‘We need the bravery to change the process’: How AI is impacting physical environments
Sifted
AI is moving from chatbots into labs and factories, tackling materials science and drug discovery. European startups in these fields have already raised billions this year, more than all of last year combined.
Forget another chatbot update. The more interesting AI story right now is happening in beakers and wind tunnels, not browser tabs. Startups building physical-world AI tools have pulled in serious money in 2025: European advanced materials companies have raised €3bn so far this year, nearly double the €1.6bn raised in all of 2024, according to Sifted data. Drug discovery startups are tracking similarly, at €4bn already versus €4.7bn for the whole of last year.
The pitch from founders in this space is that language models solved the easy problem. Chad Edwards, who runs CuspAI and uses AI to hunt for climate-friendly materials, points out that code can be written and tested in seconds, but materials science doesn't work that way. You still have to synthesize something, run it through a lab, wait. That slow feedback loop is exactly why AI in physical science has trailed behind software for so long, and why simply throwing a language model at a chemistry problem doesn't cut it.
What's changing, according to Edwards and Anthony Bradley at DaltonTx, is that models are being built with actual physics baked in rather than just pattern-matching on data. Bradley's example is antibody design: an AI generating new molecules has to know what's biologically expressible, not just what looks statistically plausible. That requires a mix Bradley calls non-negotiable — deep domain science, strong AI, and genuinely good engineers, because these are hard science problems dressed up as software problems.
Iraia Ibarzabal at Multiverse Computing flags a different bottleneck: infrastructure. A model that behaves in a lab often falls apart in the field, whether that's a factory floor or a defense application running without cloud access. Her answer is building for the real deployment target from day one, not retrofitting AI onto existing workflows. Bradley makes the same point more bluntly — sprinkling AI onto old processes doesn't transform anything, and organizations need the nerve to actually redesign how they work.
None of this displaces scientists, at least not yet. Edwards frames it as removing drudgery — instant literature search instead of manually trawling papers, freeing researchers for actual thinking. And Bradley offers a useful corrective to the hype: AI is good at recombining what's already been measured, weak at anything genuinely novel outside that data. The industry's real test over the next five years, Edwards says, won't show up on a dashboard. It'll show up in solar panels, chips, and materials quietly doing their job in the physical world.
My take
The refreshing part here is founders admitting AI's limits out loud — Bradley saying outright that models can't invent things from thin air is the kind of honesty missing from most AI marketing. Europe's advantage in this race won't come from chasing bigger chatbots; it'll come from pairing serious domain scientists with AI tooling in materials and pharma, areas where the continent already has research depth. Anyone still measuring AI progress by benchmark scores instead of actual solar panels or drugs reaching patients is watching the wrong scoreboard.
Read more about this at: Sifted
