Extropic’s p-bit chips aim to slash AI energy use by 1000x using thermodynamic randomness

Extropic’s probabilistic computing approach aims to disrupt the AI chip market through revolutionary energy efficiency as the industry grapples with rising power demands. The startup is developing silicon chips that harness naturally occurring thermodynamic fluctuations in electronic circuits to perform calculations with probabilities, potentially reducing energy consumption by three to four orders of magnitude compared to today’s hardware. This innovation comes at a critical moment when AI companies are building datacenters near nuclear power stations and facing increasing scrutiny over their environmental impact.

The big picture: Extropic is developing a fundamentally different computing architecture that leverages randomness rather than fighting against it, positioning their technology as a potential Nvidia challenger in AI chip markets.

• The company’s “probabilistic bits” (p-bits) harness thermodynamic fluctuations that occur naturally in silicon circuits—phenomena traditionally considered problematic in conventional computing.
• Unlike previous approaches that required extreme cooling or superconducting materials, Extropic’s technology works with standard silicon manufacturing processes at normal temperatures.

Key technical breakthroughs: Extropic can control the probability of a bit being in either a 1 or 0 state and engineer interactions between these p-bits to perform complex probabilistic computations.

• The company has demonstrated a working p-bit on an oscilloscope, showing they can manipulate the probability states of these bits.
• By controlling these thermodynamic effects in conventional silicon, Extropic performs calculations without the extreme cooling requirements of quantum computing.

Why this matters: The approach could dramatically reduce AI’s growing energy footprint when companies are facing both technical limitations and environmental concerns about powering next-generation AI systems.

• Extropic claims its hardware is particularly well-suited for Monte Carlo simulations crucial in finance, biology, and building reasoning capabilities in AI models like OpenAI‘s o3 and Google’s Gemini 2.0 Flash Thinking.
• If successful, their chips could make a “sizable dent” in future emissions from AI computing infrastructure.

What they’re saying: “The world’s first scalable, mass-manufacturable, and energy-efficient probabilistic computing platform,” is how Guillaume Verdon, Extropic’s CEO, describes their innovation.

The competitive landscape: Challenging Nvidia’s dominance in AI chips might seem far-fetched, but Extropic’s founders believe the market is ripe for disruption.

• Nvidia currently dominates the AI acceleration market with its GPUs optimized for machine learning.
• The enormous energy demands of current AI infrastructure create an opening for radically more efficient approaches, especially as AI models continue to grow in size and complexity.