Heard in the Circle: What Physical AI Builders Argued About This Quarter

The founders, engineers and investors in my circle build and back Deep Tech & Physical AI, and most of what I learn comes from arguing with them. Between July and September the arguments kept returning to the same questions: whether robots have enough data, which kind of model has the higher ceiling, what the public filings say about humanoid economics, where the reliability bar sits, who certifies a robot as safe, and whether the humanoid is the right shape at all. This is a write-up of what I heard here and there, with names left out. Nothing below is attributed to a person or a company.

Data: enough already, or nowhere near

The quarter's data argument started with two public results. On August 10, Dyna Robotics released Dyna-2, a world-action model pretrained on a million hours of first-person human video, with a scaling ladder behind it: across 14 robot tasks, mean normalized performance rose from 20 percent of the attainable maximum at 1,000 hours to 28 percent at 10,000, 45 percent at 100,000 and 53 percent at a million, and the company said it will explore ten million hours if the curve holds. On September 17, Figure reported that pretraining on its Index dataset of paid first-person video lifted its humanoid's zero-shot success across 30 unseen Bay Area homes from 9 percent to 56 percent, with everything else held fixed.

The reactions split three ways. One camp read Figure's result as the end of the data debate: there is clearly enough data already, and the remaining work is modeling. The reply came from the driving analogy: Waymo needed millions of hours to solve one task in one kind of environment, and 56 percent in a living room is a research number, not a shift. An investor who backs construction companies added the part both camps skip, which is that most construction work has never been recorded at all, and "the best machine ever created is still a human."

The scale question came back in October with a conversion. One founder mapped Dyna-2's million hours onto the language-model timeline and put it somewhere around GPT-2.5, with a GPT-5-equivalent near two billion hours, which at tenfold growth a year lands around 2030. Another pointed out what the hours contain: first-person video, with no gripper data, so the conversion flatters the robot side.

Generalist's one-shot result in August, a new task learned from a single demonstration, drew the shortest responses. "Confirmation that scale is scaling." One founder compared it with in-context learning in language models and with voice cloning, which went from hours of audio to thirty seconds in three years. Another said several labs are taking the same "bitter-lesson-pilled" approach and the result was expected by the people building toward it.

Models: which ceiling is higher

A debate at a Saturday Robotics session in Beijing produced three positions on world-action models against vision-language-action models, and the people I put them to split the same three ways. World-action models have the higher ceiling because they learn from unlabeled video. The ceilings are the same and world-action models only start higher. Or the label matters less than how much physics a model extracts from whatever it is fed. Nobody changed sides, which is itself the finding: the architecture argument is about which base model you start from and how you spend a fixed data and compute budget.

Economics: what the filings say

Unitree listed on Shanghai's STAR Market on August 19 at 150.8 yuan a share, opened up 629 percent and closed up 460 percent, per AFP. An investor who read the prospectus pulled out three things the price did not: steep erosion in unit prices, a great deal of manual assembly, and vertical integration of components as the one durable moat. Another drew the line on who buys: the customers are predominantly research labs, often government-linked, while industry "is still dominated by arms, cobots, and AGVs."

Agility Robotics' S-4, filed for its SPAC merger, reported $1.8 million of net sales in 2025 against a $140 million operating loss, per The Robot Report, and the filing lists Digit at $8,500 a month on a robots-as-a-service contract. A roboticist did the division: $1.8 million is about 8 robots sold outright or 17 on subscription for a year, and at full uptime the subscription works out to $11 to $12 an hour.

The longest argument of the quarter started with an investor's question: has anyone seen a Skild robot in production? Skild said in September that it had crossed a $100 million revenue run rate with more than 60 customers and "hundreds of robots" deployed. One investor had passed on the last round at 140 times revenue. The arithmetic followed: hundreds of robots against $100 million implies more than $50,000 per robot per year, which pays off against $30,000 to $70,000 a year in wages per shift only when the robot runs two or three shifts. Nobody had seen one running.

Reliability: 99 against 97

The question of why Cognex and Keyence still own industrial vision, after a decade of deep learning, got the most useful answer of the quarter from an engineer who built a package-sorting cell that handled 3,500 packages an hour. A rejected package costs about seven times as much to fix by hand as a package the cell handled, so a cell that is right 85 percent of the time saves no labor at all, and the difference between 99 and 97 percent accuracy is about two years of payback. A startup selling into the same customers says it is taking share from both incumbents, on speed of deployment rather than on accuracy.

Hands got the same treatment. When 1X showed new tendon-driven hands in July, one investor said tendons are harder to produce and to quality-check, and another read the video of a technician re-tensioning them as rework. In September a roboticist put a number on it: no tendon-driven hand that anyone could name has survived two million cycles, and "a million seconds is 12 days."

Safety: demonstrated against assured

San Mateo County's supervisors voted 4-0 on September 15 to require permits for businesses in the county's unincorporated areas that run autonomous, battery-powered robots around the public or workers, with manufacturer and safety-record disclosures and a workforce plan for displaced staff. A roboticist expects the inspectors to ask for UL certificates, which almost no learned-control robot has. An engineer who has shipped robots into warehouses separated two things the industry blurs: demonstrated safety, which is miles without incident, and assured safety, which is a formal hazard analysis validated by those miles. Waymo did the second. "There is a chasm," the engineer wrote, between matching Waymo's miles and being safe like Waymo.

The argument continued in October at a smaller scale: code. A robotics engineer reported pull requests that had passed an AI-assisted review and merged with basic safety errors. One reply was that test harnesses will soon be obsolete; the answer was "Safety != Harness." The practical proposal was to reserve senior human review for the parts of the stack that arbitrate safety and let the assistants handle the rest, on the argument that "the moat in a post-AI world is reliability and quality past the 99 percent threshold."

Form factors: crabification

The humanoid argument ran all summer and produced the quarter's best word. After a humanoid company launched at a $1.1 billion valuation in July, a founder predicted "a Darwinian battle of the form factors" ending in "robocrabs," and crabification became the running joke for any machine that starts human-shaped and evolves toward the task. The serious version came in August: raise as a humanoid company, then build "literally any other form factor." An engineer offered a support group for anti-humanoid practitioners. When a video showed a humanoid climbing out of a car, the fine print said "jointly controlled by a human operator and AI," which I filed under teleoperation plus.

A few things to remember

Data has become a portfolio question. The people I talk to hold both results at once, Figure's 56 percent and Waymo's millions of hours, and the sharpest of them ask which source, in which order, against which budget.

The filings are the best data on humanoid economics. Unitree's prospectus and Agility's S-4 gave us more to work with than any launch video, and the arithmetic on both points the same way: research buyers, thin unit economics, and subscription prices that only clear against multiple shifts.

Reliability is priced in years of payback. The 99-against-97 arithmetic explains the incumbents better than any model comparison, and tendon cycle life explains why hands ship slowly.

Safety is moving from demonstrated to assured. San Mateo's permits, UL questions and formal hazard analysis are arriving before most learned-control robots can answer them.

References

• Helix 2.5: Zero-Shot 30-Home Generalization - Figure AI, September 2026. figure.ai
• Dyna-2: A 1-Million-Hour Scaling Law for World-Action Models - Dyna Robotics, August 2026. dyna.co
• Unitree surges more than 600% on Shanghai debut - AFP via Kuwait Times, August 2026. kuwaittimes.com
• Agility Robotics reports $1.8M revenue ahead of humanoid SPAC - The Robot Report, September 2026. therobotreport.com
• Skild AI hits $100m revenue run-rate - Silicon UK, September 2026. silicon.co.uk
• San Mateo County Supervisors pass new public safety, worker protections for autonomous robots - County of San Mateo, September 2026. smcgov.org

~Bogdan