The Most Interesting AI Startup This Week Isn't Trying to Replace You
A British ex-DeepMind researcher just raised $50 million to do the least glamorous thing in AI right now: make a new liquid that keeps server racks from cooking themselves. No chatbot. No agent that books your flights. A coolant. And I think it's one of the more honest bets in the whole field. Here's the setup. The AI boom runs on GPUs, and GPUs run hot. Jonathan Godwin, Orbital's CEO, describes a modern GPU rack as packing "basically a supermarket's worth of energy into a filing cabinet." You have to pull that heat out somehow, and one of the best ways the industry found uses a class of chemicals called PFAS. You've heard the other name for them. Forever chemicals. That's the part of this story nobody puts in the funding headline. PFAS don't break down. They show up in roughly 45% of tap water samples tested across the US. They build up in your liver and kidneys over years, and they've been linked to liver damage, thyroid disease, and kidney and testicular cancer. The companies that made these chemicals profitable were rarely the ones drinking the runoff. nihThe Cooling Report So when Orbital says its AI screened hundreds of thousands of candidates to find a coolant that does the job without the forever chemicals, that's not a vanity product. That's pointed straight at a problem that's already in somebody's well water near a chemical plant in North Carolina. Fortune Now the skeptic in me has to speak up, because this field has a track record. Two years back, DeepMind announced its AI had found 2.2 million new crystals, 380,000 of them stable enough to maybe synthesize. Sounded like a new world. Then actual materials scientists opened the database and found over 18,000 of the "discoveries" contained radioactive elements so scarce they're useless in the real world. One reviewer called the whole thing a random walk through the periodic table. SubstackWinsomemarketing The harder truth, from MIT Technology Review back in December: despite all the money pouring in, AI materials discovery still has no convincing big win. No AlphaFold moment. The field is still waiting. MIT Technology Review That's the context Orbital is walking into. And here's why I don't lump it in with the hype. Most of these startups find a compound and license the idea to a chemicals giant. Orbital decided to do the boring, expensive, unsexy thing instead. It's building the manufacturing. It found a contract maker, it's running molecules through qualification with chip providers, and it wants to ship the coolant alongside the next GPUs in 2027. Godwin claims a new cooling fluid normally takes ten years and $100 million, and they did it in months for far less. FortuneFortune That claim is doing a lot of work, and I'd want to see it survive contact with a real factory. Predicting a molecule on a GPU is one thing. Making it at scale, cheap, and getting it certified is where the bodies are usually buried. But notice the shape of the bet. He's not promising to discover a thousand miracle materials. He's promising one specific product, with a ship date, aimed at a real toxic mess. If it lands, he says, it'd be the first AI-designed molecule to hit the commercial market in any industry. Even in drug discovery, where people have chased this for years, nothing AI-found has made it through trials yet. And the guy isn't shy about where this goes. He wants to build the largest industrial conglomerate in Europe, an AI-native answer to the chemical giants that rose a century ago. The reason those old companies still rule, he says, is entrenched moats, and the only way through a moat is a radical technology. FortuneFortune Maybe. Or maybe the moat is the factory, the certification, the decades of process knowledge, and AI just gets you to the starting line faster than before. I keep thinking about the family near the Chemours plant whose water got contaminated for years before anyone told them. A non-toxic coolant doesn't undo that. But if the next data center going up doesn't need the poison in the first place, that's worth more than every agent demo I've watched this year. Whether the molecule survives the factory floor, we find out in 2027.
