This is a very interesting take on why AI gobbles up excessive power, making the case that it does not need to do so.
__________
Christian Schmidt
Why do GPUs burn 10,000x more energy than they should for generative AI?
Because we built algorithms to fit hardware—not the other way around.
Extropic just proved the opposite works better.
━━━━━━━
The Pattern Everyone Misses:
GPUs dominate AI because they crushed gaming graphics a decade ago.
Neural networks happened to run well on them.
Since then: every AI breakthrough optimized for GPU architecture.
▸ LLMs? Designed for matrix multiplication.
▸ Diffusion models? Retrofitted to fit CUDA cores.
▸ Sampling tasks? Faked using deterministic math.
The Real Cost:
Digital systems spend massive energy generating pseudo-randomness.
They suppress thermal noise—the very thing probabilistic tasks need.
Quantum chips go further: cool to near absolute zero to avoid noise.
Both waste energy fighting physics.
━━━━━━━
What Extropic Actually Built:
✅ Thermodynamic Sampling Units (TSUs) that sample probability distributions natively using thermal electron fluctuations.
✅ Denoising Thermodynamic Model (DTM)—a generative AI algorithm purpose-built for TSUs, not retrofitted from GPUs.
✅ XTR-0 development platform already beta-tested and shipping to researchers.
✅ Open-source Python library (`thrml`) so developers can prototype algorithms before commercial chips arrive.
The Efficiency Gain:
Simulations show 10,000x better energy efficiency than GPUs on small benchmarks.
Standard semiconductor processes—no exotic materials or cryogenics.
━━━━━━━
Here’s the fix most won’t see:
Stop asking “How do we make GPUs faster?”
Start asking “What hardware fits the algorithm?”
Extropic bet on energy bottlenecks 3 years ago.
They were right.
━━━━━━━
I used advanced research tools to break this down in under 7 minutes.
Extropic’s CEO Guillaume Verdon: “harness stochastic thermodynamics to execute certain algorithms.”
WIRED: chips that achieve “energy efficiencies thousands of times greater than current models.”
The academic paper is public.
Independent replication code is open-source.
━━━━━━━
👉 Action tool: News analysis on steroids
This is my unfair advantage
⏳ https://lnkd.in/daEhWsh6
🔥 Save/Share now—most people won’t see this again.
♻️ Repost if this changed your view.
➕ Follow me Christian Schmidt
━━━━━━━
Sources:
https://lnkd.in/dDaSsGN4
https://lnkd.in/dEAuMY4H
https://lnkd.in/d2_skR3x
