Is AI compute still scaling? (the OOM bet, graded)

The bet

A core driver of the AGI-2027 forecast is that effective compute — raw compute combined with algorithmic efficiency gains — keeps climbing at roughly half an order of magnitude per year, sustained over the decade. If that curve breaks, the whole timeline slips.

The evidence so far

As of mid-2026 the curve has held up well. Nathan Delisle's quantitative one-year audit (June 2025) checked the math driver-by-driver and found the pace "roughly supported," with individual launches scattered within about ±0.5 OOM of the trend. Epoch AI's tracking is broadly consistent. This is one of the predictions where Aschenbrenner bet on a trend and the trend obliged.

What would flip it

The pre-registered condition: this moves to Wrong if the next frontier generation lands more than 1 OOM below the trend line, or if a major lab publicly abandons the scaling thesis. Neither has happened.

Why it matters

Compute scaling is the engine under the hood of every downstream claim — capability, AGI timing, the intelligence explosion. Its holding is a big reason the aggressive end of the forecast distribution (where Aschenbrenner sits) hasn't been ruled out.

Frequently asked questions

Yes. Aschenbrenner predicted effective compute would scale ~0.5 OOM/yr, and as of mid-2026 the pace has roughly held — an independent audit calls it 'roughly supported.' The prediction is graded On track.

Not as of mid-2026. Effective compute (raw compute plus algorithmic efficiency) has continued climbing near the ~0.5 OOM/yr pace Aschenbrenner projected, with launches scattered ±0.5 OOM around trend.

It flips to Wrong if the next frontier generation lands more than 1 order of magnitude below the trend line, or if a major lab publicly abandons the scaling thesis.