AI’s hunger for power is outpacing common sense. If current trends hold, energy use could rival dozens of reactors by the end of the decade. I argue that the answer is not more megawatts, but a new way to compute. We should stop forcing machines to fake randomness and start using the real thing.
A seasoned chip designer recently laid out a case for hardware that treats noise as fuel, not a flaw. Their thesis is simple and sharp: modern AI is probabilistic at heart, so our hardware should be, too. I agree. The promise is huge, and the risks are clear. Both deserve attention now.
The Case for Thermodynamic Computing
Digital chips were built for certainty. They burn energy to keep ones and zeros clean, even when the task is to sample from a probability curve. Generative models do not seek a single fixed answer. They weigh options and pick one. That is controlled randomness, and we pay dearly to simulate it.
“What if the way we compute is fundamentally wrong?”
Physics tells us that information has an energy cost. Landauer’s idea still cuts deep: erase a bit and you add entropy. AI piles on more bits than ever. So the old plan—fight noise with more power—hits a wall.
Extropic proposes a different path. Run transistors in a “relaxed” region where thermal noise nudges them across an energy barrier. Each device becomes a probabilistic bit, or P‑bit, that flips between zero and one. Nudge the voltage and you shape the odds.
“They lean into this instability… and use it for computation.”
This flips the script. Instead of spending energy to imitate randomness, the chip harvests it from physics. Connect many P‑bits with weighted links and the system samples from a Boltzmann distribution. It spends most time in low-energy states—the good solutions—and only sometimes climbs to higher ones.
What The Numbers Say—And Don’t Say
The headline claim is eye-catching: up to 10,000 times better energy efficiency than top GPUs. If true at scale, the economics of AI change overnight.
“If that holds, it changes the whole economics of AI.”
I like the direction, but we need to separate promise from proof. The strongest results so far come from small tests such as image generation and simulations. That is progress, not a verdict. Still, there are clear advantages: this approach uses standard CMOS, runs at room temperature, and needs no exotic cooling. That makes scaling at least plausible.
There are also real hurdles. Analog coupling can corrupt randomness when many P‑bits sit close together. Software is another mountain. The entire AI stack assumes deterministic hardware. CUDA took years to mature. A new stack must win mindshare fast while GPUs keep improving.
- Physics advantage: randomness “for free,” not simulated.
- Silicon reality: standard processes, no cryogenics.
- Scaling risk: noise you want vs. noise you don’t.
- Tooling gap: new algorithms, new abstractions.
- Incumbents sprint: GPUs get leaner every year.
These points can coexist. The idea can be right and still hard to ship. That is why independent testing and open benchmarks matter now.
Where This Hardware Belongs
This will not replace deterministic computing. Banking, flight control, and medical devices need certainty. But the probabilistic world is large and growing. It includes the heaviest AI workloads we run today.
- Generative AI inference and sampling
- Energy‑based models and optimization
- Anomaly detection and search
- Monte Carlo and scientific simulations
- Self‑supervised learning loops
For these tasks, computers that ride entropy can beat computers that fight it. The chip designer put it well:
“The most powerful computers of the future won’t fight entropy and uncertainty. They will run on it.”
My Take
I believe this path deserves serious trials in real workloads, not just labs. Start with inference at the edge and data center pilots where power is tight and latency matters. Demand apples‑to‑apples energy per token, per sample, and per solved instance. Reward wins with procurement, not just applause.
And one more thing: do not underestimate the software climb. The winners will pair silicon with a clean developer story—libraries, converters, and training wheels for teams steeped in PyTorch and CUDA.
Stop treating noise as a tax. Treat it as a resource. If Extropic’s Z1 delivers even a slice of its promise, AI’s power curve can bend. If not, we still gain sharper tools for sampling and the best hardware random source we’ve ever seen. Either way, the experiment is worth it.
It’s time to fund open benchmarks, back pilot deployments, and set energy targets that reward real efficiency. Let’s measure what matters and push computing to match the math our models already live by.
Frequently Asked Questions
Q: How is a probabilistic bit different from a normal transistor?
A standard transistor is forced into strict on/off states. A probabilistic bit operates near a thermal threshold, flipping between states with a controllable probability.
Q: Why could this cut AI’s energy use?
Many AI tasks require sampling. Today we simulate randomness with lots of operations. Thermodynamic chips tap physical noise directly, reducing wasted work.
Q: What workloads are the best early fit?
Generative inference, optimization, anomaly detection, and Monte Carlo methods benefit first. They rely on drawing many samples rather than producing one fixed answer.
Q: What are the main technical risks?
Scaling introduces unwanted coupling in analog networks and can bias results. Keeping “good” noise while suppressing the rest at large sizes is tough.
Q: How would developers actually use this?
They will need new libraries and abstractions that map models to sampling hardware. Conversion tools from common frameworks will be key to adoption.
