Twelve quarters ahead, with the assumptions on the page
A forecast is worth exactly what its assumptions are worth, so they are printed beside the lines they produce and scored against a naive baseline on quarters we already know the answer to.
Forecast, backtest and momentum index, rebuilt weekly · rebuilt 2026-08-17 13:30 UTC ·
method
Cheapest on-demand H100 per GPU-hour right now:
AWS $6.88 · Azure $6.98 · GCP $4.20
(GCP's GPU SKUs exclude the host VM — see tracker methodology).
Cheapest-H100 moves since last snapshot: AWS +0.0%, Azure +0.0%, GCP +0.0%
2026Q1 cloud revenue: AWS $37.6B ·
Intelligent Cloud $34.7B ·
Google Cloud $20.0B; Google Cloud is growing roughly
twice as fast as the other two.
Combined parent capex hit $111B in 2026Q1 —
the AI buildout is still accelerating, not peaking.
If the last-8-quarter trends simply continued, Google Cloud's run-rate would cross AWS's around 2030Q1 — treat as an illustration of the growth gap, not a prediction.
If you are buying GPU compute today
Bursty training, price-sensitive: GCP has the lowest H100 list floor
($4.20/GPU-hr, GPU-only SKU — add host-VM cost);
combine with Azure's spot/preemptible tier, which carries the deepest median discount
(82%) if your jobs checkpoint well.
Latency / data-residency constrained: AWS offers GPU compute in the most
regions (94) — the widest footprint for compliance-bound workloads
(see the regional deep-dive for the China exception).
Negotiating leverage: regional price dispersion is real — the same GPU lists at
materially different prices across regions (tracker, dispersion chart). If your workload is
region-flexible, quote the cheapest region's floor in negotiations; list prices are ceilings,
not floors, at committed volume.
Generated from the latest weekly snapshot — every number above recomputes
automatically as prices move.
Revenue outlook
Assumptions, stated plainly. Log-linear extrapolation of the last
8 quarters: constant growth, no saturation, no competitive or macro response — useful for sizing
the growth gap, not for betting. Azure is fit only on FY25-basis quarters (the segment was
re-defined 2024Q3) and is a segment proxy, not Azure alone.
On these trends Google Cloud's quarterly revenue would cross AWS's around 2030Q1.
How good is this forecaster? (rolling-origin backtest)
Before trusting any extrapolation, test it on quarters we already know the answer to:
re-fit the model at six past origins and score out-of-sample errors against a naive
last-value baseline.
Provider
Horizon
Log-linear MAPE
Naive (last value) MAPE
AWS
1 quarter ahead
2.5%
5.1%
AWS
4 quarters ahead
4.4%
17.2%
Google Cloud
1 quarter ahead
5.5%
8.9%
Google Cloud
4 quarters ahead
7.5%
27.5%
MAPE = mean absolute percentage error across 6 rolling origins.
The model needs to beat naive convincingly at 4 quarters out to justify the fan chart above;
1-quarter-ahead is nearly unbeatable by anything (revenue is highly persistent).
AI Momentum Index
A documented composite built only from this project's own datasets — weekly GPU pricing
snapshots, SEC-filed capex, and segment revenue. Each component is normalized to the best
performer (=100) and weighted as shown; for frontier price, cheaper scores higher.
Frontier set: B200, B300, GB200, H100, H200, MI300X.
Component (weight)
AWS
Azure
GCP
GPU breadth (models) (20%)
18.0 (score 100)
13.0 (score 72)
10.0 (score 56)
Frontier reach (regions) (25%)
18.0 (score 39)
46.0 (score 100)
46.0 (score 100)
Frontier price ($/GPU-hr, median) (15%)
8.9 (score 90)
10.6 (score 76)
8.1 (score 100)
Capex acceleration (YoY %) (20%)
62.2 (score 69)
58.5 (score 65)
90.4 (score 100)
Cloud revenue growth (YoY %) (20%)
28.4 (score 45)
40.0 (score 63)
63.4 (score 100)
AI Momentum Index (0–100)
66.0
76.4
91.1
Index-design caveats: components correlate (capex buys frontier regions),
the price component inherits the bundling wedge (GCP's per-GPU SKUs exclude the host VM),
and breadth counts SKU variety, not installed capacity. Weights are judgment calls — the
table shows raw values so you can re-weight.