AI demand, forecast in advance
and scored in public.
Every week we publish a dated, tamper-evident forecast of how much AI the leading apps will use over the next eight weeks, then score the forecast in public against the standard methods and against a large language model alone. Our forecast blends the WORKCAST simulation, a large language model and a statistical trend model. Misses stay up.
Accuracy is 100% minus token-weighted error across the leading apps. The axis zooms in on the range shown and does not start at zero.
What we forecast, against what happened
The quantity we forecast: what the 100 leading apps used and spent at list prices over each 8-week window, 7 Jul 2025 to 31 Aug 2026, next to the forecast made at the start of the window from data before it. Spend: the simulation missed by $9.8M a window on average (16%); the best standard forecast, no change, by $18.9M (30%). Tokens: the simulation missed by 23.1T a window on average (28%); the best standard forecast, market growth, by 20.3T (24%). Simulation alone: these windows start before the language model's training data ends, so the blend is not scored here.
Table view
| Method · what the method sees before each cutoff | Average miss | Share of actual |
|---|---|---|
| Market growth | 20.3T | 24.3% |
| WORKCAST simulation alone | 23.1T | 27.6% |
| Recent trend | 29.1T | 34.8% |
| No change | 29.6T | 35.4% |
| Method · what the method sees before each cutoff | Average miss | Share of actual |
|---|---|---|
| WORKCAST simulation alone | $9.8M | 15.7% |
| No change | $18.9M | 30.3% |
| Market growth | $20.2M | 32.3% |
| Recent trend | $26.3M | 42.0% |
This week's call
The full record
Weights fixed on earlier weeks, then scored once on 7 new weekly cutoffs (7 Sep 2026 to 19 Oct 2026), all after the language model's training cutoff. Highest average accuracy of all 9 methods (76.1%). At 1 and 2 weeks ahead another method was marginally ahead. The 1, 2 and 4 week columns cover 7, 6 and 4 cutoffs. The two Timestral rows were tested later on the same weeks. Eight-week results come from the live board.
The 18 weekly cutoffs (4 May 2026 to 31 Aug 2026) used to choose designs and fit the combined method's weights, so these scores flatter the combined method. Context only, not proof.
16 cutoffs four weeks apart, 7 Jul 2025 to 31 Aug 2026, 1,600 forecasts. These weeks reach back before the language model's training cutoff, so the language model and the combined method are not scored here. The simulation alone beats the three simple standards at every horizon, but the damped trend beats the simulation at every horizon, with either history length. That result is why the published forecast blends the simulation with the language model and the damped trend.
| Method · what the method sees before each cutoff | 1 wk | 2 wk | 4 wk | Average |
|---|---|---|---|---|
| WORKCAST combined (our forecast) | 77.0% | 74.4% | 77.0% | 76.1% |
| Timestral, adjusted by a language model | 77.1% | 74.4% | 76.4% | 76.0% |
| Damped trend | 77.0% | 74.6% | 76.1% | 75.9% |
| Large language model alone | 76.3% | 73.5% | 76.0% | 75.3% |
| WORKCAST simulation alone | 76.4% | 73.0% | 76.0% | 75.1% |
| Timestral alone | 76.3% | 72.5% | 72.7% | 73.8% |
| Recent trend | 69.6% | 70.4% | 71.9% | 70.6% |
| No change | 71.6% | 69.8% | 65.3% | 68.9% |
| Market growth | 68.1% | 67.2% | 66.3% | 67.2% |
| Method · what the method sees before each cutoff | 1 wk | 2 wk | 4 wk | 8 wk | Average |
|---|---|---|---|---|---|
| WORKCAST combined (our forecast) | 81.2% | 78.1% | 73.3% | 64.8% | 77.6% |
| Timestral, adjusted by a language model | 81.3% | 77.6% | 72.3% | 64.4% | 77.1% |
| Damped trend | 80.3% | 77.8% | 72.8% | 63.5% | 77.0% |
| Large language model alone | 81.0% | 77.1% | 72.0% | 64.3% | 76.7% |
| WORKCAST simulation alone | 77.3% | 74.4% | 67.9% | 56.0% | 73.2% |
| Recent trend | 75.1% | 72.1% | 66.4% | 55.9% | 71.2% |
| Timestral alone | 79.1% | 72.8% | 60.9% | 29.4% | 70.9% |
| Market growth | 70.3% | 67.7% | 63.5% | 56.9% | 67.2% |
| No change | 70.6% | 67.9% | 62.9% | 53.3% | 67.1% |
| Method · what the method sees before each cutoff | 1 wk | 2 wk | 4 wk | 8 wk | Average |
|---|---|---|---|---|---|
| Damped trend, full history | 77.8% | 74.6% | 68.8% | 57.4% | 73.8% |
| Damped trend, last 12 weeks | 77.5% | 74.5% | 68.8% | 57.9% | 73.6% |
| WORKCAST simulation alone | 74.2% | 71.0% | 65.2% | 56.4% | 70.1% |
| No change | 67.3% | 65.0% | 60.3% | 54.9% | 64.2% |
| Recent trend | 66.9% | 64.3% | 55.6% | 34.2% | 62.3% |
| Market growth | 65.7% | 63.2% | 56.8% | 46.9% | 61.9% |
Week by week
Each point is the accuracy at one weekly cutoff; higher is better. The leading methods trade places from week to week, so a small gap in the average is not decisive. Move across the chart for every method's accuracy, or click a name to hide a line.
Table view
Forecast against actual
Held-out weeks only. Every dot is one app. Dots on the line were forecast exactly; the further from the line, the bigger the miss. Both axes are on a log scale. Apps that fell outside the weekly top 200 are counted at half the 200th app's volume, which makes the flat row along the bottom. Switch the method or the week and watch the dots move.
Table view
How the forecast works
Our forecast blends three forecasts with weights fixed per horizon on earlier weeks: the WORKCAST simulation, a large language model and the damped trend. At each Monday cutoff, WORKCAST carries every leading app forward 8 weeks from data before the cutoff: the app's growth path, whether the app stays among the leaders, and the models the app's workloads move to as new models launch and prices change. The simulation range is the 5th to 95th percentile of 20 runs. The language model reads each app's recent weeks and returns a forecast of the same quantities. The damped trend is the standard statistical forecast for series like these (Holt's method): the method projects each app's current level and growth, and lets the growth fade a little each week, capped at 15% a week. Accuracy is 100% minus the token-weighted error across the leading apps, so higher 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. The record's SHA-256 is anchored in the Bitcoin blockchain with OpenTimestamps before the weeks the forecast predicts. Download a record and the .ots proof, then run ots verify. A forecast published more than two days after the cutoff is skipped, never backdated.