Definitions

All capacity and demand figures on this platform are counted in thousands of H100-equivalent GPUs (k GPUs) — physical accelerators normalized to the sustained useful throughput of one NVIDIA H100 SXM.

Why — compute is not a standardized commodity: a fleet of 10,000 A100s and a fleet of 10,000 B200s differ in useful output by roughly 5×, so raw GPU counts cannot be added, compared, or priced against each other. Energy markets solve the same problem by normalizing volumes to energy content (gas is traded in MMBtu, not cubic metres). H100-equivalence is that normalization for compute: it makes supply, demand, and balances additive across heterogeneous hardware. The H100 is the reference unit because it is the most widely deployed and most benchmarked accelerator of the buildout era.

How computed — each GPU type is assigned an equivalence factor: its sustained effective throughput on a representative mixed workload, relative to the H100. The factor is not peak-spec FLOPS; it is precision-weighted delivered FLOPS × achievable utilization, degraded for memory-bandwidth and interconnect limits at cluster scale. Fleet capacity is thenH100-eq = Σ (units × factor)summed over hardware types.

Factors above are illustrative placeholders pending benchmark-derived values — see Methodology. Caveats: equivalence is workload-dependent (training vs inference weight memory and interconnect differently), precision-dependent (FP8/FP4 gains are not uniform), and cluster-dependent (network topology can dominate at scale). A single scalar is a deliberate simplification — the price of an additive market balance.

A GPU class groups accelerators of similar generation, architecture, and capability that are close market substitutes — they rent at similar prices and serve the same workloads. All class-level figures are deployment-weighted averages over member SKUs:

• A100-class — A100 40/80GB, A800
• H100-class — H100 SXM/PCIe, H800
• H200-class — H200, GH200
• B200-class — B200, GB200
• B300 & next-gen — B300, GB300, successors
• AMD MI-series — MI300X, MI325X, MI355X
• Custom silicon — TPU, Trainium, in-house ASICs (not openly rented; excluded from rental pricing)

Equivalence factors (above) are defined per specific SKU; a class's factor is the deployment-weighted average of its members.

• On-demand— uninterruptible pay-as-you-go: start and stop at will, no commitment, the provider cannot reclaim the capacity while you hold it. The most expensive way to buy compute — you pay for flexibility. The market's prompt price.
• Spot — interruptible (preemptible) capacity: the provider can reclaim it at short notice (typically seconds to minutes) when a higher-value customer appears. Sold at a discount to on-demand; suits fault-tolerant, checkpointed, or batch workloads. The discount is the price of interruption risk — the spot−reserved spread is a market-tightness signal.
• Reserved (1yr) — a one-year commitment starting the stated month: guaranteed capacity at a discounted rate in exchange for taking volume risk.
• Long-term — multi-year offtake contracts (the Oracle–OpenAI shape): deepest discount, longest commitment; economically a forward strip (see Time spreads).

Source of truth — "spot" and "on-demand" are industry-standard terms, defined here per provider documentation (interruption notice varies by provider: ~5s on RunPod, 2min on AWS). "Reserved (1yr)" and "long-term" are this platform's modelling conventions — the market offers many commitment products with no standard tenor. See Methodology › Sources.

Forecast supply is classified by how certain it is to arrive, adapted from oil & gas reserves classes (1P/2P/3P) and grid interconnection-queue stages. Balance lines are named by the cumulative supply they assume (e.g. "Balance · S2" — see the notation note below).

Notation: by abuse of notation, for brevity, Sndenotes the cumulative stack S1 + … + Sn — so S2 = online + committed ("committed supply"), S3 adds probable, S4 = max supply. Where a single tier is meant in isolation we write "tier Sn". Pipeline tiers (S2–S4) are forward-looking, measured as of today — history shows "—" because everything already built is S1. Only ~15–20% of capacity in power-grid interconnection queues historically reaches operation — the reason tier S4 is quarantined.

Demand history is measured and shared; forecasts fan out as five growth scenarios, ordered high → low: D1 Extreme high · D2 High · D3 Most likely · D4 Low · D5 Extreme low. Charts default to D3; the "all scenarios" view draws the fan with opacity falling as probability falls. Each supply scenario × demand scenario pair yields its own balance line.

The merit order's posted offers are not the whole supply curve. As price rises, additional capacity surfaces — the compute market's equivalent of peaking plant. Shown as the dashed top end of the pricing stack:

• Secondary resale — holders of reserved/long-term contracts releasing unused capacity onto secondary markets (capacity-block resale, compute marketplaces). The cheapest flexibility; kicks in first.
• Own-use release— enterprises and labs renting out first-party capacity when the market bid exceeds its internal marginal value (a research cluster's "price" is the value of the experiments it runs). Heterogeneous and high — a steep upward tail.
• Priority displacement — spot preemption: reclaiming interruptible capacity converts low-priority consumption into supply for high-priority demand. Already inside the posted stack.
• Demand response — the mirror image: deferring training, quantizing/distilling, migrating regions. Belongs on the demand side; modelling it as a sloped demand curve is recorded future work.

• Effective supply — capacity online, powered, and sellable now (not nameplate/announced), in H100-eq terms.
• Effective demand— demand that would clear at prevailing prices: realized consumption + contracted backlog + estimated queued demand. Latent "unbounded" demand is excluded.
• Utilization — realized consumption ÷ effective supply.
• Balance— effective supply − effective demand; positive = surplus, negative = deficit. The platform's primary object.
• Cluster — a group of compute servers networked together; the atomic unit of supply in the cluster registry.

Full glossary (spreads, performance normalization, pricing-stack terms) in progress — added as their pages are built.