | Built the largest contractor network for language data; listed on the ASX; peaked above a A$3B market capitalisation with revenue in the hundreds of millions. | Google, its largest customer, cut and insourced spend. Revenue fell sharply, margins compressed and the share price collapsed by roughly 90% from its peak. | The cautionary case. Labour-heavy supply plus concentrated customers plus no proprietary layer is the exact shape World Engine must avoid. |
| Scale AI | Moved from autonomy labelling to an enterprise data engine, and became strategically important to the model labs. | Reported to have reached a valuation around $14B when Meta took a large stake in 2025, despite no longer being the cheapest option. | Shows how valuable the position can become when a data supplier owns proprietary tooling and workflow, not just a workforce. |
| Surge AI | Bootstrapped a human-data business aimed at frontier labs rather than at enterprise buyers. | Reported to have grown to very substantial revenue with unusually high margins for a data company, without raising a large war chest. | Selling to a handful of frontier labs is extremely profitable when it works, but the price is negotiated with those same handful of buyers. |
| Expert-data marketplaces | Platforms that match subject-matter experts to model builders, priced per task. | Reported valuations in the billions on relatively modest revenue, funded on growth rather than margin. | Competition for the same talent pools World Engine hires as operators, which is a cost risk at scale. |
| Tesla and Waymo fleets | Own the collection asset: vehicles that generate data simply by operating. | The strongest version of this business, because the data arrives as a by-product of a commercial activity. | The end state World Engine is aiming at with deployments. Reaching it requires robots in the field, which is a bet on customers, not on its own execution. |
Data revenue funds the infrastructure, the infrastructure creates lock-in, and deployments generate the real-world feedback that makes the model and the data better.
Every technical term used in this memo, defined so the analysis can be read without a robotics background.
| Term | What it means | Why it matters here |
|---|---|---|
| Physical AI | AI that acts in the physical world rather than only on text or images: robots, vehicles, drones. | It is the category World Engine sells into. |
| Humanoid | A robot built to the human form, so it can use tools and spaces designed for people. | The end customer's product, and the reason human-shaped data is valuable. |
| Foundation model | A large model trained broadly first, then adapted to many specific tasks. | The buyer of World Engine's data is building one. |
| VLA model | A vision-language-action model: it sees, reads an instruction and outputs movement. | Ψ₀ is a VLA model. |
| Ego-centric data | Video and motion captured from the point of view of a person doing a task. | The high-volume, lower-cost tier of the data pyramid. |
| Teleoperation | A person remotely controlling a robot to perform a task, recording the motion. | The highest-fidelity, highest-cost tier used to align a model. |
| Loco-manipulation | Moving through a space while also manipulating objects with hands. | The hard capability humanoid labs are trying to master. |
| Retargeting | Translating a human's motion into the joint movements of a specific robot body. | Decides whether cheap human video is usable for a robot. |
| Post-training | Teaching a pre-trained model a specific skill with smaller, higher-quality data. | Where the sensorised glove tier is used. |
| Reinforcement learning (RL) | Improving a model by letting it try tasks and learn from outcomes. | Drives demand for data generated at deployment. |
| Sim-to-real | Training in simulation, then transferring to the real world. | Where simulation helps, and where real-world data is still essential. |
| DaaS and IaaS | Data as a Service: selling data. Infrastructure as a Service: selling the system that produces and runs it. | World Engine's first two revenue lines. |
| Run-rate | The latest period's revenue expressed as an annual figure. | The $100M headline is stated as a run-rate. |
| TAM, SAM, SOM | Total, serviceable and obtainable market: how big the opportunity is, how much applies, and how much a company can realistically win. | Used in chapter 9 to size the data layer three ways. |
| Concentration | How much of a company's revenue depends on a small number of customers. | Eight frontier-lab customers, each a material contract. |
Ten chapters on the company, the market, the competition and the business model, with every source listed at the end.
We use cookies to improve your experience and analyse site traffic. By clicking "Accept", you consent to our use of cookies in accordance with our Privacy Policy.