THE POST
Thermodynamic chips will help decentralize compute again. The power will be in your hands. No more paying rent to the big model labs forever. https://t.co/kxaY2xwxUR
QUOTED
the tiny corp@__tinygrad__
The future depends on the direction this trend takes. https://t.co/ET5lxm8uTr
https://x.com/__tinygrad__/status/2109167903463686502The quoted post carries one image attachment. It is not copied into this file and Observatory did not read it, so this analysis does not say what trend it shows.
ENGAGEMENT
- likes379
- reposts26
- replies22
- quotes2
- bookmarks56
- views13,650
AS RETURNED BY THE X API · FETCHED 2026-10-13 05:21 UTC
WHAT IT SAYS
Quoting a tiny corp post about the direction of an unnamed trend, Beff predicts that thermodynamic chips will decentralize compute again, putting AI capability in individual hands instead of renting it from the big model labs. It is a forecast about hardware and market structure. It does not cite a benchmark, a shipping date or a power figure.
CONTEXT
Extropic is the company building thermodynamic chips, which it calls Thermodynamic Sampling Units (TSUs). Its Z1 chip, as described by Extropic, has 269,568 pbits across eight cores, samples at more than 50 MHz and draws under 1 W on a die under 12 mm per side. Extropic plans an M.2 stick with two Z1 chips, a PCIe card with 16, and a billion-pbit cluster, with early access in 2027. It says that for the right workloads its stack can be up to 10,000 times more energy efficient than GPUs. That figure is Extropic's own claim, and the workloads are probabilistic sampling tasks rather than general model inference.
On 29 July 2026 Extropic 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, demonstrate performance on generative-AI benchmarks, and fabricate a Z1.5 chip at a US foundry. Extropic notes that TSUs run on mature process nodes, so they would not compete for leading-edge wafers. The funding still depends on definitive agreements and milestones.
Beff has said on X that he is spending long days on Extropic milestones, so this is partly a founder describing where his own company is headed. That doesn't make it wrong, but it is a prediction from an interested party.
WHAT IT MEANS FOR K
Speculative for K. Intelligence per watt matters to the climb: if a class of AI work can run on chips that draw under a watt, the same harnessed power buys far more computation, and cheap local hardware spreads that capability widely. But no Z1 systems are in users' hands yet, the efficiency gains are company claims on selected workloads, and decentralization is a forecast, not a measurement.
CLAIMS
- Thermodynamic chips will help decentralize compute again (Beff, forecast).
- People will no longer have to pay rent to the big model labs forever (Beff, forecast).
- For the right workloads, Extropic's stack can be up to 10,000x more energy efficient than GPUs (Extropic, company claim).
EVIDENCE
- Extropic describes Z1 as taped out: 269,568 pbits, 16-neighbor connectivity, more than 50 MHz sampling, under 1 W.
- Extropic plans Z1 sticks, cards and clusters for early access in 2027.
- Extropic shipped dozens of XTR-0 desktop systems built on its earlier X0 chip to early adopters.
- Commerce letter of intent for up to $75M (29 Jul 2026) is non-binding and milestone-dependent.
REASONING
- Real silicon exists and a funded scaling plan is public, so this is more than a slogan.
- No independent benchmark of Z1 on mainstream AI workloads has been opened for this read, and broad availability is a 2027 plan.
- Whether compute decentralizes depends on software, models and markets as well as chips, so the verdict stays speculative until hardware ships and is measured.
SEE ON THIS SITE
TAGS
- THERMODYNAMIC COMPUTING
- EXTROPIC
- COMPUTE
- ENERGY EFFICIENCY
- E/ACC