Physical Superintelligence turns data centers into AI research tests

Physical Superintelligence earns its current revenue from AI tools that optimize data centers, even as the company reported a complete Lean-checked proof on the third day of a run on November 6, 2026. The company reports more than $58 million raised and more than 25 researchers and engineers, with Breakthrough Energy Ventures leading the funding.

TechCrunch reported the split on November 10, 2026. Chief executive Matt Pines described the objective as converting theoretical-physics breakthroughs into actual technologies. Data centers are the first test environment, where PSI’s “virtual physicists” can work on live physical systems with measurable operational constraints.

PSI says it operates as a public benefit corporation. Its charter establishes safety, verification, responsible development and broad public benefit. The company’s immediate business remains facility optimization, while the larger plan depends on validated discoveries that can become devices or other technologies.

That approach faces a clear commercial boundary. Pines said mathematics and computer science can express results in validated code, but physics and life sciences still require experimental contact with nature. He summarized the distinction with an observation: “nature gets a vote in your models of reality.”

What does PSI sell today?

According to PSI, its current optimization work simulates heat, power, water and sound. The company says it builds high-fidelity models of live facilities so virtual physicists can inspect complete systems, uncover hidden constraints and test interventions. Those models connect facility performance and community impact to the site’s underlying physics.

PSI lists the intended gains from that work: increased computational density, more efficient energy use, improved thermal management and higher capital efficiency. The company presents these as intended results, not promised outcomes. Its present business can therefore stand on operational gains, while the research program searches for findings with wider applications.

Each commercial campaign also supplies research feedback. PSI says live facilities impose constraints that simulations would not propose, and each campaign trains the virtual physicists used in later work. A deployment can therefore reveal a practical limitation, feed that limitation into the next system and generate revenue at the same time.

The data-center business is both a revenue source and a research input. PSI says operating constraints expose weaknesses that simulations may miss. The same campaigns can generate revenue and improve the system used in later work.

Takeaway: Data-center work supplies both revenue and real-world constraints for PSI’s research system.

How does the research system test a discovery?

Pines said an effective research system combines software harnesses around models, tools, memory systems and simulation infrastructure. In his account, that structure lets PSI pursue a problem at breadth and depth while considering a wide range of alternatives. The machinery is designed to carry work beyond an initial answer and into testing, engineering and deployment.

PSI formalizes each continuous effort as an Autonomous Campaign. Its sequence runs from initial research through use in the physical world:

• Problem definition
• Hypothesis
• Verification
• Experiment
• Engineering
• Deployment

According to PSI, every campaign has measurable milestones, hard verification gates and a path toward application. Human researchers choose the missions and their success criteria. Their role sets the target, while the autonomous system searches for and tests possible routes to it.

Takeaway: Human researchers set the missions; the Autonomous Campaign supplies the repeatable process.

How does PSI check the result?

PSI says its architecture separates candidate generation from verification. A proposed result must survive checks that span formal mathematics, simulated systems and measurements of physical systems. The verification set includes:

• Conservation laws
• Symmetry
• Dimensional consistency
• Formal proof
• Simulation
• Digital twins, or virtual counterparts of physical systems
• Physical measurement

The company says the verifiers that certify a result can also become reward functions, or training signals, for later virtual physicists. Verification therefore does more than reject weak answers. It can supply the positive examples used to improve subsequent research runs, connecting evaluation directly to system development.

Takeaway: Verification can train later virtual physicists instead of serving only as a final check.

What did the November proof demonstrate?

On November 6, 2026, PSI announced a concrete test of that process. The company said an internally developed system, running a then-non-public Anthropic model, produced a complete proof of the Aaronson–Ambainis conjecture on the third day of a run.

Lean, the proof assistant that checks every step, accepted the proof. It later accepted a final version after two improvements.

PSI says the conjecture had been open since 2008. It concerns the conditions under which quantum computers can be exponentially faster than ordinary computers on random inputs. The company said David Turturean led the work with early access to Claude Fable 5.2, which Anthropic publicly introduced on November 9, 2026.

PSI presented the result as evidence for its combined generation-and-verification architecture. Lean supplied a step-by-step check outside the research system, while PSI said its internally developed system produced the proof. The company’s claim is concrete: a run produced a complete mathematical artifact that a proof checker accepted.

The result does not, by itself, show that the same process can handle physics problems with commercial value. Nor does it establish that a model can repeatedly make experimentally validated discoveries. It does show a full path inside mathematics: propose a proof, check each step in Lean, revise the work and secure acceptance of the final version.

Takeaway: The Lean proof is verified, but it does not yet show a repeatable route to commercial devices.

What has to happen before commercialization?

Pines said PSI is working on efficient ways to validate computational results through experiments. He also said theoretical physics is not yet mostly the domain of machines, despite a stronger, conditional pitch to investors six to nine months earlier. That qualification sits at the center of the business.

Facility optimization can earn revenue under present constraints, while the larger ambition depends on discoveries that survive experiments and can support devices. The company would need repeatable, experimentally confirmed results before it could move from optimizing AI infrastructure to creating technologies based on new physics.

Pines connected a validated fact of nature to the possibility of building a device with capabilities that were not previously available. That is the standard he described for the research system. A successful result must clear formal checks, match physical measurements and survive contact with nature.

The next milestone is therefore practical rather than another formal proof. PSI would need to show that its virtual physicists can make reliable, experimentally verified discoveries at a repeatable pace, and that those discoveries can support useful products.

Takeaway: PSI’s next test is repeatable, experimentally verified discoveries that can support products; data-center optimization remains its revenue engine meanwhile.