The Lattice
An international consortium used machine learning to discover two superconductors, the first ever found this way. Both conduct at less than one degree above absolute zero, 292 degrees from room temperature. On the same day the paper was published,...
An international consortium used machine learning to discover two superconductors, the first ever found this way. Both conduct at less than one degree above absolute zero, 292 degrees from room temperature. On the same day the paper was published, a materials AI startup raised $400 million at a $2.6 billion valuation. The venture capital is pricing the method. The method just proved it works.
Eleven physicists spread across four universities published a paper in Physical Review Research on June 17 reporting the discovery of two new superconducting materials. The materials are yttrium-ruthenium-boron and lutetium-ruthenium-boron, designated YRu3B2 and LuRu3B2. Both derive their superconductivity from electrons forming flat bands within a kagome lattice, a geometric pattern named after traditional Japanese basket weaving. Both superconduct below one degree above absolute zero. They are, by any practical measure, useless. They are also the first superconductors in history discovered through machine learning.
The critical temperatures are 0.81 Kelvin and 0.95 Kelvin. Liquid helium boils at 4.2 Kelvin. Room temperature is 293 Kelvin. The distance between what the SuperC consortium found and what civilization needs is 292 degrees.
What matters is not the destination but the speed of arrival. A machine-learning algorithm screened an enormous chemical search space and ranked candidates by their predicted likelihood of superconductivity. Quantum-geometry calculations then refined the shortlist to compounds where electronic structure theory predicted flat-band behavior favorable for Cooper pairing. Collaborators at Rice University, led by Professor Emilia Morosan, synthesized the two top candidates and confirmed bulk superconductivity in both. The pipeline from computational screening to physical verification compressed what traditionally takes years of laboratory trial and error into months.
The SuperC consortium was founded in 2023 by Aalto University's Paivi Torma and an international team of ten other researchers, including Princeton theorist B. Andrei Bernevig and Ruhr University Bochum's Miguel A.L. Marques, who developed the screening methodology. The consortium's stated goal is to discover a room-temperature superconductor by 2033. It runs on approximately 2.7 million euros in philanthropic funding from the Kavli Foundation, the Klaus Tschira Stiftung, and four smaller donors.
That goal requires closing a large gap on a short timeline. Superconductivity was discovered in 1911, when Heike Kamerlingh Onnes cooled mercury to 4.2 Kelvin and measured zero electrical resistance. The record critical temperature at ambient pressure stood at approximately 133 Kelvin for three decades, set in 1993 by the cuprate compound HgBa2Ca2Cu3O8. In March 2026, University of Houston physicists raised that record to 151 Kelvin by pressure-quenching the same compound, briefly compressing it to 30 gigapascals and rapidly releasing the pressure to lock in a metastable phase. Even with that advance, the gap from 151 Kelvin to room temperature is 142 degrees. From SuperC's materials, it is 292.
They are not the only ones making that bet. On the same day the SuperC paper was published, the Financial Times reported that CuspAI, a Cambridge-based AI materials discovery company, was closing a $400 million funding round at a $2.6 billion valuation. The round was co-led by Jeff Bezos' family office, Bezos Expeditions, and Kleiner Perkins. Nine months earlier, CuspAI had been valued at $520 million. Its advisory board includes Geoffrey Hinton and Yann LeCun. The company describes itself as building a search engine for the material world.
CuspAI is not an outlier. XtalPi, a materials AI company focused on pharmaceuticals, carries a $2.5 billion valuation. Flagship Pioneering launched Lila Sciences with a $200 million seed round. Periodic Labs raised $200 million at a $1 billion valuation. The global market for AI in materials discovery was estimated at $2 billion in 2025 and is projected to reach $17.9 billion by 2034, a compound annual growth rate of 28 percent.
Google DeepMind demonstrated the scale of what computation can find. Its GNoME project, published in Nature in November 2023, used graph neural networks to discover 2.2 million new stable crystal structures, equivalent to approximately 800 years of traditional experimental discovery. External researchers subsequently synthesized more than 700 of those predicted structures and confirmed their stability. The NIMS SuperCon database, the most comprehensive catalog of superconducting materials, contains roughly 33,000 entries accumulated over a century. Machine learning produced two orders of magnitude more crystal structure candidates in a single project.
SuperC's contribution is narrower but more pointed. GNoME predicted stability. SuperC predicted a specific, useful property: superconductivity. The kagome lattice that appears in both YRu3B2 and LuRu3B2 was not stumbled upon. The algorithm specifically searched for quantum-geometric signatures that theory identifies as favorable for flat-band superconductivity. It found structures matching those signatures, and synthesis confirmed they work. The method is not accidental discovery at machine scale. It is targeted hunting.
A room-temperature superconductor would be worth hunting for. The United States loses approximately five percent of its generated electricity during transmission and distribution, roughly 207 million megawatt-hours per year, valued at more than $6 billion at wholesale prices. Globally, transmission losses run 8 to 9 percent of total production. Superconducting power lines would eliminate resistive losses entirely. Medical MRI machines currently require liquid helium cooling at 4 Kelvin; removing that requirement would cut costs and expand access to clinics that cannot maintain cryogenic infrastructure. Quantum computers operate in dilution refrigerators near 15 millikelvin; room-temperature superconducting qubits would transform them from laboratory instruments into deployable systems.
The investment disparity is the telling fact. SuperC published the first machine-learning-guided superconductor discovery on the same day CuspAI closed a funding round 150 times larger than the consortium's entire budget. The five largest technology companies committed more than $600 billion in capital expenditure in 2026, nearly all of it for training and serving language models. The AI materials discovery market, at $2 billion, amounts to roughly one third of one percent of that figure. A room-temperature superconductor would restructure the global power grid. The search for one is funded like a postdoctoral fellowship.
That ratio will not hold. The same venture capitalists who valued language model companies at trillions of dollars are now arriving at materials science, carrying the same thesis: that AI can compress search from decades to months, and that whoever finds the right material first captures the value. CuspAI's valuation quintupled in nine months. Periodic Labs went from founding to a $1 billion valuation in a single round. The capital is beginning to follow the method into domains where the payoff, if the method works, is measured in infrastructure rather than subscriptions.
The lattice that matters is not the kagome pattern inside the crystal. It is the computational search grid that can now evaluate billions of elemental combinations for a specific physical property and surface the candidates most likely to work. The first superconductor ever discovered by that grid conducts at less than one degree above absolute zero. That is 292 degrees from changing the world. The method that found it runs at room temperature, and it scales.