Artificial intelligence turns to physical laws to cool computing facilities

A newly funded physics laboratory aims to simulate thermodynamics and electrical loads before construction crews pour concrete for computational facilities.

Modern computing facilities generate immense thermal energy as microprocessors calculate billions of operations each second. That heat does not vanish on its own. It moves through metal heatsinks, enters circulating coolant or ambient air, and forces mechanical cooling equipment to consume vast quantities of electricity just to keep chips below damaging temperatures. When electrical current flows through dense networks of power delivery lines and processors, power losses compound at every physical interface. Balancing heat dissipation, fluid motion, and electrical resistance across a facility requires managing interconnected physical forces simultaneously.

Addressing these thermal and electrical bottlenecks has become essential as large-scale computing clusters expand. If a facility runs even slightly too hot or routes electricity inefficiently, operational expenses surge and electronic hardware degrades prematurely. The fundamental constraint on computing growth is no longer merely chip fabrication, but the classical laws of heat transfer, fluid dynamics, and power distribution.

To model those constraints before heavy infrastructure is built, Physical Superintelligence, a Cambridge, Massachusetts startup, launched on September 1, 2026, with $58 million in seed funding.¹² The funding round was led by Breakthrough Energy Ventures, an investment firm founded by Bill Gates, with participation from venture funds including Dragon Global, Robot Ventures, Solari, Susa, SV Angel, and Valkyrie, as well as individual investors from technology firms such as OpenAI, NVIDIA, and Oracle.²¹³ The venture intends to deploy computational models that simulate multi-physics environments to optimize cooling and electrical flows for terrestrial and orbital data centers.¹³

How can algorithmic models simulate real world physics?

Simulating complex physical systems requires calculating how different natural forces interact over time rather than evaluating isolated components. The company organized its software platform, named Emmy after mathematician Amalie Emmy Noether, as a collection of automated research agents designed to reason through physical laws and run specialized numerical simulations.¹ While an individual engineer evaluates a single mental model, Emmy breaks research problems into branching trees of verifiable hypotheses and tests them concurrently across simulations.⁴

The system applies physics-based reasoning to evaluate interactions among power distribution, liquid and air cooling, networking layout, and raw computing hardware. Co-founder and chief executive officer Matt Pines told Semafor that the platform incorporates strict verification mechanisms designed to catch errors when computational models generate false conclusions with high statistical confidence.²¹ The company intends to test these simulations on infrastructure design before physical construction begins, starting with an planned project at a large computing site in Texas, as well as retrofitting operating facilities.²¹

What are the limits of automated physical reasoning?

Computational simulations of physical laws remain mathematical approximations rather than direct experimental proofs. The outputs generated by automated reasoning systems depend strictly on the boundary conditions, numerical assumptions, and simulation fidelity programmed into their environments. While automated agents can verify internal mathematical consistency against established equations, they cannot substitute for empirical testing in operational engineering environments.

This limitation is explicit in the company's technical work outside commercial facilities. Physical Superintelligence served as the founding technical partner on the Fermi Explorer mission, an initiative run by a non-profit organization that aims to plan a small interstellar probe to the Alpha Centauri star system.⁴² In a feasibility assessment prepared in July 2026, the company used its automated platform to calculate orbital trajectories, but noted in the document that the quantitative assessment had not undergone comprehensive human peer review.⁵

Where will automated physics platforms expand next?

Optimizing industrial energy use across computing infrastructure represents the initial commercial testing ground for automated physics software. Carmichael Roberts of Breakthrough Energy Ventures said that artificial intelligence has the potential to alter the pace of scientific discovery, pointing to physics as a domain where computational reasoning could influence hardware design.¹ Beyond facility cooling, the company stated that its seed funding will support research into underlying physical hardware, spanning advanced sensors and experimental computing substrates.¹⁴

The venture has also released an open-source copilot tool called Get Physics Done to assist researchers with theoretical formulations.⁴ The primary open question is whether computational models tested across simulated environments will deliver measurable efficiency gains when applied to operating power grids and high-density computing buildings under fluctuating environmental conditions.

This piece was prepared from announcement documents and public records; the founders have not been interviewed.

References

This article is based on 5 sources, listed in the order they are cited.

1. 1 H hpcwire.com third party · 1 Sep 2026 Physical Superintelligence Raises $58M to Develop AI Physics Platform - HPCwire See the source
2. 2 JC J.D. Capelouto announcement · 2 Sep 2026 Startup launches AI physics lab to maximize data center efficiency See the source
3. 3 U unite.ai Physical Superintelligence 获得 5800 万美元种子轮融资 See the source
4. 4 U unite.ai Physical Superintelligence získává 58 milionů dolarů v seed kole See the source
5. 5 U unite.ai Physical Superintelligence залучає $58 млн посівного раунду See the source