The Humanoid Robot Industry Shipped 13,000 Units Last Year. It Promised 100,000 This Year.

Seven. That's how many humanoid robots Toyota Motor Manufacturing Canada will deploy at its Woodstock, Ontario plant in April. After a full year of piloting three Agility Digit units loading totes onto automated tuggers, TMMC signed a Robots-as-a-Service contract for four more. Toyota's largest plant outside Japan, builder of 11 million vehicles since 1988, evaluated the state of humanoid robotics and bought seven.

Meanwhile, Goldman Sachs projects 50,000 to 100,000 humanoid units shipping globally in 2026. Tesla targets a million per year by late 2026. Analyst decks from McKinsey, Bain, and every venture firm with a robotics portfolio describe a market poised to explode.

Something about these two realities doesn't fit.

What Actually Shipped

Omdia's General-Purpose Embodied Intelligent Robot 2026 report, released in January, counted approximately 13,000 humanoid robots shipped worldwide in 2025. That's the total. Not per quarter. Not per company. Thirteen thousand units across every manufacturer, every country, every application.

Shanghai-based AgiBot led with 5,100 units and 39% market share. Unitree followed with roughly 4,160 units at 32%. Together they account for 71% of global shipments. Leju Robotics shipped 4%. Fourier Intelligence, 2%. Agility Robotics, the only Western manufacturer with paying factory customers, captured 1%.

One percent of 13,000 is about 130 robots. Agility's public announcements account for units at Amazon, GXO, Spanx, Schaeffler, and now Toyota. Seven here, a handful there.

Chinese manufacturers dominate volume because most of their units aren't doing factory work. AgiBot's 5,100 robots span hospitality, entertainment, security, and research. Unitree sells the G1 for $13,000 to $16,000, cheap enough for university labs and corporate lobbies. Walk into enough hotel lobbies in Shenzhen and you'll find a humanoid greeting guests. Count those toward the 100,000-unit 2026 target if you want. Most analysts do.

What "Deployed" Means

Language matters. "Deployed" in a press release covers a range of situations. A robot running 30-minute shifts before a 9-minute recharge at Amazon is deployed. A robot in a trade show booth is deployed. A thousand Optimus units doing kitting tasks inside Tesla's own factory, reported by Tesla's own communications team, are deployed. Here's what the major players actually claim.

Tesla Optimus: Elon Musk stated that over 1,000 Gen 3 units are operating at Fremont and Giga Texas as of January 2026. Parts kitting, material handling. These numbers come from Tesla itself and have not been independently verified. Zero units sold to external customers. Tesla is both the manufacturer and the only buyer. Musk targets 1 million units annually by late 2026, with a dedicated factory under construction at Giga Texas. Getting from 1,000 internal units to 1 million annual production requires a roughly 1,000x ramp in 12 months. For context: Tesla took three years to ramp Model 3 production from initial units to 350,000 per year, and that was a car it already knew how to build.

Agility Digit: Seven units at Toyota Canada, expanding from a three-unit pilot. Additional units at Amazon, GXO, Spanx, Schaeffler. Factory capacity at RoboFab in Salem, Oregon: 10,000 units per year. Actual shipped units in 2025: roughly 130. Factory utilization: about 1.3%.

Figure AI: Figure 02 robots ran 10-hour shifts five days a week at BMW Spartanburg over 11 months, loading 90,000+ parts and contributing to 30,000+ X3 assemblies. Total logged runtime: 1,250 hours. Figure claims 400% efficiency gains in complex assembly tasks with 5mm tolerance at 37-second cycle times. BMW Leipzig is starting a second pilot for European production. Figure raised $675 million at a $2.6 billion valuation. It has not publicly disclosed unit sales or how many robots operated at Spartanburg simultaneously.

Boston Dynamics Atlas: Electric Atlas production underway. Fleet deployments to Hyundai RMAC and Google DeepMind pilots. Partner-only access. No retail sales. No unit numbers disclosed.

Chinese manufacturers: AgiBot and Unitree ship volume, primarily into hospitality, entertainment, and education. BYD targets 1,500 units in 2025 ramping to 20,000 in 2026. UBTECH has a $1 billion financing facility and government contracts including border and public service deployments. Real units. But not the factory floor revolution the projections describe.

90 Minutes Between Charges

Battery life explains much of the gap. IEEE Spectrum reported that Agility Digit operates in 30-minute intervals at Amazon, with 90 minutes maximum between charges. Figure 02 manages 2 to 3 hours. Sanctuary AI's Phoenix demonstrated 43.5 hours of cumulative operation at Hannover Messe 2025, with charging breaks.

A factory shift runs 8 hours. Even the best humanoid battery covers a fraction of that. Bipedal walking burns enormous energy just maintaining balance. Lithium-ion energy density improves 5 to 8% annually. Solid-state batteries might double that density by 2028 or 2029. Neither timeline helps with the 2026 projections.

Sanctuary AI addressed this directly: they switched Phoenix to a wheeled base. Wheels are roughly 10x more energy-efficient than legs for locomotion. If a humanoid robot needs to do useful factory work for 8 hours, the most honest engineering solution is to stop pretending it needs to walk like a human.

But a robot on wheels isn't a humanoid. It's a mobile manipulator, a category that already exists and ships in the tens of thousands. Humanoid form factor is the product being sold, even when it's the engineering constraint holding performance back.

60% Task Completion

Autonomy is the other constraint nobody discusses at conferences. Current humanoid robots complete 60 to 80% of assigned tasks in real-world factory scenarios. Industrial customers require 95% or higher for unsupervised operation. One widely reported case started at 46% and improved to 78% after training on over a million robot trajectories across 217 distinct tasks.

78% sounds like progress until you run the math. A robot failing 22% of tasks on an assembly line doing 500 operations per shift generates 110 interventions per shift. At that rate, you need almost as many human minders as you'd need humans doing the work. A 95% completion rate still produces 25 failures. Even 99% yields 5 per shift that stop the line. Industrial reliability isn't "pretty good." It's near-perfect or it's expensive.

Toyota's year-long pilot ending in a 7-unit purchase makes sense in this context. They're not buying robots to replace a production line. They're buying them to load totes onto tuggers, a single repetitive task with high tolerance for error. If a tote gets dropped, nothing breaks. If a tote gets missed, a human puts it on the next run. The task was selected because it's boring, physical, and forgiving.

What the Projections Actually Project

Goldman Sachs forecasts 2.6 million humanoid units shipped annually by 2035. That projection assumes roughly 80% compound annual growth from the 2025 base of 13,000. Decade-long exponential growth curves in hardware have precedent. Smartphones went from niche to a billion units per year in about eight years. Industrial robots took 30 years to reach 500,000 annual installations.

Humanoid robots are probably closer to the industrial robot trajectory than the smartphone one. Smartphones rode an existing cellular infrastructure. Every humanoid robot deployment requires custom integration, charging infrastructure, safety fencing or co-bot certification, staff training, and often physical facility modification. Every factory is different. Every task requires training data.

Getting from 13,000 to 100,000 units in one calendar year requires 7.7x growth. Nothing in the humanoid industry's history comes close. Industrial robot installations peaked at about 30% year-over-year growth during the post-pandemic manufacturing automation surge in 2021. Automotive factories, the primary target for humanoid deployment, are running 85 to 90% capacity utilization. No slack exists in the schedule for extended robot pilot programs.

Who Benefits from the Gap

AgiBot's 39% market share supports its next funding round. Tesla's 1,000-unit internal deployment supports a stock price that now incorporates humanoid robotics as a future revenue line. Goldman's 2035 projection supports investment banking fees from robotics IPOs. Every party in this ecosystem has a financial interest in the larger number.

Not conspiracy. Incentive structure. Companies report what raises the next round. Analysts model what sells the research note. Both are choosing which numbers to spotlight, and the numbers that move capital aren't the numbers that describe factory floors.

Limitations

Several qualifications. Omdia's 13,000-unit count may undercount Chinese shipments; smaller manufacturers don't always report to Western research firms. Tesla's internal deployment figures have not been independently audited. Battery runtimes reflect 2025 hardware; newer platforms shipping later this year may improve materially. Some tasks genuinely suit humanoid form factors over wheeled alternatives, particularly stair climbing and accessing human-designed spaces. Both smartphone and Segway analogies apply here: nobody predicted the iPhone's adoption curve in 2006, and nobody predicted the Segway's irrelevance.

None of which changes the basic arithmetic. Multiplying last year's actual shipments by 7.7 requires something that isn't yet visible in order books, factory capacity, or customer contracts. The industry's most advanced external deployment just went from 3 robots to 7.

Where This Leaves the Bet

Humanoid robotics is real. Money is real. Hardware improvements are real. AgiBot shipped 5,100 units and somebody bought them. Agility's Digit handles totes at Toyota 30 minutes at a stretch. Figure's robots logged 1,250 hours on BMW's production line. Each of these is progress.

Progress is 7 robots at Woodstock, 30 minutes at a stretch, loading totes. Projections are 100,000 units worldwide, full factory shifts, general-purpose autonomy. Between those numbers sits unsolved battery chemistry, an AI generalization challenge no one can put on a Gantt chart, and a lot of investor capital that needs the projections to be true long before the engineering catches up.

Toyota spent a year watching three robots load totes. Then it bought four more. When billions are on the line, conviction moves in single digits.