Winning Strategies in Physical AI

A start-up that builds solar farms with robots won its first four customers without selling labour savings. The contractors it works for had end customers, data centres building their own solar farms, asking for two gigawatts in half the time first promised. No human crew can work that fast. Raj Kapoor of Climactic, whose fund backed the company, told the story at the physical-AI Ripple co-hosted with Sean Petersen of Wellington Management. It made the case for going narrow. What stays defensible afterwards was harder to settle.

The hosts opened with the split in the market. Some companies raise 300–400 million in their first rounds to build a world model of everything, Physical Intelligence among them. Others pick one vertical and nail it. Asked to choose, seven people raised a hand for vertical first and nobody for horizontal, with many undecided. Strategy matters more here than in software, Raj said. A software company can change course with every release. In physical AI there are atoms and processes built into customers' operations, so a change of strategy is expensive. Founders pitching a vertical usually promise to spread out later, and Raj wanted an expansion logic they actually believe.

A painkiller built from commodity parts

The solar company uses a Caterpillar base and a standard robot arm, weatherised. Its value lies in the full-stack software and the data it has collected. The contractors agreed to work with it only because they pay per panel placed, and all four saw a fourfold speed-up. The company raised 5 million and reached 25 million of bookings, Raj said, while horizontal companies have raised hundreds of millions and are not yet at 20 million of revenue. The deadline moved the robot "from vitamin to painkiller". Raj would rather back a smaller market with a burning need than a pitch for a 100-billion-dollar market where the need is unclear.

The founders at the table had drawn their own narrow lines. One makes high-precision metal parts for aviation, semiconductors and energy grids in lots of about a hundred, not ten thousand, and still sees a 150-billion industry. Another founder's company keeps wildlife away with a device that has onboard AI and a speaker to talk to the animals. It had to build its own hardware because no such trail camera existed, and the founder's rule was to build hardware only when the problem cannot be solved without it.

One of the hosts sees the opposite choice spreading in Silicon Valley. Many start-ups stick with commodity hardware so they can raise debt to scale and avoid supply-chain and safety trouble. Raj added that more of Climactic's companies now sell robots as a service, so the customer worries about uptime and not equipment. Elon Musk made the same kind of choice by forcing engineers to rely on cameras for self-driving and drop lidar.

being the throat to choke could be the new moat.

Layers in between

A co-founder of a firm that buys legacy service businesses and brings robots in said value will gather in the deployment layer. One host saw the appeal in certified niches such as aircraft maintenance, where a bought business comes with an approval and a customer book that are hard to win. The other asked what stops an incumbent that can see the same return from doing it alone. Middleware drew more doubt. An early-stage investor had backed an orchestration layer so that robots never crash or hurt anyone, but another investor never invests in such companies, because their defensibility does not survive the investment committee. Raj, who ran self-driving at Lyft, recalled that the winners there built everything themselves, and the one supplier that broke through, Scale AI in data labelling, did so much later.

After the commodities

One of the hosts asked what gets commoditised, since a superintelligence might one day assemble a robot for any job from a rack of parts. Hardware already is, one investor said, and a pure hardware play is very hard for a European VC to back. A hardware founder pointed to Lovable. People call it moat-less every week, the founder said, yet it makes billions. Get going, earn trust and build a brand. Another participant cited Gecko Robotics, which uses largely off-the-shelf hardware to inspect infrastructure but has gone so deep into its customers' data that it now informs whether they build new assets or maintain old ones.

That pointed to the customer as the moat. Customers will not find anyone at OpenAI or Anthropic who understands their business, one of the hosts said, and they will be overwhelmed by the pace of change, so "being the throat to choke could be the new moat." A pre-seed investor took the point to the factory floor. Last year every pre-seed fund was pitched an operating system for manufacturing, and such companies wait a long time for data from their corporate customers. The investor's latest bet buys small injection-moulding firms outright and rebuilds them step by step, because changing a process needs ownership.

Raj's last provocation reopened the first question. If the next humanoids can do a million jobs, four of them might build a solar farm as fast as a specialised machine, at a lower marginal cost. One of the hosts answered on behalf of the robotics professors they know. The human form is not the ideal shape for most jobs, they said. It just happens to be ours.

This Ripple was hosted by Raj Kapoor (Climactic) and Sean Petersen (Wellington Management) at The Drop 2026 on 16 September.