PSI's physics-lab bet: $58M seed led by Breakthrough Energy

Physical Superintelligence (PSI), a Cambridge, Massachusetts startup founded by Matt Pines, Alexander Wissner-Gross, Alex Klokus and Alessandro Morari, came out of stealth on September 1, 2026 with $58 million in seed funding led by Breakthrough Energy Ventures. Its bet: AI can industrialize physics discovery through Emmy, a platform of 'virtual physicists' that generates hypotheses, writes code, runs simulations and tests designs, with every result forced through independent verification gates such as conservation laws, symmetry, dimensional consistency, formal proof and physical measurement.

PSI argues that twenty-first-century technology still runs on twentieth-century physics, and AI is now hitting those walls: energy caps compute, heat caps density. Its first commercial wedge is deliberately pragmatic — model the coupled physics of power, cooling, networking and compute in AI data centers, optimize facilities before construction and inside live ones, and sell measurable efficiency. Investors say some initial customers are data centers, with early work reported at a large Texas site; neither PSI nor its backers have named customers or published measured savings.

PSI also frames itself as an AI-native physics laboratory meant to run hundreds of research campaigns at once and eventually discover new physical laws, organized as a public benefit corporation with safety and verification in its charter. It is founding technical partner to Fermi Explorer, a 501(c)(3) effort to send a privately funded probe to Alpha Centauri — a trip of roughly 70,000–75,000 years by conventional means — by end-2029 with at least 1 kg of payload for under $15 million, and says Emmy validated the mission's physics and found a more efficient trajectory.

As of early September 2026 the company has no public performance data, pricing, or independently measured efficiency results — its own July 2026 feasibility assessment passed internal audit and adversarial proof review but not full peer review. The bet is therefore unproven but structurally coherent: verification gates address the most common failure of scientific AI, and the data-center market gives the research agenda a paying customer before it tries to rewrite the laws of the universe.