# Thermodynamic Computing Could Slash AI Energy Consumption by Orders of Magnitude
Researchers from Extropic and MIT have proposed a classical computing architecture that could dramatically reduce the energy demands of artificial intelligence workloads. The Denoising Thermodynamic Computer Architecture uses probabilistic computing — calculations based on manipulating probability distributions — built from conventional transistors to perform AI tasks with roughly 10,000 times less energy than GPUs on benchmark tests.
Rather than deterministic calculations, the system harnesses controlled randomness and draws on concepts from diffusion models, breaking complex computations into simpler denoising steps that progressively transform noise into structured data. The team fabricated and tested a transistor-based random-number generator to validate feasibility, though the complete architecture remains theoretical.
While promising, the work faces scaling challenges. Current benchmarks use simple datasets like Fashion-MNIST, far simpler than modern large language models. The researchers acknowledge that efficiently handling increasingly complex data and integrating probabilistic hardware with conventional neural networks will require additional advances before this approach could replace existing AI accelerators.
Read the full article at [The Quantum Insider](https://thequantuminsider.com/2026/07/03/researchers-propose-thermodynamic-computing-architecture-that-could-dramatically-reduce-ai-energy-use/).