THE POST
We're reaching the top of the GPU cycle s-curve it's urgent we scale thermo compute to start the next 5 OOMs of scaling that would be physically impossible simply scaling the current paradigm and grid it's time to scale @extropic https://t.co/uODQtd4mR5
QUOTED
Wall St Engine@wallstengine
ANTHROPIC IPO VALUATION EXPECTATIONS DROP TO ~$1.5T FROM $2T: RTRS Some existing and prospective investors are reconsidering the Claude maker's valuation ahead of its anticipated November IPO, according to Reuters. Anthropic reported Q3 revenue of $16.8B, up 45% QoQ, compared with growth of more than 100% in Q2, when revenue reached $11.6B. Despite reporting adjusted profitability, the slowdown disappointed some investors, with a broader tech selloff and concerns over rising AI infrastructure costs adding pressure. The company has committed to spending at least $518B on cloud computing and infrastructure over the next decade. Source: Reuters.
https://x.com/wallstengine/status/2110838686061564068Text is the full long-form text the X API returned for the quoted post. It carries one photo, which is not copied here. Observatory did not open the underlying Reuters report.
ENGAGEMENT
- likes176
- reposts7
- replies6
- quotes0
- bookmarks17
- views11,355
AS RETURNED BY THE X API · FETCHED 2026-10-20 05:22 UTC
WHAT IT SAYS
Beff, who founded Extropic, says GPU scaling is near the top of its S-curve and that the next five orders of magnitude of compute can't come from more GPUs and more grid alone. His answer is to scale thermodynamic computing, and Extropic with it. He quotes a post relaying a Reuters report that investors are trimming Anthropic's expected IPO valuation as revenue growth slows and AI infrastructure costs rise.
CONTEXT
Extropic says it builds thermodynamic sampling units for probabilistic AI workloads and calls them radically more energy efficient than GPUs. It describes Z1, its first production-scale chip, as holding over a quarter million probabilistic bits on a die under 12 mm a side and drawing under one watt. Those are company claims; I found no independent benchmark of Z1 against GPUs on a real AI workload.
GPU efficiency is still climbing, not flat. The June 2026 Green500 leader, KAIROS at CALMIP in Toulouse, runs on NVIDIA GH200 superchips at 73.282 GFlops per watt. On this site's Green500 series that is 22.8 times the June 2013 leader, about 27% a year. By Observatory arithmetic, five orders of magnitude at that pace would take roughly 48 years, so Beff's point that today's curve alone won't deliver 100,000x soon holds up.
The grid half of the claim is visible right now. Microsoft reportedly walked away from anchoring Nscale's Loughton data centre in Essex after the site was told not to expect power until the 2030s, and Nscale's Norway site runs 6.4 MW of a planned 30 MW.
WHAT IT MEANS FOR K
Speculative. More compute per watt does not raise K by itself, since K counts the power we harness. But it changes what each watt is worth, and when grid connections are the bottleneck it lets more useful work happen inside the power we already have. If thermodynamic chips deliver even part of what Extropic claims, that is a real accelerant. There is no public evidence yet that they do at scale.
CLAIMS
- GPU scaling is near the top of its S-curve.
- The next five orders of magnitude of compute are physically impossible by scaling GPUs and the grid alone.
- Thermodynamic computing, and Extropic in particular, should be scaled now.
EVIDENCE
- Green500: the June 2026 leader reaches 73.282 GFlops/W, 22.8 times the June 2013 leader on this site's series.
- TNW: Loughton was told not to expect grid power until the 2030s; Microsoft reportedly pulled out.
- Extropic: Z1 holds over 250,000 probabilistic bits on a die under 12 mm a side at under one watt (company claim, no independent benchmark found).
REASONING
- The grid constraint is documented and the GPU efficiency trend, while steady, is far too slow for 100,000x in the near term by Observatory arithmetic.
- The proposed fix comes from the author's own company and has no published head-to-head results on production AI workloads.
- A credible problem with an unproven solution is speculative under this site's rules.
SEE ON THIS SITE
TAGS
- COMPUTE
- THERMODYNAMIC COMPUTING
- EFFICIENCY
- GRID
- EXTROPIC