PriceCompute.com

Research

Economics of AI Compute

AI compute is becoming a distinct economic market.

  1. Energy
  2. Chips
  3. AI Compute Infrastructure
  4. Models
  5. Applications

PriceCompute focuses on the middle of that stack: the GPUs, servers, datacenters, power, cooling and networks that produce usable computing capacity for AI.

Gartner forecasts worldwide AI spending of approximately $2.67 trillion in 2026, including roughly $1.48 trillion on AI infrastructure. These are spending forecasts, not measures of compute-market value.

Bandi and Su (2026) estimate that compute investment and the service flow produced by AI infrastructure have already reached macroeconomically significant scale.

The institutions used to measure and price this capacity are still developing. Rental-market indexation remains an open research problem; a nominal GPU-hour does not fully describe the capacity a customer receives. Our current provider listings illustrate price dispersion without treating different service bundles as equivalent.

Compute is also becoming a financial asset. CME has announced H100 and B200 compute futures scheduled for launch on 5 October 2026, subject to regulatory review.

This combination creates a research agenda at the intersection of economics, finance, information systems, machine learning and infrastructure policy.

The economic problem

Why AI compute is academically interesting

01No stable commodity unit

What should a unit of compute measure? A GPU-hour describes time on hardware, but an economically comparable unit must also account for the service delivered. Bandi and Su examine the compute rental market and its indexation.

  • GPU generation and memory
  • Interconnect and cluster size
  • Software environment and workload
  • Provider quality
  • Geography
  • Reliability and availability
  • Contract duration
What is the economically equivalent unit of AI compute, and how should that unit be benchmarked?

We study how benchmark choices affect futures settlement, asset valuation, lending and hedging. The objective is to distinguish changes in price from changes in the quality or composition of capacity.

02Rapid technological and economic obsolescence

How quickly does functional hardware lose economic value? Technological progress and economic obsolescence need not coincide. Ornn’s The Economics of Open-Weight Inference examines how older GPUs can remain competitive for suitable workloads. Its results are conditional on workload, model architecture, utilization and serving configuration.

Our research asks how workload economics and energy efficiency should enter depreciation schedules, residual values, collateral assessments, lease pricing and project finance. It also examines hardware replacement cycles and workload migration across GPU generations.

03Financialization before market maturity

AI infrastructure financing spans hyperscalers, third-party developers, private credit and structured finance. These financial exposures coexist with unresolved questions about rental-market indexation and futures pricing, while regulated compute futures are scheduled to launch.

What happens when financing and financial exposure develop faster than the market institutions used to measure and hedge the underlying asset?

We examine whether fragmented spot markets, different benchmark methodologies and incomplete standardization complicate the valuation and hedging of equity, credit, leases, datacenter financing, purchase commitments and long-term compute offtake.

Research program / 01

Price Discovery and Benchmark Formation

How does a market price for AI compute emerge from fragmented and heterogeneous provider markets, and what benchmark best captures its underlying economic value?

We investigate common latent price factors across GPU generations and providers, and whether benchmarks represent comparable economic services.

  • Benchmark fragmentation
  • Transaction prices versus posted prices
  • Spot, term and forward relationships
  • Basis risk
  • Cross-provider price dispersion
  • Derivatives price discovery
  • Hedging effectiveness
  • Predictive validation of benchmark quality

Machine-learning forecasting is a tool, not the objective. The broader question is whether a benchmark captures the economic value of compute capacity.

Current study

The first H100 derivatives markets

H100 GPU-hour derivatives began trading on crypto-native venues before the scheduled launch of regulated exchange-traded compute futures.

PriceCompute has assembled historical and live data covering Lighter, Gate and Bitget, alongside Ornn physical/reference pricing and Silicon Data reference pricing. Historical decentralized compute markets are included where verified. CME data will be added once regulated futures begin trading.

  1. 01 →Physical benchmarks
  2. 02 →Crypto-native derivatives
  3. 03 →Competing venues
  4. 04Regulated futuresScheduled: 5 October 2026
  • Which venue incorporates information first?
  • How does liquidity migrate across venues?
  • Do prices converge?
  • Does benchmark choice fragment price discovery?
  • Do crypto-native derivatives contain information about later regulated futures prices?
  • What changes after regulated futures begin trading?
  • Do 24/7 crypto markets lead regulated markets during nights and weekends?

Research program / 02

Where Should AI Infrastructure Be Built?

AI datacenters combine specialized chips with power infrastructure and cooling systems. Our location research asks how electricity, scarce grid capacity, land and networks shape investment decisions.

Private location choices

Electricity prices, grid availability, fiber connectivity, land, cooling, tax incentives, latency and regulation.

Public allocation questions

The opportunity cost of grid capacity, electricity-price effects, grid reliability, employment, local productivity, tax revenues, environmental effects, strategic value and alternative industrial uses of power.

How do AI infrastructure developers choose locations?
How should governments allocate scarce grid capacity and incentives?
Where and why do private and social values diverge?

One objective is to estimate the economic and social value per MW of scarce infrastructure capacity allocated to AI, relative to alternative uses. Spatial econometrics, geospatial analysis, optimization and machine learning can help connect location choices to local outcomes. Portugal and Sines are possible empirical settings for this broader question.

Research program / 03

Financialization Before Market Maturity

Van Nieuwerburgh’s Financing the AI Buildout examines the financing structures behind AI infrastructure and their implications for leverage, risk allocation and financial stability.

Our research considers corporate equity, project finance, private credit, leases, GPU financing, purchase commitments, long-term compute offtake and derivatives.

How does financialization before mature price discovery and hedging affect valuation, capital allocation and the transmission of risk in AI compute markets?

Possible transmission channels include datacenter investment, financing costs, collateral values, credit spreads, GPU residual values, provider cash flows and wider infrastructure investment. These are hypotheses to test, rather than established causal effects.

As regulated futures enter, we ask what changes in price discovery, basis, hedging, financial-market integration and risk transmission.

Research resources

Data

PriceCompute is building a research dataset spanning the AI compute market. The categories below describe its scope; availability and historical coverage vary by source.

Compute prices

  • GPU rental-price histories
  • Provider-level listings
  • Benchmark and index histories
  • Forward curves
  • Multiple GPU generations

Derivatives

  • H100 perpetual prices
  • Individual trades where available
  • Funding rates, volume and open interest
  • Mark and index prices
  • Live order books
  • Regulated futures data after launch

AI demand

  • Model and token-market activity
  • OpenRouter model usage
  • Model and provider characteristics

Infrastructure

  • Electricity prices and power-market data
  • Datacenter locations
  • Grid and regional variables
  • Hardware economics

Finance

  • Infrastructure commitments
  • Financing structures
  • Leases and offtake agreements
  • Corporate disclosures
  • Asset and credit-market data

Where possible, raw observations are preserved with retrieval metadata, source provenance, hashes and reproducible transformation pipelines.

Research approach

Methods

The research combines economics and finance with modern data science.

  • Panel econometrics
  • Time-series analysis
  • Asset pricing
  • Market microstructure
  • Price-discovery models
  • Index construction
  • Machine learning
  • Forecasting
  • Conformal prediction
  • Natural language processing (NLP)
  • Spatial econometrics
  • Geospatial analysis
  • Optimization
The method follows the economic question, rather than the other way around.

Foundations

Academic context

The research sits within the broader Economics of Artificial Intelligence, including the questions collected in The Economics of Artificial Intelligence: An Agenda, published by the University of Chicago Press for the NBER.

For compute specifically, Bandi and Su’s (Early) AI Compute Asset Pricing provides an early academic framework for compute as a financial asset, with open questions around rental-market indexation, futures pricing and risk premia.

Market formation

Why now?

Fragmentation

Provider-level GPU listings show differing prices for nominal GPU capacity. Contracts, hardware configurations and service bundles must be considered when comparing them.

This creates an unusual opportunity to study the formation of a new commodity and financial market while it is happening.

PriceCompute

About the research

PriceCompute.com is an independent research project focused on the economics and financialization of AI compute.

The research combines experience in institutional finance, trading and risk with academic work in economics, machine learning and information management.

The objective is to produce rigorous, reproducible research on how AI compute should be priced, benchmarked, financed, allocated and hedged.

Explore the market data →