Global demand for computing power for artificial intelligence is growing exponentially, and along with it, the industry's appetite for electricity. In response to this challenge, a group of researchers from Extropic and the Massachusetts Institute of Technology has proposed a concept that could overturn our understanding of the physics of computing. This is about a thermodynamic computer—an architecture that uses thermal noise not as an obstacle, but as a working tool.
Modern GPUs and ASIC accelerators spend enormous resources on maintaining determinism—the precision of every bit. However, many AI tasks, such as text generation or optimal pathfinding, are inherently probabilistic. The authors of the study demonstrated that for such tasks, the thermodynamic approach could be up to 10,000 times more energy-efficient than classical circuits.
How It Works
Instead of suppressing random thermal fluctuations in transistors, the new architecture uses them as a source of entropy for computations. Physical noise turns from an enemy into an ally: the system naturally "iterates through" probabilistic states, finding the most likely answer without the cost of precise mathematical operations. This is especially relevant for large language models, where each generation step is essentially a choice from a probability distribution.
AI Energy Crisis: Is There a Solution?
Major technology corporations are already investing tens of billions of dollars in building data centers, and the demand for electricity to train models continues to grow. If thermodynamic computing proves its viability in practice, it could reduce not only energy consumption but also capital costs for infrastructure. Instead of giant clusters with liquid cooling, we might see compact chips operating on principles of randomness.
When to Expect Commercial Solutions?
For now, the work is fundamental in nature—simulations and architectural sketches have been presented. It could be years before commercial samples appear. However, the very fact of such research emerging amid the AI boom indicates a serious shift in industry thinking. Quantum and neuromorphic computing are developing in parallel, and the thermodynamic approach could carve out its niche in this spectrum of alternatives.
Cryptalist Expert Opinion: From a long-term infrastructure investment perspective, any breakthroughs in AI energy efficiency will directly impact the value of tokens for projects related to decentralized computing and mining. If thermodynamic chips lower the barrier to entry for training models, we could see a new wave of decentralization of AI computing power. However, until then, the market will continue to seek a compromise between performance and electricity costs.