The house I grew up in was being knocked down. Rollins House, a rundown Art Deco former factory in south-east London, was the kind of building that looks like nothing until you know what it is: unglamorous on the surface, full of history underneath. It was where my mother lived and where I was raised, and a developer wanted the site for flats. We had to fight, but couldn’t afford lawyers.
I decided to teach myself planning law in between school exam revision. I pored over archives nobody else had opened, putting together a dossier of precedents that would help our case.
I stood in front of a council panel, making our case opposite a law firm with more experience than I had years alive. We won.
The experience taught me about the value of detail, how research wins out, and that outsized odds shouldn’t be intimidating.
A decade later, I had to channel that upstart teenager once more, sitting across from Thomson Reuters, which wanted to buy my AI research startup, Safe Sign Technologies.
The journey to get there had been bruising. I had trained as a solicitor at Allen & Overy while building Safe Sign on the side. I was up at 4:30 am to build the company, spent full days on legal training, before returning to startup life after I got home. A&O were remarkably tolerant, but it wasn’t sustainable.
Many nights Safe Sign had effectively run out of money, needing funding again by morning. Our first plan failed; we built a consumer legal product nobody would invest in. Local investors said no over and over again, so I got on a plane to New York with £200 to my name.
In the end, nearly all our capital came from North America, rather than the UK. American investors tended to ask “how can I help?”; British ones asked, “how will this fail?”. We managed to raise enough to survive, before abandoning revenue-chasing entirely to build a proprietary AI model instead. This bet, which involved investing everything in a small team drawn from Cambridge, MIT and Harvard, meant telling investors, repeatedly, that we wouldn’t have meaningful revenue for a long time. Most lost interest.
Our research on safety, robustness and reliability created a model with differentiated performance, on a shoestring budget. When we shared strong internal results, Thomson Reuters’ venture arm replied in minutes. As with Rollins House, conventional wisdom might have told me I had a losing hand; my startup hadn’t yet sold a single subscription. But, again, the substance was there.
Twenty months after founding, we were Thomson Reuters’ first pre-revenue acquisition in 174 years, and one of the more significant European deals of 2024. It was a sum that changed lives, for a company that had never taken a penny in revenue.
I have been turning that fact over ever since, because it contradicts almost everything the industry tells founders to want. We are trained to admire the loud things: revenue climbing on a chart, capital raised, a founder on a stage. Institutions, it turns out, reward none of that. They reward you for doing the work so thoroughly that, when someone finally looks closely, there’s nothing to find but substance.
I invest now and see my main responsibility as parsing between style and value. The mistake I watch the market make, again and again, is the one those early British investors made with me: judging a startup by its pitch deck, not its core defensibility.
Since then, the market has learned to price science better, but only at one level: the big models, the famous labs, the scientists whose names move valuations. But the opportunity in this AI revolution is the layer underneath the one that everyone is talking about: the technologies solving difficult problems for the frontier AI labs themselves.
It might be the infrastructure for testing what AI systems can do. Every claim about a model nailing an exam or outperforming a doctor rests on somebody having built that exam, and once systems learn to beat the test rather than master the job, building exams that can’t be gamed becomes one of the hardest problems in the field. It might be memory: today’s systems start every conversation from zero, when almost everything valuable about human expertise comes from accumulation. The fix gets dismissed as a database bolted onto a chatbot, when continual learning is a fundamentally unsolved machine learning problem. Or it might be neolab enablement, the specialised tools early AI-for-science labs need but can’t build themselves. These are dismissed as a services business when in reality they demand the same domain depth as the science they support.
Style calls these tooling, storage, and services. Substance says they’re three of the harder problems in the field, each one quietly blocking the frontier labs themselves. And almost none of the capital swarming into AI is going in this direction. That gap is where I’d look.
It’s Rollins House again: the dossier nobody had read, the case everyone assumed was lost. It is the zero in pre-revenue: the value that’s real long before anyone looks closely enough.
My mother still lives in the building. It stands defiantly; unremarkable yet worth more than people thought on first glance.
The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.