In a lab at MIT, Caio Silva watches a silicon chip do something that would alarm most engineers: heat flows through it, and the chip computes. The undergraduate physics student and his collaborators have built a system that encodes data as temperature and performs matrix vector multiplication—the core mathematical operation underlying large language models—with over 99% accuracy. No transistors switched. No binary logic gates fired. Just heat doing the thinking.
The work, led by Giuseppe Romano at MIT's Institute for Soldier Nanotechnologies, represents a fundamental inversion of how we design computing systems. "Most of the time, when you are performing computations in an electronic device, heat is the waste product," Silva says. "You often want to get rid of as much heat as you can. But here, we've taken the opposite approach by using heat as a form of information itself."
The mechanism is elegant. Input data maps to temperatures across a network of silicon structures, each geometry optimized by a physics-based algorithm developed by the team. As waste heat from any electronic device naturally flows through these microstructures, the thermal distribution and power collected at output terminals complete the calculation. The system performs matrix vector multiplication without electricity, relying instead on the physical behavior of heat transfer.
The implications extend far beyond energy efficiency. This approach could embed AI directly into physical infrastructure—no datacenters required. A power plant's turbines could run real-time combustion models using their own waste heat. A factory floor's machinery could optimize its own maintenance schedules through the thermal signatures it already generates. Temperature itself becomes the interface between computation and the physical world.
Scaling remains formidable. The team must tile millions of these structures together to approach modern deep-learning capacity. Accuracy degrades as matrices grow more complex, particularly when input and output terminals span large distances—the thermal signal blurs into noise. These are not trivial obstacles.
Yet the more immediate application bypasses them entirely. The same structures that compute can sense: detecting problematic heat sources and measuring temperature changes in electronics without any additional energy draw. This eliminates the multiple discrete temperature sensors currently consuming valuable chip real estate. Energy-free thermal intelligence, embedded directly into silicon.
The deeper significance lies in what this work reveals about computation itself. By developing algorithms that exploit physical heat dynamics, the team has shown that the boundaries of computing are not fixed by transistor geometry or semiconductor physics. Information processing can be distributed into the infrastructure around us, not merely housed in dedicated computing hardware. Heat becomes not a burden to manage but a medium to harness. Every warm chip, every thermal gradient in industrial equipment, every heat signature in a building or vehicle becomes a potential substrate for intelligence. The researchers have transformed the very thing that costs the computing industry billions in cooling expenses into a resource.