Methodology
What is measured, what is modeled, and what is — for now — synthetic. Stating this explicitly is the point.
Data provenance (current build)
Market series (supply, demand, balances, prices): synthetic — deterministic formulas shaped to be directionally realistic, for layout and demonstration. They are not measured market data and must not be read as such.
Timeline events: real, curated facts with month precision; figures as publicly reported at the time. See Timeline.
H100-equivalence factors: illustrative placeholders pending benchmark-derived values (MLPerf-class benchmarks, precision-weighted, utilization-degraded). Definition: Definitions › H100-equivalent.
Full methodology — sources per series, tier-classification rules, scenario calibration, and the modeled-vs-measured boundary per chart — lands as real data replaces the synthetic layer.
Price formation — posted prices, not auctions
GPU compute is a posted-price market: providers set rate cards and adjust them by utilization-targeting (yield management), competitive benchmarking, and cost-plus floors; large deals are bilateral OTC. There is no central auction — AWS Spot itself has been a posted, administratively adjusted price since 2017. Marketplaces (RunPod-style) rank decentralized host prices cheapest-first, which approximates auction outcomes over time without running one.
Modeling consequence: prices lag utilization. Series here are modeled as posted prices adjusting toward clearing with a lag (the tightness term); excess demand manifests first as queues and allocation, then as price. This is why a demand line can sit beyond the posted stack on the merit order — and why the flexibility tranches (Definitions › Sources of flexibility) mark where latent supply surfaces as price rises.
Sources — contract-type definitions
Spot / on-demand definitions follow provider documentation: RunPod — spot vs on-demand (spot interruptible any time, ~5s SIGTERM, up to ~60% below on-demand), AWS EC2 Spot (2-minute interruption notice; posted price, no longer a true auction), and Google Cloud Spot VMs. Observed spot discounts run ~40–70%; this platform's synthetic spot multiplier (0.62× on-demand) sits at the conservative end.
Commitment products in the real market are heterogeneous — AWS Capacity Blocks (defined start + duration), GCP committed-use discounts, bespoke neocloud contracts, and bilateral multi-year offtakes. The "reserved (1yr)" and "long-term" buckets used here are modelling conventions collapsing that landscape; mapping it properly is recorded future work (see SPEC › Contract-terms map).