According to Wired, Extropic has developed its first working thermodynamic sampling units (TSUs), a new type of processor that uses probabilistic bits (p-bits) instead of conventional binary bits. The startup, founded by former Google quantum computing researchers Guillaume Verdon and Trever McCourt, claims its approach could be thousands of times more energy efficient than current chips from Nvidia and AMD when scaled up. The initial hardware, called XTR-0, combines field programmable gate arrays with the company’s first probabilistic chips, and has been shared with select partners including frontier AI labs, weather modeling startups, and government representatives. Extropic is also releasing TRHML software that simulates chip behavior on GPUs, allowing developers to test applications before accessing physical hardware.
Table of Contents
The Thermodynamic Computing Revolution
Extropic’s approach represents a fundamental shift from deterministic computing that has dominated the industry for decades. While traditional processors rely on precise binary states (1s and 0s), thermodynamic computing embraces the inherent randomness and uncertainty in physical systems. This isn’t just a different architecture—it’s a different computational paradigm that aligns with how many real-world problems actually behave. The technology leverages thermodynamic fluctuations at the electron level, essentially turning what’s typically considered noise in conventional computing into the core computational mechanism.
Addressing AI’s Unsustainable Energy Appetite
The timing couldn’t be more critical given the explosive growth in AI computational demands. Current large language models require massive data centers consuming power equivalent to small cities, creating both economic and environmental sustainability challenges. Extropic’s claimed efficiency improvements—if realized at scale—could fundamentally alter the economics of AI deployment. Rather than building ever-larger data centers, companies might achieve similar results with dramatically smaller physical footprints and energy consumption. This becomes particularly important as AI models grow more complex and computational requirements continue their exponential trajectory.
The Scaling Challenge and Technical Hurdles
While the concept is promising, the path to commercial viability faces significant obstacles. The current XTR-0 system contains only “a handful of qubits,” which represents an extremely limited scale compared to the billions of transistors in modern CPUs and GPUs. Scaling probabilistic systems while maintaining coherence and managing error rates presents formidable engineering challenges. Additionally, the entire software ecosystem for probabilistic computing needs development from the ground up—existing AI frameworks and programming models built around matrix multiplication won’t directly translate to this new architecture.
The Controversial Foundation
Extropic’s leadership brings both technical credibility and philosophical baggage. CEO Guillaume Verdon’s background as “Based Beff Jezos” and his association with the effective accelerationism movement adds an unconventional dimension to what’s typically a conservative hardware development space. While his quantum computing experience at Google provides technical validation, the e/acc philosophy’s controversial stance on rapid technological development without constraints could influence both investor perception and regulatory scrutiny. In the cautious world of hardware investment, this philosophical positioning represents both a branding opportunity and potential liability.
Potential Market Disruption Scenarios
If Extropic succeeds, the implications extend far beyond just another startup entering the AI hardware market. The company is targeting specific applications where probabilistic modeling provides natural advantages—weather prediction, financial modeling, drug discovery, and generative AI. These are domains where uncertainty is inherent and traditional deterministic computing struggles with complexity. Rather than competing directly with Nvidia across all AI workloads, Extropic could carve out specialized high-value niches where their thermodynamic approach provides unbeatable advantages in both performance and efficiency.
The Long Road to Commercialization
History suggests revolutionary computing architectures face decade-long journeys from proof-of-concept to mainstream adoption. Even if Extropic’s technology works as promised, the company must navigate manufacturing scalability, software ecosystem development, and customer education. The current strategy of partnering with select organizations for testing is smart—it allows real-world validation while building use cases. However, the ultimate test will be whether Extropic can transition from demonstrating isolated capabilities to running production-scale AI models that compete with established hardware on both performance and total cost of ownership.
