AI demand,
forecast before it happens.
Every week we publish a dated, tamper-proof forecast of how much AI the leading apps will use over the next eight weeks, then score it in public against the standard forecasts and against GPT-6 Astra on its own. Our forecast combines the WORKCAST simulation with GPT-6 Astra and a statistical trend model. Misses stay up.
This week's call
Track record
Held-out test: the combined method
Weights fixed on earlier weekly cutoffs, then scored once on 7 new weekly cutoffs (2026-09-07 to 2026-10-19), all after GPT-6 Astra's knowledge cutoff: the lowest mean error of every method (23.9%). At 1 and 2 weeks ahead the damped trend was marginally ahead. Eight-week results come from the live board.
Every method on the held-out weeks
Token-weighted error across the leading apps, lower is better. Cutoffs 2026-09-07 to 2026-10-19; the 1, 2 and 4 week columns cover 7, 6 and 4 cutoffs. Every method sees only data published before each cutoff.
| Method · what it sees before the cutoff | 1 wk | 2 wk | 4 wk | Mean |
|---|---|---|---|---|
| WORKCAST combined (our forecast) | 23.0% | 25.6% | 23.0% | 23.9% |
| Timestral, adjusted by GPT-6 Astra | 22.9% | 25.6% | 23.6% | 24.0% |
| Damped trend | 23.0% | 25.4% | 23.9% | 24.1% |
| GPT-6 Astra alone | 23.7% | 26.5% | 24.0% | 24.7% |
| WORKCAST simulation alone | 23.6% | 27.0% | 24.0% | 24.9% |
| Timestral alone | 23.7% | 27.5% | 27.3% | 26.2% |
| Recent trend | 30.4% | 29.6% | 28.1% | 29.4% |
| No change | 28.4% | 30.2% | 34.7% | 31.1% |
| Market growth | 31.9% | 32.8% | 33.7% | 32.8% |
Backtest: the simulation on its own
16 cutoffs four weeks apart, 2025-07-07 to 2026-08-31, 1,600 forecasts. These weeks fall before GPT-6 Astra's training cutoff, so Astra and the combined method are not scored here. The simulation alone beats the three simple standards at every horizon, but the damped trend beats it at every horizon, with either history length. That is why the published forecast blends the simulation with Astra and the damped trend.
| Method · what it sees before the cutoff | 1 wk | 2 wk | 4 wk | 8 wk |
|---|---|---|---|---|
| Damped trend, full history | 22.2% | 25.4% | 31.2% | 42.6% |
| Damped trend, last 12 weeks | 22.5% | 25.5% | 31.2% | 42.1% |
| WORKCAST simulation alone | 25.8% | 29.0% | 34.8% | 43.6% |
| No change | 32.7% | 35.0% | 39.7% | 45.1% |
| Recent trend | 33.1% | 35.7% | 44.4% | 65.8% |
| Market growth | 34.3% | 36.8% | 43.2% | 53.1% |
How it works
Our forecast blends three forecasts with weights fixed per horizon on earlier weeks: the WORKCAST simulation, GPT-6 Astra and the damped trend. At each Monday cutoff, WORKCAST carries every leading app forward 8 weeks from data before the cutoff: its growth path, whether it stays among the leaders, and the models its workloads move to as new models launch and prices change. The simulation range is the 5th to 95th percentile of 20 runs. GPT-6 Astra reads each app's recent weeks and returns its own forecast. The damped trend is the standard statistical forecast for series like these (Holt's method): it projects each app's current level and growth, and lets the growth fade a little each week, capped at 15% a week. Error is token-weighted across apps (lower is better).
How to verify a forecast
Each forecast is a read-only JSON record carrying the code version and a hash of every input table. Its SHA-256 is anchored in the Bitcoin blockchain with OpenTimestamps before the weeks it predicts. Download a record and its .ots proof, then run ots verify. A forecast published more than two days after its cutoff is skipped, never backdated.