Basecamp Research raises $140M to train AI on evolutionary data

What happened

London-based Basecamp Research has raised $140 million from investors including Nvidia and Anthropic's Anthology Fund. The company trains AI models on genetic material collected from rainforests, oceans and hot springs, with the aim of designing antibiotics and tools for cell therapies.

In an interview with The Decoder, chief technology officer Philip Lorenz explained why he sees biology as a considerably harder problem for AI than language. He also argued that strong results on paper do not guarantee that a designed molecule will perform well in practice.

Why it matters

The funding round signals continued investor appetite for AI applied to scientific discovery, a field where the technology is being asked to produce physical outcomes rather than text or images. For organisations exploring AI in research and development, the interview points to a practical constraint: the usefulness of a model depends on the quality and breadth of the data behind it, and on whether its outputs are tested against real-world results.

That has implications beyond the laboratory. Any organisation deploying AI to make decisions or recommendations faces the same question of whether internal evaluation metrics reflect what happens when the output meets a customer, a patient or an operational process. Building that connection is a transformation task as much as a technical one, touching data pipelines, validation steps and the way teams share feedback.

The Renascence take

Most readers will focus on the headline figure and the investor names. The more useful signal is the gap Lorenz describes between a strong score and a working result, which is a familiar problem in service design.

A benchmark is a proxy, and a proxy that is optimised for its own sake will drift away from the outcome it was meant to represent. Teams building AI systems should treat every model score as a hypothesis to be tested against real behaviour, whether that behaviour is a molecule's performance in a lab or a customer's response to an automated decision. The customer-obsessed operator will fund the validation loop as seriously as the training run, and will measure success by the result that reaches the person at the other end.