In April 1956, Malcom McLean’s Ideal X sailed from Newark to Houston with 58 shipping containers on deck. At the time, loading loose cargo by hand cost about $5.83 a ton. McLean’s experts put the cost of loading the Ideal X at under 16 cents. The container didn’t make ships any smarter. It fixed the shape of everything around the cargo, and global trade reorganised itself around that standard.
What it never standardised was the inside. Seventy years on, the cartons stacked loose in many import containers are still unloaded by hand, one at a time, because no two loads are the same. That combination of a fixed environment and work that keeps changing is where physical AI is finding its first paying customers and, I think, where the lasting value will sit. The market has never been louder. “Every industrial company will become a robotics company,” Jensen Huang told NVIDIA’s GTC conference in March, where I was in the audience. Investors appear to agree.
Robotics and physical AI startups raised $33.4 billion in the first half of 2026, more than in all of 2025 ($28.7 billion), according to PitchBook’s latest robotics report. That capital is unusually concentrated. Three rounds, for Anduril, NEURA Robotics, and Physical Intelligence, made up 40% of the second quarter. The last of those valued a general-purpose robot “brain” at $11.7 billion, roughly double its late-2025 level, and Bloomberg reports it is now in talks to raise again at around $20 billion. Logistics and warehousing took around 6% of the half-year, and its funding fell 68% from the first quarter to the second. Yet established automation businesses such as Symbotic and Rockwell Automation offered some of the strongest evidence of customer demand in the sector. As the report puts it, “funding rankings are not the same thing as customer demand rankings.” Capital is chasing the biggest platform bets, while customers pay for automation that works in a bounded space.
Humanoids show that gap between capital and demand most clearly. Figure AI alone was valued at $39 billion last year. The premise is intuitive: the world is built for people, so build robots shaped like people. That argument is strongest in the home, where nobody will redesign a kitchen around a robot, and the evidence there is thin. 1X says its first year of home robots sold out in five days, but it has not said how many have reached customers’ homes, and early chores lean on remote human operators. In industry, a few paid deployments exist, such as Agility Robotics’ Digit at Toyota, but PitchBook found humanoid makers rarely disclose fleet sizes, uptime, or cost per task. Factories, meanwhile, installed a record 603,307 conventional industrial robots in 2025, nearly all of them robot arms, according to the International Federation of Robotics. In commercial settings the environment can be designed around the task, which makes the human form a premium you rarely need. That leaves the humanoid with its best argument in its hardest market.
The environment matters because of data. Language models learned from an internet’s worth of text that already existed. Physical work has no equivalent. Robot data has to be generated, through teleoperation, simulation, human video, or robots doing real jobs. UC Berkeley’s Ken Goldberg calls this the “100,000-year data gap”, after the time it would take a person to read the text language models are trained on. That gap was the thread running through the robotics rooms at SF Tech Week in October, where teams argued over which of those sources to bet on. At a showcase hosted by SAP and Plug and Play, the panel’s answer to how you get from impressive demos to reliable systems was simple: get hardware into the field and collect the data. Robots doing paid work generate the data that improves the next model, and the customer funds it. The industry shorthand has become “deployment is the moat”.
It’s close, but not quite right. PitchBook’s latest note on robot data points out that more hours of footage don’t automatically make a robot better. As the roboticist Chris Paxton (Agility Robotics) has observed, a deployment narrow enough to be reliable can stop producing anything new to learn. A robot that does the same thing perfectly all day is a good product but a poor teacher. The moat is an environment fixed enough to deploy in reliably today, with work varied enough that every shift throws up something new, and a system that turns those surprises into a better model. A trailer is always a trailer. No two loads are alike.
None of this works if the operation can’t be captured as data. Bessemer’s panel on AI for the factory floor at LA Tech Week started from the premise that most shop-floor deployments stall for reasons that have nothing to do with the model: fragmented data, decades-old equipment, and know-how that has never been written down. Either the robot captures the work itself, or someone has to digitise the operation first.
Much remains open. First, timing. I expect general-purpose robot brains to become a layer others build on. The question is when. Physical Intelligence, which has been releasing open weights since early 2025, launched a developer API in October 2026, while Bessemer sees them as a 2028-and-beyond story. At a Magnus Robotics session I attended in San Francisco, the view was that they are around five years from scaling, and that startups are better off starting in one vertical. Second, the data recipe: nobody yet knows the right mix of teleoperation, human video, simulation, and deployment data, or how much is enough. Third, reliability. Production lines typically need better than 99.9%, and, as a16z has pointed out, a robot that succeeds 95% of the time in the lab can drop to 60% in deployment. Safety standards were written for programmed machines, not learned ones. Fourth, cost. Each robot needs its own real-time compute, but the bigger bill is for surprises, since every one a robot can’t handle needs a remote operator or a service visit. The loop only pays if it brings that bill down.
Two recurring questions came up in the Tech Week rooms. Who owns the footage a robot records inside a customer’s facility? And when a learned system fails, who is responsible: the hardware maker, the model provider, or the operator? The first matters less to the moat than it sounds. As the same note on robot data argues, a vendor can sell the loop that turns field failures into better models even when the customer keeps every raw recording. The real risk is contracts that stop what one customer’s robots learn from improving the next customer’s. The second has no settled answer, and until it does, I’d expect cautious buyers to favour vendors who take responsibility for the whole system. Then there is China, where around 90% of humanoids sold last year were made. Who can build at scale is as open a question as who can build the best brain.
I think physical AI will follow the path we at Quadri described for language models in The Commoditisation of Intelligence. The robot brain will commoditise, perhaps more slowly, as open models and APIs spread. The obvious objection is that general models trained on more data tend to beat specialists, and that scarce robot data favours whoever pools the most. I agree, which is why the brain will be rented rather than built. Several labs will make good ones, and competition will keep them cheap. What no general model can get is what sits inside a specific operation: the installed fleet, a place in the customer’s daily routine, a steady supply of new situations, and the loop that turns them into a better model. Even the companies building general-purpose brains seem to agree. When Skild AI bought Zebra’s Fetch Robotics business this year, it was buying installed fleets, enterprise workflows, and the operating data that comes with them. Models will commoditise. Environments won’t.
At Quadri, this is where we have concentrated our physical AI investing: companies that own a hard, specific environment, and the experience it produces.
Pickle Robot works inside the box. Its robots unload floor-loaded trucks and containers, a punishing, high-turnover job in an industry where labour makes up 50% to 70% of warehouse operating costs. Customers, including UPS, can be live in under five days, and Pickle says most deployments pay back within 18 months. No two trailers are loaded the same way, so the fleet keeps learning from loads it hasn’t seen before.
Moonware, one of our most recent investments, takes the other route on the airport ramp: digitise the operation first. The ramp is a bounded space where the schedule changes by the minute, yet much of a turnaround is still coordinated by radio. Moonware’s HALO platform coordinates turnarounds in real time, assigning tasks and dispatching crews and equipment. Before baggage tractors and pushback tugs can drive themselves at scale, someone has to orchestrate the ramp. That is the layer Moonware is building.
For both, a better robot brain is a tailwind. What they own is the inside of the box.