AINEWS
Gallium, the soft metal in high speed chips and radar, is almost entirely refined in China. CuspAI's $450M Series B funds a search engine for substitutes.
Gallium is a soft, silvery metal that shows up in the high-speed semiconductors behind radar, LEDs, and the accelerators that train frontier AI models. By widely-cited industry estimates, the world refines roughly 99% of it in China, and the International Energy Agency projects data-center demand for the metal could exceed supply by about 10% by 2030.
The Cambridge, UK-based startup CuspAI closed a $450 million Series B on 20 July 2026 at a post-money valuation of about $2.6 billion, roughly five times its level from September 2025. Kleiner Perkins and New Enterprise Associates co-led the round; Bezos Expeditions, AMD Ventures, and the UK government's Sovereign AI Venture Fund joined. Founded in 2024 with no public products, CuspAI is carrying a $2.6 billion mark on a category, not a revenue line.
Its product, a platform called MIRA, runs the kind of material-discovery work that a graduate lab would otherwise grind through on a bench for years. "CuspAI built a search engine that changes that," Kleiner Perkins partner Josh Coyne told CNBC at the announcement. The platform simulates candidate materials in software, then screens them for the electronic, optical, or thermal properties a chipmaker or a defense supplier would want, including potential gallium substitutes.
CuspAI is launching what it calls an "AI Materials Foundry," a coalition of 45 founding partners, including Nvidia and Meta, that will feed real industrial problems into the platform and stress-test its output. The Guardian frames the round as part of a broader UK bet on sovereign AI infrastructure, with the British government's participation meant to keep at least one materials-discovery pipeline anchored in Europe as the dependency tightens.
The IEA's projection is the load-bearing forecast here, and it is a forecast. A roughly 10% gap between data-center gallium demand and supply by 2030 is a modeled shortfall, not a measured shortage, and the agency's underlying assumptions about AI build-out, cooling architectures, and the rate of gallium-arsenide and gallium-nitride substitution all move the number. Capital is pricing the direction, not the precision.
Two cautions matter. Materials discovery is a long-cycle problem even with machine learning; a single platform does not re-route a supply chain in a quarter. The "99% refined in China" figure that anchors the geopolitical frame comes from secondary write-ups of USGS and IEA data, and should be read as a widely-cited approximation rather than a precise measurement.
For the past two years the chip industry has been answering the supply question with capacity: more fabs, faster. The CuspAI round, and the consortium built around it, points at a different lever. The constraint is the materials one country controls, and the bet is that new chemistry, not new capacity, will route around them.