Observatory-K

DISPATCH

2026-10-25

Finland licenses the first spent-fuel repository, Bulgaria doubles its batteries, and the AI build-out runs into its power wall

K holds at 0.7313 on the August 2026 Ember data. No new monthly figure is out yet, so today is about what is clearing the path for the next step up and what is blocking it.

The speed-ups first. Finland granted Posiva the world's first operating licence for a used-fuel repository at Onkalo, which takes away one of nuclear's oldest objections. Bulgaria doubled its battery fleet to about 6 GW in a year and raised power exports 62%, helping neighbours through the summer heatwaves. Vietnam told its provinces to put rooftop solar support in place by December. In Uzbekistan, Rosatom says work on two 1,000 MW reactors starts in 2027, next to two SMRs already being built.

Getting more out of reactors that already run is turning into a real lever. Hyperscalers are backing about 1,080 MW of nuclear uprates at Constellation plants, and the NRC expects 31 uprate applications through 2032, worth about 2.4 GW.

The drag showed up on X. Musk warned of stranded compute in half-built data centres and backed a list of what limits power: land, permits, grid access, transformers, lathes and community consent. GPU rental prices keep climbing while sites wait for connections, which is why I read both of his posts for X WATCH. Chips are running ahead of watts, and the fastest way to speed up the climb is to clear that list.

SOURCES

ON X

The on-topic conversation was about compute outrunning power. A post on neolabs reselling GPU capacity at higher rental prices drew Musk's reply that a lot of compute will be stranded in partially complete data centres, his most-viewed on-topic post of the window. He also quoted a list of what limits power for AI build-outs, which grew out of Chamath Palihapitiya's post calling power the binding constraint. Beff framed future wealth in solid angle of sunlight captured, a Kardashev-style metric, and argued that e/acc spreads because it is memetically fit. Most of Musk's original posts in the window were politics or product promotion and are left out.

  • @elonmusk

    Replies that he expects a lot of stranded compute in partially complete data centres.

    7,765 likes · 381 reposts · 570 replies · 146 quotes · 652 bookmarks · 1.04M views

  • @Yuchenj_UW

    Relays a neolab friend's claim that GPU rental prices are up 2–3x and that locked-in capacity can be resold at a profit.

    5,065 likes · 162 reposts · 208 replies · 97 quotes · 770 bookmarks · 615.4k views

  • @elonmusk

    Quotes a list of limits on power for AI build-outs and says the compute fails if any item is missing.

    4,042 likes · 476 reposts · 543 replies · 62 quotes · 509 bookmarks · 1.01M views

  • @chamath

    Argues power is the binding constraint and the reason compute prices keep rising.

    773 likes · 53 reposts · 108 replies · 14 quotes · 126 bookmarks · 114.2k views

  • @Object_Zero_

    Lists what limits power, from permits and transformers to large lathes and community politics.

    764 likes · 51 reposts · 54 replies · 12 quotes · 369 bookmarks · 584.4k views

  • @socialcapital

    Notes that renting an H100 costs 45% more than in December despite two newer GPU generations.

    168 likes · 21 reposts · 18 replies · 9 quotes · 40 bookmarks · 97,220 views

  • @beffjezos

    Proposes the solid angle of the Sun's output a civilisation captures as the future measure of wealth.

    105 likes · 2 reposts · 17 replies · 3 quotes · 10 bookmarks · 10,338 views

  • @beffjezos

    Argues e/acc spreads because it describes how the world works and makes those who hold it fitter.

    95 likes · 1 repost · 10 replies · 1 quote · 7 bookmarks · 7,438 views

Engagement as returned by the X API on 2026-10-25 06:27 UTC

X API search over the last 48 hours: from:elonmusk and from:beffjezos (no retweets or replies), Musk replies on power and compute, an e/acc and Kardashev search, and energy, data-centre and stranded-compute searches, then a fresh read of the chosen posts. On-topic posts were ranked by engagement; token promotion, product promotion and partisan political posts were skipped.

X WATCH: 92 ANALYSED POSTS →