Robotics in 2026 looks like language models around 2020. The technical breakthrough is real. The money arrived before the products.
The breakthrough is the vision-language-action (VLA) model, which reuses the vision-language pretraining behind LLMs to control a robot's motors. Through 2025 and 2026, five labs (Figure, Dyna, Skild, Generalist and Xiaomi) separately reported steady gains from training on more human video rather than robot data. Generalist trains on more than 500,000 hours from wearable cameras, and Dyna on more than a million hours of first-person video. That is why so many pitches have moved from "we collect robot data" to "we record people doing the work."
The capital came faster. Crunchbase counted about $15B of robotics venture funding in 2025 and $18.8B by mid-June 2026; broader definitions reach $41B. F-Prime's analysis of PitchBook data finds that all of the growth came from rounds of $50M or more, while early-stage funding has been flat for five years. Founders from the right lab now raise $100M+ seed rounds: Genesis raised $105M, and Walden raised $300M at a $1.1B valuation. About 13,000 humanoids shipped in 2025, roughly 90% of them Chinese.
The products are behind the money. The best 2026 models succeed 40% to 66% of the time on tasks or homes they have never seen, as the labs themselves report it. Figure's Helix 2.5 hit 56% across 30 unseen homes; Skild's S1 hit 66% on unseen tasks. 99% shows up only on narrow tasks after extra training. A production line needs roughly 99.9%, and at 95% per step a ten-step task succeeds only about 60% of the time. Bessemer puts the field at "GPT-2.5" while NVIDIA says robotics' ChatGPT moment has arrived. Both are right, about different layers.
When the cost of starting a company collapses and the cost of finishing one does not, you get exactly this market: dozens of well-funded teams at the start, very few near the finish. The clearest sample is the 42 labs and startups that 55 voters named on Frontier List's robotics talent-density ranking. I researched every one of them for this piece.
Only two companies on the list have large, disclosed revenue. Unitree, now listed in Shanghai, made about $0.24B in 2025 and was profitable. Skild says it is at a ~$100M annual run-rate. Everyone else is valued on research speed and demos. At $39B, Figure's valuation is already higher than Goldman's 2024 estimate for the entire 2035 humanoid market.
Units tell a different story than valuations. The companies actually shipping humanoids are Chinese, and the most valuable Western ones have built a few hundred.
a16z has published more on this than anyone, and its view has moved in a clear order: a horizontal platform in 2024, geopolitics and the hardware underneath in 2025, deployment realism and a dedicated hardware fund in 2026. Its money followed the writing. a16z is not an investor in Physical Intelligence, Skild, Figure, Genesis or Dyna. It led Atoms' $1.7B round (Kalanick's company, built on specialized robots rather than humanoids), co-led Mind Robotics' $500M Series A (using Rivian's production lines as a data source), and backed XDOF (robot training data), Westmag (US-made motors and actuators), World Labs, and defense autonomy (Anduril, Saronic, Shield AI, Skydio).
Oliver Hsu. Foundation models could create the first shared robotics platform, with many body types rather than one best one.
Ryan McEntush. Minerals, batteries, power electronics, motors and compute form one stack, and China dominates the middle of it.
Martin Casado and Anne Neuberger. Whoever deploys first gets the data to deploy more.
McEntush. EVs, drones and robots share one component design, and whoever controls the modular middle wins.
Hsu. Six gaps separate lab demos from production: distribution shift, reliability, latency, integration, safety and maintenance. Value sits in closing them.
Hsu. Embodied AI repeats the LLM scaling pattern, earlier on the curve. Simulation moves the bottleneck to designing environments.
A $1.1B hardware-only fund covering chips to robots. Hardware is now over 20% of a16z's deal flow.
Read together, a16z's writing and portfolio say the robot brain alone is not a moat. The value they back is deployment data, reliability, middleware, safety, and hardware components that aren't made in China.
Greylock has published almost nothing on robotics since its 2022 autonomy podcasts, and I found no stake in any of the 42 companies. The signal is its one 2026 bet: leading Antioch's $32M Series A in September, for simulation tooling used by robot and drone teams. Partner Saam Motamedi framed it as advantage going to teams that shorten development cycles, not to teams with better models.
My reading of how a Greylock-style lens (network effects, enterprise distribution, compounding advantage) would score this market: horizontal robot brains do not yet have a network effect. There is no developer ecosystem, no two-sided market, and switching costs are low once open models are good enough. The durable position is the toolchain every robot team adopts early and can't easily rip out: simulation, evaluation, data operations and fleet observability.
The playbook is getting consistent. Don't pretrain a generalist model. Fine-tune an open one (π0.5, GR00T N1.7) and let reinforcement learning from real deployments do the rest, as in Physical Intelligence's Recap method behind π*0.6. Pick tasks where 99% is reachable and the buyer pays per outcome. Ship with remote operators behind the robot, and measure the business in human-minutes per task. If you sell in the US, design for the FCC's domestic-content rule from the start.
One model for every body, monetized by licensing and fine-tuning.
Own a paid workflow and its data. Horizontal robot APIs come later.
Everyone agrees data is the bottleneck. They disagree about where it comes from (simulation, recordings of people, or working fleets) and who ends up capturing it.
On 8 October Lightspeed published "Mapping the Robotics Landscape", drawn from Amber Yang's discussion with Claire Zau on Lightwork, Lightspeed's weekly show. It sorts robotics companies on two axes: how general the model is, and how general the robot's body is. It names one category outright: DeployCos, general-purpose hardware running a model tuned for one paid job, with Ultra as the example. The chart marks Lightspeed as an investor in two of its eleven examples, Skild and Dexterity. Here is the same frame with every company from the Frontier List placed in it.
The crowding is on the right. 20 of the 42 build general models: ten as shared brains meant for many bodies, ten inside their own robots. The shared-brain quadrant also holds the big-tech labs giving their models away, which is the pricing pressure the original chart does not draw.
The paid work is on the left. DeployCos and vertical full-stack companies make up only 11 of the 42, but they hold most of the paid deployments: Ultra at Highline Commerce, Bedrock's operator-free excavation, Watney at Meta, Medra's biopharma partners.
The right sells to the left. Ultra's OP1 runs on Physical Intelligence's π0.6, so the top-right quadrant's customers are the top-left quadrant's companies. Whether Ultra pays PI a licence, and how much, is the cleanest early test of whether a horizontal brain can be a business.
A quarter of the list doesn't fit the frame. 11 of the 42 build data, evaluation, simulation and developer tools, sell robot bodies to others, or don't build robots at all. A model-by-body map can't see that layer, and it is where I think the least crowded opportunity sits. China is missing too, though Unitree belongs here as a body supplier.
Put the calls on one stack, from the parts inside a robot up to the paid job it does, and the pattern is plain. The value I'd back sits at the top and the bottom. The crowded middle is where I'd hold back.
The data that stays scarce comes from real deployments: failures, recoveries, and high-precision contact. Only fleets doing paid work produce it. Look for a structured environment, a task where 99% is reachable, a buyer who pays per outcome, and remote-operator costs already counted in the margin. On the list: Bedrock (paid excavation with nobody in the cab since Aug 2026), Dyna (Din Tai Fung expanded after the robots cleared its ROI bar), Watney (data-center work for Meta), Ultra (~30% of one customer's order fulfilment on three units), and Skild's shift to 60+ paying customers.
Every company deploying robots needs intervention analytics, independent benchmarks, safety cases and a remote-operator floor. Almost nobody builds these as products. Robocurve has a $10M seed; RoboArena is academic. It is the least crowded box on the map, and every other box needs it.
Actuators are more than half a humanoid's bill of materials, and China holds ~63% of actuators, motors and reducers. Planetary roller screws are the tightest bottleneck. The FCC rule creates a protected US market for domestic-content parts, which is a16z's Westmag bet. It is a policy moat: real for now, erasable by a policy change.
Open models are now good enough to build on. The closed frontier is about six labs with $0.5B to $3.5B each, and Google and NVIDIA give models away to sell cloud and chips. A new entrant needs a data source nobody else has, not a better architecture.
Labour costs $1 to $25 an hour, 100,000-hour datasets are open, DoorDash can reach 8M Dashers, and Figure's Index programme pays creators directly. This is the commodity end of the data market. It lasts only with curation, action labels and quality control.
The long-term prize is huge, but today these are beta fleets (Sunday is in 50 homes), $20K builds, and remote operators behind every product. The next 24 months carry the wrong risk for the price.
Every robot product shipping in 2026 has people in the loop. 1X's "Expert Mode" hands the robot to a remote operator, Weave has remote specialists finish folds, Watney describes a floor of 2,000+ teleoperators, and OpenAI runs a 400+ person data-collection lab. That is a contact center, and it has the same problems: queueing, routing, quality scoring, handle time, escalation, workforce management. It also has one new job, turning every intervention into training data.
Contact centers took decades to build their software. Robot fleets need it now, and nobody sells it as a horizontal product yet. Much of that operator work will sit in India and Southeast Asia, where Build AI and Vision Lab already collect data. Other gaps: safety-case tooling for learned policies, and integration middleware for existing factories.
The only independent robot-brain company with real revenue (~$100M run-rate, 60+ customers) and a hardware sales channel through Zebra's former Fetch unit. If a horizontal brain becomes a business, it is this one.
Deploys first and trains its own model. Din Tai Fung is expanding on ROI. Watch whether "hundreds of robots by H1 2027" happens.
Paid, operator-free excavation for Sundt and Zachry about 27 months after founding. Construction is structured enough, and short enough of workers, to be an early profit pool.
An independent referee for robot claims: three people and a $10M seed. Evals mattered for LLMs. They will matter more when a failure breaks something physical.
The most aggressive data bet: a million hours of first-person video, a Bengaluru entity, Gurugram factories, a self-reported $100M run-rate. It is also the one to watch for labour and consent risk, which Indian press has reported on.
The strongest argument against all of this: robotics has been here before, and the money ran out before the autonomy arrived.
| Precedent | What happened | What it teaches |
|---|---|---|
| Anki | ~$100M revenue in 2017 and 1.5M+ robots sold. Shut down in 2019 when a late-stage deal fell through. | Revenue isn't survival for low-margin consumer hardware. |
| Pepper / Aldebaran | ~27,000 units sold on a $550/month subscription; only ~15% of customers planned to renew. Receivership in 2025. | Service costs kill subscription robots. |
| Berkshire Grey | Listed via SPAC at $2.7B in 2021. Taken private by SoftBank for $375M in 2023. | Demand wasn't the bottleneck. Delivering and integrating systems was. |
| Cruise | GM spent $10B+. One pedestrian incident in 2023 cost it its permits; GM ended funding in 2024. | A single safety event can end a programme, and "autonomy" can hide a remote workforce. |
| Embark · TuSimple | Listed at $5B+ and $8B+. Embark sold for $71M; TuSimple delisted ~98.5% off its peak. | Pre-revenue physical-AI valuations reprice hard when sentiment turns. |
| Covariant | A robot-brain startup valued at ~$625M. Amazon hired its founders and ~25% of staff and licensed its models (2024). | The closest precedent for how a standalone robot brain exits. |
Revenue multiples are extreme. Agility's $2.5B SPAC rests on $1.8M of 2025 sales. Skild's $14B round priced it at roughly 470x its ~$30M 2025 revenue. Unitree trades at ~125x sales even after falling ~53% from its first-day high, though it is still about 3.4x its IPO price. That is froth cooling, not a failed listing.
The repricing has started. 1X went from targeting $10B+ to reported SoftBank talks at ~$6B. K-Scale Labs shut down in November 2025 after a failed Series A. iRobot filed for bankruptcy, and its Chinese contract manufacturer is taking the equity. China's economic planning agency warned of a bubble across 150+ humanoid firms, and its securities regulator has slowed humanoid IPOs.
"Autonomy" still has a person behind it. Tesla's 2024 Optimus showcase used teleoperated robots. 1X's Expert Mode lets remote operators see inside customers' homes. OpenAI runs 400+ data collectors in three shifts. Those are wages, not software margins.
The scaling curves measure the wrong thing. Most published curves track training loss, not task success. Pooling data from different robots can make models worse. Rodney Brooks argues touch can't be learned from video at all.
The labour maths doesn't travel. A US materials mover earns ~$18/hr, a Chinese factory worker ~$5/hr, an Indian salaried worker about $1 to $1.50/hr. A $20K to $30K robot at ~80% of human productivity can pay back in the US, but not in India for general-purpose work.
Buyers aren't ready. In a Morgan Stanley survey only 23% of Chinese humanoid buyers were satisfied. Gartner expects fewer than 20 companies to deploy humanoids at scale by 2028. Unitree's average selling price fell 72% in two years.
Take the meeting with any team that has a paying fleet and a falling intervention curve, whatever its model.
Pass on the next general-purpose robot brain unless it has a data source nobody else can get.
Place an early, small bet on robot operations and evaluation infrastructure. It is the least crowded box on the map, and every other box has to buy it.
Earlier thesis in this series · Physical AI & Robotics: The India Thesis
Company sample · Frontier List: robotics talent density (42 companies, 55 voters)
a16z · Deployment Gap · Machine Age Fund · Atoms · Westmag
Other funds · Antioch (Greylock) · Bessemer Predicts · CapitalG on PI · Sequoia on Skild · Lightspeed, "Mapping the Robotics Landscape" (Lightwork, 8 Oct 2026)
Market data · Crunchbase · F-Prime · Omdia · BofA Institute · Goldman revision · Gartner · BLS
Lab claims · Figure Helix 2.5 · Figure at BMW · Skild S1 · Dyna DYNA-2 · Generalist GEN-1
Policy · FCC Covered List (Sidley)
Bear case · Unitree after IPO · Agility revenue · China reality check · NDRC warning · Covariant · Rodney Brooks · Scroll on factory data work
Company data researched in late September 2026 from company sites, job boards and press. Model success rates are as reported by each lab, not independently verified. Rounds reported but not confirmed as closed are marked. Quadrant placements, category cuts and all charts are the author's. This is analysis, not investment advice. October 2026.