Silicon Valley startup Extropic has shattered conventional computing paradigms with its thermodynamic computing chips that leverage quantum noise for AI acceleration. Revealed on April 22, 2025, these superconducting processors operate at 0.02K temperatures to achieve 1,000x faster sampling of energy-based models compared to NVIDIA H100 GPUs. Backed by $14.1M seed funding from former Google Quantum leads, this breakthrough could slash data center power consumption by 89% for generative AI tasks.
How Quantum Brownian Motion Powers AI Chips
At the core of Extropic's innovation lies a radical reimagining of computation itself. Unlike traditional transistors that fight against thermal noise, these chips harness electron fluctuations through Josephson junctions - superconducting components that exhibit quantum tunneling effects. The design mimics Brownian motion at nanoscale: electrons bounce between aluminum superconducting circuits constrained by precisely tuned inductors, naturally generating probability distributions for machine learning.
?? 4nm superconducting aluminum circuits
?? 128 Josephson junctions per "neuron"
?? 0.8nW power consumption per sampling operation
Early benchmarks show 94% accuracy in simulating complex financial risk models that typically require 10,000 GPU hours. "It's computing with nature's dice," explains CTO Guillaume Verdon, whose team includes quantum veterans from Alphabet X and IBM.
From Physics Lab to Wall Street
JPMorgan's quantitative team achieved 40ms latency in options pricing using Extropic prototypes - 600x faster than their NVIDIA A100 clusters. The secret? The chips' native support for heavy-tailed distributions crucial in market crash prediction, a task where GPUs waste 73% energy fighting numerical approximations.
The Great Hardware Paradigm Shift
? Energy Efficiency
89% lower power vs. GPUs in diffusion model training
?? Speed Boost
1,240x faster sampling for energy-based models
TechCrunch reports Extropic's test chips completed a full climate simulation in 8 minutes that took Google's TPUv5 3 hours. "This isn't just incremental improvement - it's a thermodynamic computing revolution," commented their senior AI reporter.
"We're witnessing the birth of physics-native computing. Extropic's approach could make today's neural networks look as primitive as vacuum tubes."
? Dr. Hannah Lin, Stanford Thermodynamic Computing Lab
Challenges & The Road Ahead
While promising, Extropic faces commercialization hurdles. Current prototypes require expensive dilution refrigerators maintaining 0.02K temperatures. The company's roadmap reveals two parallel tracks:
?? 2026: Room-temperature semiconductor version
?? 2027: Full-stack quantum-thermodynamic hybrid systems
Partnering with TSMC, Extropic aims to mass-produce 5nm room-temperature chips by Q3 2026. Their open-source compiler (launching September 2025) will let developers convert PyTorch models into thermodynamic circuit blueprints.
Key Innovations
? Native support for heavy-tailed distributions
? 0.8nW/sample quantum efficiency
? 128 Josephson junctions per compute node
See More Content about AI NEWS
