A market without a price — until now
Global AI inference compute is a multi-billion dollar market trading daily without a common pricing standard. Enterprises, AI application developers, and institutional allocators make capital allocation decisions across a fragmented landscape of vendors, architectures, and geographies — with no benchmark against which to measure cost, risk, or value.
The Global Compute Price Index (GCPI) is the first systematic attempt to construct a standardized, reproducible benchmark for AI inference costs. Modelled on the methodology of established commodity indices — drawing from energy price benchmarks in its treatment of supply-side shocks, and from CPI construction in its standardization approach — the GCPI provides a single reference rate for the cost of running one million tokens of AI inference across a normalized basket of providers and model tiers.
The GCPI is constructed as a weighted geometric mean of token-normalized inference prices across provider tiers. Weights reflect market share estimates derived from reported API traffic and public funding data. Prices are standardized to a common unit: USD per 1,000,000 input tokens, using the Llama 3.1 family as the reference model architecture, enabling like-for-like comparison across heterogeneous hardware.
GCPI = ∏ (P_i)^(w_i) | ∑w_i = 1 | P_i = USD / 1M tokens (normalized)
Geopolitical Risk Premium (GRP) is estimated using an event-study approach: we identify discrete regulatory and supply-chain events, measure the mean price response of affected provider cohorts over a 30-day window, and express the cumulative effect as a percentage premium over the pre-event baseline. This approach draws directly from event study methodology in empirical asset pricing.
The inference market: floor, ceiling, and everything in between
Our inaugural dataset covers 8 inference providers across Tier S (Llama 3.1 8B reference architecture). The most striking finding: a 10.0× spread between the cheapest and most expensive providers serving the same underlying model weights. This dispersion is not explained by model quality alone — it reflects hardware efficiency, go-to-market strategy, and in several cases, subsidized pricing to acquire market share.
Supply shocks are not priced — yet
The semiconductor supply chain underpinning global compute has never been more exposed to geopolitical disruption. Between June 2025 and April 2026, we identify four discrete policy events that materially affected compute supply conditions. Applying an event-study framework, we estimate a cumulative Geopolitical Risk Premium of +18% embedded in current spot pricing for GPU-backed inference relative to a counterfactual no-disruption baseline.
This premium is structurally analogous to the energy risk premium in oil markets: it compensates suppliers for the option value of supply restriction and compensates buyers for replacement cost uncertainty. As export control policy continues to evolve — the January 2026 BIS revision introduced case-by-case licensing for H200 exports to China, creating sustained uncertainty — this premium is unlikely to compress in the near term.
Three structural observations
1. The inference market is in a race to the floor, but the floor has a geopolitical ceiling. Commoditization pressure is real: DeepInfra prices Llama 8B at $0.03/M tokens, near or below cost for many providers. Yet this floor is contingent on continued access to NVIDIA silicon — a contingency that US export controls make increasingly uncertain for non-Tier 1 providers.
2. Price dispersion encodes hidden quality signals. The 10.0× spread between providers is not random. Groq's premium reflects specialized LPU silicon with deterministic latency — a genuinely different product. Fireworks AI's premium reflects custom inference kernels delivering 4× faster structured output. Dispersion is a feature of a market with real product differentiation, not a bug.
3. The market lacks forward pricing entirely. Unlike energy, agricultural, or financial commodity markets, there are no futures contracts, no forward curves, and no implied volatility surface for compute. This absence imposes real costs on both sides: suppliers cannot hedge capacity investments, buyers cannot lock in input costs. The economic argument for compute derivatives is identical to the argument that gave rise to oil futures in the 1970s.