Index Methodology

Axon indices aggregate raw pricing observations from 20+ independent GPU cloud providers and marketplaces spanning 7 global regions.

Data is collected via authenticated REST APIs, GraphQL endpoints, and structured browser-based extraction. Each observation is timestamped at the source and stored with full provenance - source ID, raw price, raw unit, currency, GPU model, region, and availability count.

Sources are classified into three tiers by proximity to real transactions. Market-clearing prices (spot and preemptible instances) receive the highest weight, followed by marketplace bid/ask prices, with published on-demand list prices weighted minimally - reflecting that list prices are ceiling rates rarely paid at scale.

Raw observations arrive in heterogeneous formats - per-instance-hour, per-GPU-hour, monthly reserved rates, and per-GB pricing. All compute prices are normalized to a standard unit of US dollars per single GPU per hour ($/GPU-hr).

Multi-GPU instance prices are divided by GPU count. Reserved pricing is amortized to hourly equivalents. Non-USD prices are converted using daily ECB reference exchange rates.

Stale observations are filtered based on configurable freshness thresholds. Minimum availability thresholds ensure only genuinely purchasable capacity contributes to the index.

A multi-step filtering pipeline identifies and excludes anomalous observations before aggregation.

Observations beyond 3 standard deviations from the rolling 7-day median for that source-model-region combination are flagged and excluded. This guards against data entry errors, stale cache artifacts, and promotional pricing that doesn't reflect true market conditions.

An anti-manipulation guard ensures no single data source may contribute more than 25% of the weight in any published index value. A minimum number of independent sources is required for publication - indices with insufficient source diversity are marked as provisional.

Surviving observations are grouped by region and aggregated into regional reference prices using a liquidity-weighted methodology.

Each observation's weight is a function of its reported availability (scaled by square root to dampen outlier capacity claims) and its proximity to the regional median (exponential downweighting for prices far from consensus). This approach reflects actual market depth - a provider with 1,000 available GPUs at a given price contributes more to the reference price than one with 5.

Regional sub-indices are only published when a region has sufficient contributing observations from multiple independent sources.

Regional reference prices are combined into a global index value, weighted by observed market depth across regions. Regions with insufficient real transaction data receive reduced weight in the global index, ensuring that list-price-only regions do not distort the benchmark.

For model-specific indices (e.g., GCI-H100), the global value reflects the liquidity-weighted average across all qualifying regions. For the headline GCI composite, individual model indices are combined using dynamic weights that reflect each model's current market liquidity and lifecycle stage.

Memory indices (GMI family) use a simpler source-weighted median methodology, as memory markets have different liquidity characteristics than GPU compute spot markets.

Official daily fixing values are published at 16:00 ET, aligned with the US equity market close. The fixing window uses observations collected during the 24-hour period preceding the fixing time.

Each published value is accompanied by metadata: source count, observation count, methodology version, and publication status (published, provisional, or under review). Indices that experience day-over-day changes exceeding defined thresholds are flagged for manual review before publication.

Full calculation logs - including every intermediate step, parameter used, and observation included or excluded - are retained for 5 years in compliance with IOSCO Principles for Financial Benchmarks.