Conference Presentation, Keynote
Guillaume Verdon, Extropic AI: Thermodynamic Computing and the Energy Efficiency Crisis of AI
- Energy demand will increase exponentially as training models become more costly, rendering current scaling approaches unsustainable for agentic chat and video models.
- The industry is projected to generate more heat from AI compute than the Earth currently radiates, culminating in a thermodynamic wall in the 2030s where transistor scaling fails.
- Algorithms will shift from single forward pass methods to probabilistic approaches utilizing test-time compute to capture the complexity of nature.
- Physical substrate development will focus on reverse engineering the physics underlying biology rather than mimicking biological forms to create the densest possible space, time, and energy substrate.
- Quantum computing is deemed too slow for the fundamental scaling of AI, prompting a pivot to thermodynamic hardware with probabilistic degrees of freedom.
- Manufacturing plans include launching a chip and board this fall containing millions of probabilistic degrees of freedom, with a target 1,000x year-over-year scale-up.
- A hybrid diffusion model using energy-based models (EBMs) is scheduled to run on this hardware in 2026, projecting a 10^8 energy efficiency gain at equivalent performance.
- Commercial deployment plans involve releasing initial customer devices by the end of summer, followed by larger devices for the subsequent year.
- Industry expectations include GPU build-out disruptions within five years, with market pricing expected to account for risks associated with half-trillion dollar infrastructure investments.
- Technology scaling is anticipated to progress significantly within five years, with the hardware designed to remain highly energy efficient regardless of input complexity.