糖心视频

By

I sold my AI research company while I was qualifying as a lawyer in the U.K. I built with researchers from , , and who believed in the mission enough to trust a 21-year-old law student to lead the ship.

, when acquired us, it was the first time in its 170-year history it bought a company pre-revenue. Thomson Reuters acquired us for the science.

Getting there was painful, though. Our published papers put the model among the best in the world at legal reasoning, and we trained it for a fraction of what the large labs were spending. We had been a quieter version of 鈥渢he story,鈥� developing very capable models using novel algorithms with huge capital efficiency.

None of that counted for much in the rooms I walked into. Investors always asked about the product and the traction. U.K. investors passed, and I ended up raising most of our funding in the United States.

I back founders now, and the things I weigh have stayed consistent. As a founder, I was told again and again that science meant little until it was bolted onto a product. That test was wrong then and I believe it is fatal now.

Backing founders in the foundational layer

In the first quarter of 2026, foundational AI startups raised around $178 billion. The market is realizing that foundational AI is where the long-term value sits, and this is the year we may see the exits and IPOs that prove the bet right.

However, the capital and the conviction have also pooled around a few names that were already incumbent. , and took roughly 97% of it, and every other foundational AI company in the world shared what was left.

For a deep-tech founder starting out now, that might push them toward a tempting but dangerous read of the market: that the race is over, and that the sensible move would be to build on top of one of these giants.

I鈥檓 looking for founders who move the other way.

Most application-layer companies, built on a model they do not own, adapt to the pricing and access decided for them by the firms upstream, and compete in categories that the same firm can absorb whenever it chooses.

The more durable place to build is the layer underneath. The cost, speed, reliability, interpretability and safety of AI systems remain unsolved and genuine scientific challenges, and they decide what everything that sits above them can do.

A real advance in training efficiency, model architecture and inference cost is the work that will still matter in five years, long after most of today鈥檚 wrappers have been priced out or absorbed.

Asking the right questions

So the questions I ask AI founders are:

• Is your technical team, scientist-for-scientist, equal to or better than the team at DeepMind?
• Does the problem sit at the level of the model and the system, or is it one more thing stacked on someone else鈥檚?
• Will the 鈥減roduct鈥� get harder to live without over the next five years, or is it looking to reach for early revenue like every other startup?

Some of the companies that ended up mattering most in the AI era are those that survived this line of thought. DeepMind and OpenAI began as research efforts with no obvious product, and both would have looked uncomfortable to a conventional early-stage software investor. Their importance is obvious in hindsight, but the foundational problem-solvers tend to look unfundable right up until it looks inevitable.

Do not build to look fundable this quarter. Build something that the whole stack will depend on in the future. Hire the best team you can find to do it, and solve the hard, foundational problem while it is still unfashionable.

The deep-tech capital market is slow, and it will keep chasing familiar names for a while yet. The work still comes first, and the founders who dare to do it early are the ones the market eventually has to come and find.

The future lies in deep tech, not in the surface wrappers that pass for most products nowadays.

was the founder and CEO of , which was acquired by in 2024. He is an angel investor and senior director at Thomson Reuters Labs.

Illustration: