Are AI scaling laws dead?

Where the "scaling is dead" claim comes from

The skeptic case usually bundles three real observations: raw pre-training runs stopped delivering the dramatic jumps of the GPT-3→4 era; high-quality training data is increasingly scarce; and each additional order of magnitude costs enormously more. All three are genuine constraints. The question is whether they add up to "scaling laws are dead" — and the tracked evidence says they don't.

What the scorecard's data shows

This scorecard grades the underlying prediction — that effective compute keeps scaling at roughly half an order of magnitude per year — as On track. The trend was never only about bigger pre-training runs: it combines physical compute, algorithmic efficiency, and unhobbling. The mix has shifted hard toward the latter two, which is a change in composition, not a death.

The steelman — and what would actually confirm it

The strongest version of the skeptic case is economic: capex has exceeded projections while revenue lags the spend, and a funding pullback would slow the compute trend for financial rather than technical reasons. That is the credible path to "scaling ends." What would confirm the claim on this scorecard: a sustained, multi-year drop below the ~0.5 OOM/yr pace, or capability benchmarks flatlining despite continued compute growth. Neither has happened as of mid-2026.

Why the answer matters

The entire OOM argument behind AGI-by-2027 is an extrapolation of these curves. If scaling truly broke, the 2027 case would collapse with it. So far the curves have held — which is precisely why the AGI-by-2027 verdict stays Open rather than Wrong, even with autonomous AI research still undemonstrated.

Frequently asked questions

Not on the tracked evidence. Through mid-2026, effective compute has roughly held its ~0.5 OOM/yr pace and capability benchmarks kept climbing. The gains have shifted from raw pre-training scale toward reasoning, tools, and agents — a change in composition, not an end to scaling.

Because raw pre-training jumps have shrunk, quality data is scarcer, and each order of magnitude costs far more. Those constraints are real, but the combined effective-compute trend — hardware plus algorithms plus unhobbling — has still roughly held.

A sustained multi-year drop below the ~0.5 OOM/yr effective-compute pace, or capability benchmarks flatlining despite continued compute growth. A capex pullback driven by the revenue gap is the most credible route there.

It collapses — the 2027 forecast is an extrapolation of these curves. That's why this scorecard tracks the compute verdict so closely: it is the load-bearing input for the headline claim.