Observatory-K

X WATCH

SPECULATIVE
Beff (e/acc)@beffjezos
ORIGINAL ON X ↗

THE POST

If you ask neo-cloud what they want, they say "Faster GPUs" People don't realize what they really want is thermodynamic computing: a probabilistic inference machine operating at the very physical limits of computation, both in terms of thermodynamic efficiency and speed. https://t.co/liYl2t35i2

QUOTED

Brendan Carr@BrendanCarrFCC

Frequently, we underestimate the significance of a technological innovation. We often assume that the next big thing will just be a better version of what we have today. This is not a new phenomenon. Henry Ford reportedly said that, if asked what they wanted, people would have said “faster horses.” Indeed, we called the first cars “horseless carriages.” In the future, these probably won’t be called “flying cars.”

https://x.com/BrendanCarrFCC/status/2111280971756573084

Quoted post includes a short video; Observatory did not copy the video file into this analysis.

ENGAGEMENT

  • likes33
  • reposts2
  • replies5
  • quotes0
  • bookmarks2
  • views3,191

AS RETURNED BY THE X API · FETCHED 2026-10-17 05:22 UTC

WHAT IT SAYS

The post says that when neoclouds are asked what they want, they say faster GPUs, but what they really want is thermodynamic computing: a probabilistic inference machine running near the physical limits of computation in both thermodynamic efficiency and speed. It quotes a post comparing flying cars to Henry Ford's 'faster horses' line about underestimating new technology.

CONTEXT

  • The account's X profile links to extropic.ai and calls its owner "thermo king @extropic". Extropic builds thermodynamic computing chips, so the post argues for the category the author's company sells. Extropic says its Thermodynamic Sampling Units use the natural thermal fluctuations of standard CMOS transistors to sample from programmable probability distributions, and that on key generative-AI workloads they can be orders of magnitude more energy-efficient than GPUs. In July 2026 it signed a non-binding letter of intent with the US Department of Commerce for up to $75 million in CHIPS R&D funding to bring the first Z1 clusters online and fabricate a Z1.5 chip at a US foundry.

  • The strongest published evidence is Extropic's own paper. It estimates that a device based on its architecture could match GPU-based generative models on binarized Fashion-MNIST using about 10,000 times less energy per generated sample. The authors say the experiments were simulations of the hardware on a classical computer, the GPU energy model is simplified, and the results are a design template rather than an optimized end-point. All authors work at Extropic.

WHAT IT MEANS FOR K

SPECULATIVE

Speculative. Better intelligence per watt does not add harnessed power by itself, but it decides how much useful work each new gigawatt buys, and cheaper compute tends to pull more power into use. The efficiency claims are so far simulations and company estimates on a simple benchmark, not measured results on production workloads.

CLAIMS

  • The post says neoclouds really want thermodynamic computing rather than faster GPUs.
  • It describes thermodynamic computing as operating near the physical limits of efficiency and speed.
  • Extropic says TSUs can be orders of magnitude more energy-efficient than GPUs on key generative-AI workloads.

EVIDENCE

  • Extropic's paper estimates about 10,000 times less energy per sample than GPU baselines on binarized Fashion-MNIST, from simulations, physical models and measurements of a test RNG chip.
  • The paper says its largest simulated model used about 50,000 nodes and that results should be read as a scaling picture, not an end-point.
  • Extropic's Commerce letter of intent is non-binding and funds work to bring first Z1 clusters online and demonstrate benchmarks.

REASONING

  • The post states a market preference without data on what neoclouds are buying.
  • The efficiency evidence comes from Extropic, the company the author's profile ties him to, and has not been shown at production scale.
  • Until cluster benchmarks and full-system power measurements exist, the effect on K is a plausible but unproven claim, so the verdict is speculative.

SEE ON THIS SITE

TAGS

  • COMPUTE
  • EFFICIENCY
  • THERMODYNAMIC COMPUTING
  • E/ACC
  • EXTROPIC

← ALL X WATCH