Observatory-K

X WATCH

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

THE POST

There is no rational reason to not invest at least ~$10B/year in alternative computing substrate research if we are going to spend a trillion a year on the current paradigm. Even if you assign 1% chance of disruption.

ENGAGEMENT

  • likes181
  • reposts7
  • replies22
  • quotes2
  • bookmarks8
  • views9,271

AS RETURNED BY THE X API · FETCHED 2026-11-01 06:17 UTC

WHAT IT SAYS

Beff argues that if roughly a trillion dollars a year is going into today's AI computing paradigm, it is irrational not to spend at least about $10 billion a year researching alternative computing substrates, even if the chance that one of them disrupts the current approach is only 1%.

CONTEXT

  • Fortune reports that Amazon, Alphabet, Microsoft, Meta and SpaceX are set to spend as much as $828 billion on capital expenditure in 2026, mostly on AI infrastructure, and that several banks see AI-related capex reaching $1 trillion in 2027. The spending fears helped knock more than $1 trillion off the Nasdaq 100 on Friday.

  • One example of the research he means: an Extropic paper on an all-transistor probabilistic computer estimates, from circuit measurements, physical models and simulation, that such a device could match GPUs on a simple image benchmark with about 10,000 times less energy. The authors say their energy model is approximate and that scaling beyond small datasets is still an open problem.

WHAT IT MEANS FOR K

NEUTRAL

This is an argument about how to allocate R&D money, not a change in harnessed power. Its relevance to K is indirect: cheaper intelligence per joule could let the same watts do more, and historically cheaper compute has raised total demand for power rather than lowered it.

CLAIMS

  • About a trillion dollars a year is going into the current AI paradigm.
  • At least $10 billion a year for alternative substrates is justified even at a 1% chance of disruption.

EVIDENCE

  • Fortune counts up to $828 billion of 2026 capex from five companies and cites bank forecasts of $1 trillion in 2027.
  • Extropic's paper gives a modelled, not yet measured at scale, 10,000x energy advantage on a narrow benchmark.

REASONING

  • The trillion-a-year scale is roughly supported for 2027 forecasts, a little ahead of 2026 figures.
  • The 1% case is an expected-value argument with no stated payoff, so it cannot be checked.
  • Nothing in the post adds or removes watts today, so the verdict is neutral.

SEE ON THIS SITE

TAGS

  • COMPUTE
  • EFFICIENCY
  • THERMODYNAMIC COMPUTING
  • AI CAPEX

← ALL X WATCH