Orbital $50M Round Builds AI That Designs New Materials

The chips inside an AI data center now run so hot that cooling them has become an engineering crisis. Orbital Industries just raised $50 million on a strange premise: use AI to invent the materials that cool the machines running AI. The recursion is not a gimmick. It is the clearest sign yet that the next bottleneck in artificial intelligence is not software at all.

What Actually Happened

Orbital Industries, a startup that uses AI to design advanced materials, raised a $50 million Series B led by the venture firm Plural, as Fortune first reported on May 28. The most strategically loaded name on the term sheet is NVentures, the venture arm of Nvidia, which had backed the company before and returned for this round alongside existing investors Radical Ventures, Compound, and Fly Ventures. Orbital runs offices in London and San Francisco, employs roughly 50 people, and recently rebranded from its original name, Orbital Materials. The new capital is earmarked to scale commercial deployment of its first two products, both aimed squarely at the data center industry, and to grow the team.

The engine underneath the company is an AI model called Orb, which predicts and simulates the quantum-mechanical behavior of atoms. Orbital says Orb is the only model that can simulate 100,000 atoms on a single GPU and that it runs roughly 10 times faster than alternatives, outperforming materials models released by Meta and Microsoft. The company used Orb to screen hundreds of thousands of candidate compounds and surface a new kind of liquid coolant for the server racks full of GPUs that now define AI infrastructure, one engineered specifically to avoid PFAS, the so-called forever chemicals facing tightening environmental restrictions in the United States and Europe. Beyond data centers, Orbital says it intends to extend the same platform to broader industrial applications.

Why This Matters More Than People Think

Materials discovery has always been the slowest loop in hard technology. Inventing a new compound, a better battery cathode, a non-toxic coolant, or a stronger alloy traditionally means years of trial and error in a wet lab, synthesizing and testing one candidate at a time. Orbital is doing to that loop what AlphaFold did to protein structure prediction: replacing physical experiments with a model that approximates the underlying physics fast enough to explore an enormous search space in silico. When a model can simulate 100,000 atoms on one GPU, a researcher can evaluate thousands of hypothetical materials in the time it once took to test one.

The choice of first market is the tell. Modern AI server racks are crossing 100 kilowatts of power draw, and air cooling has run out of headroom, which is why the industry is racing toward liquid and immersion cooling. The dominant coolants have relied on PFAS chemistry, and regulators on both sides of the Atlantic are moving to restrict those compounds. A data center operator that builds around a PFAS coolant today is building around a future ban. A drop-in replacement that delivers the same thermal performance without the regulatory time bomb is not a science project, it is a procurement decision waiting to be made, and that is why Nvidia cares enough to invest twice.

The timing also matters because the cost of getting cooling wrong is rising fast. A hyperscaler committing billions to a new campus has to choose a thermal architecture years before the chips arrive, and a wrong bet on a soon-to-be-restricted chemical could strand that investment. The opportunity for Orbital is that it sells certainty into a market drowning in regulatory uncertainty; the risk is that the same conservatism that makes operators fear PFAS also makes them slow to qualify anything new, however promising the data looks on a slide.

The Competitive Landscape

Orbital is competing against the largest labs in the world for the title of foundation model for atoms. Google DeepMind shipped GNoME, which proposed millions of stable crystal structures. Microsoft released MatterGen, a generative model for inorganic materials. Meta open-sourced its Open Materials dataset and models to commoditize the field. Orbital''s claim that Orb beats the Meta and Microsoft offerings on speed and scale is the entire basis of its valuation, because in a market where Big Tech is giving models away, a startup has to be both faster and more useful than free.

The difference is what each player does with the model. DeepMind, Microsoft, and Meta treat materials AI as research and as a way to sell cloud and tooling. Orbital is taking the opposite path: it is using Orb internally to invent specific products and then selling those products directly, starting with the data center coolant. That vertical posture puts it in a different competitive set than the labs. Its real rivals become the incumbent chemical and cooling suppliers whose catalogs it wants to replace, and the new wave of AI-for-science startups like Periodic Labs and Lila Sciences chasing the same idea that scientific discovery is now a software problem. Nvidia''s presence on the cap table is the strategic anchor: the company that sells the hot chips has a direct interest in whoever can cool them cheaply and legally.

Capital is also flooding the broader AI-for-science category, which cuts both ways for Orbital. On one hand, validation from peers and investors lifts the whole field and makes enterprise buyers comfortable that simulation-led discovery is real. On the other, it means richer rivals can fund the same atom-by-atom search and the same data center coolant, turning a clever head start into a crowded race. Orbital's answer is to move from discovery into manufacturing faster than the labs are willing to, trading the safety of pure software for the messier economics of making and selling a physical good.

Hidden Insight: AI's Next Bottleneck Is Atoms, Not Bits

For three years the AI story has been about bits: bigger models, more parameters, longer context windows. The Orbital round is a marker that the binding constraint is shifting to the physical world. The buildout of AI compute is now gated by power generation, by heat dissipation, by water for cooling, and by the supply of exotic materials for chips and interconnects. The industry spent its first act optimizing software and is entering a second act where it must optimize matter. A company that can design better materials on demand sits directly on that constraint, which is why a 50-person startup with two products can attract Nvidia''s venture arm.

There is a deeper elegance to the specific bet. Orbital is using AI to relieve a bottleneck that AI itself created. The reason data center cooling is a crisis is that AI workloads made the chips run hotter than anything before them. Orb is, in effect, AI turned back on its own infrastructure problem, compressing the discovery of the coolant that lets the next generation of GPUs exist. If that flywheel works, it suggests a pattern for the whole sector: the physical constraints on AI growth become the most valuable problems to point AI at.

The risk is that simulation is not the same as shipping, and this is where the hype gets ahead of the chemistry. Finding a promising molecule on a GPU is the first step in a long march. However fast Orb explores the search space, a candidate coolant still has to be synthesized at industrial volume, tested for stability and toxicity over years, qualified by conservative data center operators who hate changing anything that works, and manufactured at a price that beats the incumbent. The bear case is straightforward: materials startups have a long history of dazzling simulations followed by a brutal valley of death at scale-up, where the molecule that worked in a model never survives contact with a factory. Skeptics point out that Orbital''s revenue today is almost certainly a rounding error against its valuation, and that the gap between a screened compound and a qualified product can swallow years and the entire raise.

That is also why the company''s structure is the most interesting thing about it. By selling materials rather than licensing the model, Orbital is betting that the durable value is in the products, not the software. A model that beats Meta and Microsoft today will be matched within a year, because that is how this field moves. A qualified, manufactured, regulation-proof coolant with signed data center contracts is far harder to copy. Orbital is quietly choosing to become a materials company that happens to run on AI, rather than an AI vendor that happens to know chemistry, and that choice is the real strategy hiding under the headline.

Consider the order of magnitude this changes. A traditional materials program might test a few hundred candidates a year and call a single commercial compound a decade of work. A model that screens hundreds of thousands of candidates compresses the front of that pipeline from years to weeks, which means the rate-limiting step is no longer ideas but validation and production. That inversion is the quiet revolution: when discovery becomes cheap, the scarce resource becomes the ability to manufacture and qualify at speed, and the companies that win will be the ones that rebuilt their factories around an AI that never stops proposing candidates.

Separating what Orbital has proven from what it has promised matters here. The proven part is the model: simulating 100,000 atoms on one GPU at ten times the speed of rival systems is a genuine technical result that Meta and Microsoft can measure against their own. The promised part is the business, that those simulations convert into products operators will buy at a margin that supports the valuation. Most AI hype collapses the two, treating a benchmark win as if it were a revenue win. Orbital's backers are betting the team can bridge that gap, and the data center coolant is the first plank in the bridge.

What to Watch Next

In the next 30 days, watch for named customers. A funding announcement is a promise; a signed data center operator deploying the PFAS-free coolant at scale is proof. The single most informative event would be a major colocation or hyperscale provider confirming a commercial order, because that converts the simulation story into revenue. Over the next 90 days, watch the second product Orbital has hinted at and whether the company keeps Orb proprietary or moves to license it, a decision that will reveal whether management believes the moat is the model or the materials.

Over the next 180 days, the question is manufacturing. Designing a coolant is the easy half; producing it reliably and cheaply is the half that kills most materials companies, so watch for partnerships with chemical manufacturers or contract producers. Watch, too, whether Nvidia deepens the relationship beyond a venture check into a supply or qualification agreement, which would signal that the cooling problem is urgent enough for the chipmaker to pull Orbital''s products into its reference designs. If by year end Orbital is still showing benchmarks instead of shipments, the atoms-not-bits thesis will have met its first hard test.

Orbital is using AI to solve a problem AI created: the chips got too hot, so it built a model to invent the coolant that lets the next ones exist.

---