Oct 8, 2025, 9:27 AM ETConference Presentation, Keynote

Guillaume Verdon, Extropic AI: Thermodynamic Computing and the Energy Efficiency Crisis of AI

Generative AI demand is outpacing Moore's Law, with parameter counts and floating-point operations (FLOPs) growing exponentially.
The current scaling strategy relies on deterministic transistors, which are approaching thermodynamic limits where reducing voltage creates stochastic (random) errors.
Energy requirements for training large models are projected to exceed the total electricity generation capacity of the United States for future agentic and video models.
Heat dissipation from future AI compute clusters could theoretically generate more thermal radiation than the Earth currently emits.
Current hardware architectures struggle to support the probabilistic algorithms that nature and complex data distributions inherently require.

Xtropic posits that both advanced algorithms and physical hardware inherently favor probabilistic computation over deterministic logic.
Monte Carlo methods and "test-time compute" (e.g., reasoning trees) allow smaller models to achieve high performance by trading training compute for inference complexity.
Current deterministic computers are inefficient at running probabilistic algorithms, requiring energy-intensive pseudo-random number generation and Markov Chain simulations.
The solution involves reverse-engineering the physics of the brain's mesoscale, where thermodynamic fluctuations and electron jitter naturally create probabilistic behavior.
Xtropic defines its approach as "thermodynamic computing," aiming to build a substrate optimized for space, time, and energy efficiency.

Founders Trevor and co-founder transitioned from quantum computing to thermodynamic computing, identifying the former as too slow and misaligned for scaling AI.
Phase 1 (Room-sized prototype): A superconducting device utilizing three probabilistic bits (p-bits) operating at cryogenic temperatures near absolute zero.
Phase 2 (Silicon breakthrough): Successfully miniaturized the technology to run at room temperature using standard CMOS manufacturing processes.
Current Status (Lab scale): Tested a silicon chip with a few hundred probabilistic degrees of freedom that consumes only a few hundred attojoules per operation.
2025 Manufacturing: Plans to mass-produce a next-generation chip featuring millions of probabilistic degrees of freedom, representing a 1,000x year-over-year scaling increase.
The technology aims to match the brain's energy efficiency, theoretically offering a $10^8$ (100 million) times gain in AI efficiency compared to current GPU-based systems.

The platform utilizes "Energy-Based Models" (EBMs), a physics-based learning primitive that won the 2024 Nobel Prize in Physics.
The software stack includes high-level deep learning compilers designed to co-evolve with the probabilistic hardware.
Xtropic has developed algorithms that replace traditional diffusion models with "denoising thermo models," requiring 100x fewer steps to generate outputs.
Application areas extend beyond generative AI to include page ranking, recommender systems, biological simulations, robotics, trajectory planning, and optimization.
The team plans to open-source software to attract developers, acknowledging that current GPU algorithms are not optimized for probabilistic substrates.

2025 Deployment: The company expects to begin distributing early access devices to customers by the end of the summer.
2026 Deployment: Advanced applications, including hybrid diffusion models for image generation (similar to Midjourney), are scheduled to run on the hardware.
Market Prediction: XTropic anticipates that current multi-billion dollar GPU build-outs will be disrupted within five years as thermodynamic computing scales.
Risk Assessment: Investors are advised to factor in the risk of legacy GPU architectures failing to meet future energy constraints, suggesting a paradigm shift is imminent.
Scalability Thesis: If successful, the technology could enable AGI development by making intelligence per watt orders of magnitude higher, effectively removing energy as a bottleneck for civilization-scale AI.