Thesis · Tom Trounce · November 2026
Frontier AI is scaling faster than the businesses built to sell it.
For three years I sat between the model labs and the enterprises buying from them. In enterprise software, capability was rarely the constraint. The models improved every quarter, whether or not anyone was ready. What ran out was everyone else’s ability to change at the same speed: the services firms, the clients and the people doing the work. I believe that gap is the most important commercial fact in AI today. I think a version of it is coming to robotics, though the constraints there are different.
AI services: the squeeze on undifferentiated work
Demand for AI services is still growing. The margin in it is not. Most firms sold the same models through the same clouds, with almost the same accelerator decks. When every firm reads the same, the client buys on price.
The pressure comes from several directions. Adoption trails the purchase: enterprises bought seats in their thousands, then found the usage sat in a handful of teams. Compute and power are now real constraints, and they show up in the price of every workload. The agents went first for the work people enjoy, the writing, analysis and design that made the job worth having, so the resistance inside clients is real and mostly rational. And the pace leaves buyers and sellers alike feeling behind.
The economics of building have broken too. Clients now expect code to be written quickly and cheaply on LLMs. A twelve-week quote is hard to defend when their own engineer had a prototype running by Friday. So far the margin has gone to the chips, the compute and the labs. Whether it moves back up the stack is an open question. Very few services firms have a clear answer for where that leaves them.
My view is that the firms that come through this will stop reselling one partner’s capability and work across the whole ecosystem:
- the hyperscalers and hosted frontier models;
- open-weight and sovereign deployments for clients who can’t let their data leave;
- small fine-tuned models wherever they beat the frontier on cost;
- reinforcement-learning environments built from the client’s own work.
I expect them to sell governance as an ongoing service, measure themselves on how much capability they leave behind in the client, and put engineers forward into the client’s operation, because that’s where the next contract gets written.
Robotics: different constraints, the same commercial pattern
Robots already do paid work at scale: in warehouses, on production lines, in the air and under the sea. What’s still some way off is general-purpose robotics, where one machine learns many tasks. There, capability really is the constraint: dexterity, reliability across a full shift, and enough data to train on. I don’t think robotics will follow LLMs technically. I do think it will follow them commercially. The same four forces will decide the winners: distinctiveness, adoption, who controls distribution, and staying neutral across the ecosystem.
Universal brains. I believe foundation models that can run many different robot bodies will put pressure on proprietary robot software, much as frontier LLMs did to bespoke language models. The opposite bet is credible: vertically integrated companies are wagering that designing hardware and software together wins, as Apple did. I lean towards the open model for most industrial uses, because customers won’t want their operations tied to a single maker.
Learning on the job. I expect the lasting advantage to come from robots that keep training at the customer’s site, on that customer’s tasks. That raises the question I think the industry has barely started to answer: when a robot learns a skill on site, who owns it? The customer, the robot maker and the model company all have a claim, and whoever wins that argument will shape the economics of the industry.
The work and the resistance. Agents took work people wanted to keep. Robots will mostly take work people can’t or won’t do: heavy lifting, precision at microscopic scale, hazardous environments, and work across land, air and sea. That changes the shape of the resistance, but it doesn’t remove it. Work nobody wants is still work someone is paid for. I expect the hard conversations to be with unions, regulators and insurers rather than with users. They can be had, but only by people prepared to have them early.
Edge compute and mixed fleets. Compute will have to sit at the edge, because a robot can’t wait on a data centre or stop when the signal drops. Customers will run fleets from several makers, and I believe whoever coordinates them becomes the layer everyone else plugs into. The robot makers will resist that, and the software that already runs plants and warehouses may get there first.
Two questions remain open. Who will own distribution: the model companies, the robot makers, the integrators or the orchestration platforms? And what part will the frontier LLMs play? They might become the planning and language layer above motor control, or robotics firms might decide to build that layer themselves. In services, the answer to the distribution question decided who kept their margins. My bet is that it decides the same thing here.
Revenue an investor will back
I believe the robotics companies that raise their next round on good terms will be the ones with recurring revenue an investor trusts, not the ones with the best demo. Right now most can’t show it. Paid pilots don’t recur. Hardware sales arrive in lumps, at thin margins. Robots-as-a-service looks like ARR on a slide. But if it’s really the customer renting hardware the company had to finance, a careful investor will discount it, and the cost stays on the company’s balance sheet.
The revenue I’d want to build, and that I’d expect investors to pay more for, comes from the software and the work:
- fleet and orchestration subscriptions, priced per robot;
- charges tied to output: per pick, per inspection, per hour of uptime;
- licences for skills a robot has learned;
- service contracts that guarantee availability.
Each of these recurs, and each grows when the customer adds robots or sites.
That makes expansion the number that matters most. A first site proves the robot works. The second and third prove the company can deploy without heroics, and that is what drives net revenue retention. I’d watch four things closely:
- the time from signed contract to a robot doing productive work;
- what each new site costs to integrate, and whether that cost falls;
- how many sites a customer runs a year after signing;
- gross margin once field service is counted.
If those move the right way, the ARR follows. If they don’t, no pricing model will hide it.
This is where the services lesson carries over most directly. The 2,500-to-30,000 seat expansion I led was net revenue retention in practice: one team proved the value, the next team asked for it, and procurement followed. The partnership practice I built grew because partners carried the work into accounts we could never have reached alone. I believe robotics companies need both: expansion inside each customer, and integrators and channel partners who let them add sites faster than their own field teams could.
Where I come in
I’m not a roboticist. What I’ve done is structure the partnerships and distribution between model makers, delivery firms and enterprises while the terms were still being set. I built Fractal’s OpenAI and Anthropic practice from nothing to more than $15m a year, and that was the work. I’d like to do it again with a services firm deciding how to get into robotics before the partner positions are taken, with a robotics company building the recurring revenue its next round depends on, or with an investor deciding which of those companies will scale.
If any of this matches what you’re seeing, or you think I’ve got it wrong, I’d like to hear it.