LivetechWestern cycleAnchors pending

Compute-doubling-time (frontier training-run FLOPS) saturates?

As of 2 Oct · 23:18 UTCResolves 31 Dec 20280 movers0 pressure moments
Timeline A · Timeline AReality
60%
Crowd 55%Gap 0

By end of 2028, the rate at which frontier AI training runs double in computing power slows sharply—crossing from months between doublings to a year or more between doublings. Nvidia's chip roadmap (Hopper to Blackwell to Rubin) stays on schedule, but electricity grids, water supply, and chip fabrication capacity hit hard walls. OpenAI, Meta, and other labs still launch larger models, but the gap between successive training runs widens. The computing frontier doesn't stop growing; it just stops accelerating.

Timeline B · Frontier Compute-doubling ContinuesReality
40%
Crowd 45%Gap 0

Frontier compute-doubling continues every 6–9 months through 2028 and beyond. New data centers come online faster than power constraints tighten; Nvidia ships Rubin chips on cadence; the Stargate partnership and Meta's Prometheus cluster unlock enough electricity and manufacturing headroom that labs can sustain rapid scaling. Each new generation of training runs stays within the historical speed of doubling, and the exponential curve stays steep.

crowd 55
The quiet paththe one nobody’s watching

A grid collapse or regional power crisis in 2026–2027 could impose hard regulatory limits that make the doubling-time question irrelevant.

6%
The read

Where reality is heading

A, 80% bandCrowd today
Projected reality support to 31 Dec 2028A rises from 41% to about 52% with an 80% band of 33 to 71. B falls from 53% to about 44%. Quiet path near 4 to 6%. Crowd today: A 55%.100%7550250Crowd today · A 55%A ~52%B ~44%P ~4%NowJan 2027Jan 202831 Dec 2028
Reality crosses toward A in early 2028. Today’s crowd number sits at the top of A’s band, possible but optimistic. View as table

The movers, ranked by leverage

What would change our mind

Signals toward A
    Signals toward B

      Evidence