When the building where he grew up faced demolition, the founder of Safe Sign Technologies learned a lesson in perseverance that would later shape his approach to building an AI company — one that Thomson Reuters acquired before it had earned a single dollar in revenue.

Rollins House, a rundown Art Deco former factory in south-east London, was home to his mother and where he was raised. When a developer sought the site for flats, the family faced a legal battle they could not afford. As a teenager, he taught himself planning law between school exams, combed through untouched archives, and assembled a dossier of legal precedents. Standing before a council panel opposite an experienced law firm, he presented the case — and won. The episode instilled a conviction that meticulous research and an willingness to face outsized odds could carry the day.

A decade later, that same mindset was tested again in negotiations with Thomson Reuters, the global information and technology group whose legal, tax, and accounting services make it one of the largest professional-information companies in the world — and which sought to acquire his AI research startup, Safe Sign Technologies.

The path to that point was arduous. While training as a solicitor at Allen & Overy, he built Safe Sign on the side — waking at 4:30 a.m. to work on the company before full days of legal training, then returning to startup work in the evenings. Allen & Overy, he notes, were remarkably tolerant, though the schedule was unsustainable.

On numerous occasions, Safe Sign effectively ran out of money overnight, requiring fresh funding by morning. The company's initial plan — a consumer legal product — failed to attract investors. After repeated rejections from local backers, he flew to New York with £200 to his name. Ultimately, nearly all of the company's capital came from North American investors rather than the UK. He observed a cultural divide in approach: American investors tended to ask "how can I help?" while British investors asked "how will this fail?"

Having raised enough to continue operating, Safe Sign abandoned the pursuit of revenue altogether and redirected its efforts toward building a proprietary AI model. The decision meant investing everything in a small team recruited from Cambridge, MIT, and Harvard, and telling investors repeatedly that meaningful revenue was a distant prospect. Most backers lost interest.

The company's research focused on safety, robustness, and reliability — attributes that carry particular weight in legal applications, where an AI system that fabricates case citations or misreads statute language can cause direct professional harm. Producing a model with differentiated performance on a minimal budget, Safe Sign targeted the gap between general-purpose language models and the reliability threshold that legal practice demands. When Safe Sign shared strong internal results, Thomson Reuters' venture arm responded within minutes.

Twenty months after founding, Safe Sign became Thomson Reuters' first pre-revenue acquisition in the company's 174-year history, ranking among the more significant European deals of 2024. Thomson Reuters, which has been investing heavily in AI across its legal and tax platforms, saw in Safe Sign a research team and model architecture that could strengthen its position as incumbent legal-tech providers race to integrate generative AI into professional workflows. The acquisition sum was life-changing for a company that had never recorded revenue.

The founder notes that this outcome runs counter to much of the conventional wisdom in the startup industry, which emphasizes visible metrics such as rising revenue, capital raised, and founder visibility. In his view, institutional acquirers reward depth of work — substance that becomes evident only under close scrutiny.

Now active as an investor, he describes his primary responsibility as distinguishing between style and genuine value. He identifies a recurring market error — the same one early British investors made with Safe Sign — of judging startups by their pitch decks rather than their core defensibility.

While the market has grown better at valuing science, he argues this applies mainly to high-profile areas: large models, well-known labs, and prominent scientists. The greater opportunity, he contends, lies in the layer beneath — technologies that solve hard problems for frontier AI labs themselves.

He points to several areas where significant challenges remain undercapitalized. One is infrastructure for testing AI system capabilities: every claim about a model passing an exam or outperforming a professional depends on someone having built that assessment, and designing evaluations that cannot be gamed is among the field's hardest problems. Another is memory — today's AI systems begin each conversation from scratch, whereas human expertise is built on accumulated knowledge. The solution is often dismissed as a database add-on to a chatbot, but continual learning remains a fundamentally unsolved machine learning problem. A third area is what he terms "neolab enablement": specialized tools that early AI-for-science labs need but cannot build themselves, requiring the same domain depth as the science they support.

These categories — tooling, storage, and services — are often dismissed by the market as routine infrastructure. He argues they represent some of the harder problems in the field, each quietly impeding the frontier labs, and that almost none of the capital flooding into AI is directed toward them.

The founder draws the comparison back to Rollins House: the unread dossier, the case everyone assumed was lost, the value that exists long before anyone examines it closely. His mother still lives in the building, which stands much as it always has — unremarkable at first glance, yet worth far more than appearances suggest.

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.

This story was originally featured on Fortune.com.