SUBSCRIBE: Compute Heat Rate on Substack → | Foundation Research V6 | CHR Index
Working Paper & Methodology Reference

The Compute Heat Rate (CHR)

An analytical framework for measuring the electricity price tolerance of AI compute workloads, and its implications for wholesale power market price formation.
Hans Royal
Originator, Compute Heat Rate (CHR) Framework
Subscribe: Substack  |  Contact: LinkedIn
First published: February 27, 2026
Last updated: June 4, 2026
Document version: 1.3

This page summarizes the methodology and key findings of the CHR framework. The full working paper, including the complete empirical specification, identification strategy, literature review, and detailed implications, is available at: Royal, H. (2026), SSRN Abstract ID 6322318 (revised June 4, 2026). Quarterly reference values are published at the CHR Index.

On This Page Abstract Motivation Definition & Formula Methodology Results CONE-to-CHR Spectrum Implications Publications & Citations Formal Citation

Abstract

Artificial intelligence data centers represent a fundamentally new class of electricity demand with price tolerance characteristics that have no historical precedent in wholesale power markets. This paper introduces the Compute Heat Rate (CHR), a metric that quantifies the maximum electricity price at which AI computation remains economically viable, measured in dollars per megawatt-hour and expressible as a demand-side price tolerance multiple relative to traditional industrial curtailment thresholds. Drawing on publicly available data on GPU economics, cloud computing pricing, and facility cost structures, the CHR is calculated across six workload categories. Results indicate that AI workloads can tolerate electricity prices ranging from approximately $500/MWh for frontier model training to over $53,000/MWh for frontier inference services, with a blended average of approximately $6,350/MWh, representing a 127-fold multiple over the conventional gas heat rate benchmark of approximately $50/MWh.

The contributions are threefold. First, the paper formalizes the CHR as a quantitative metric derived from first principles. Second, it introduces the CONE-to-CHR pricing spectrum as a theoretical model for the range within which wholesale prices will settle in supply-constrained markets with significant AI demand. Third, it presents preliminary empirical evidence that a distinct AI load price component is emerging in wholesale markets at high data center concentration zones. The potential implications for industrial electricity consumers, forward curve construction, procurement strategy, and financial risk management are discussed.

1. Motivation

The wholesale electricity market in the United States is experiencing a structural demand shock without historical parallel. The U.S. Department of Energy estimates that data center electricity consumption could reach 12% of total U.S. electricity consumption by 2028. The financial magnitude of this demand class is evident in the equity markets: NVIDIA Corporation reported fiscal year 2026 revenue of $215.9 billion, with data center segment revenue of $193.7 billion, reflecting demand for the computational infrastructure that converts electricity into AI output.

The disconnect between equity market pricing and electricity market pricing is the central observation motivating this framework. Equity markets price massive value in AI computation. Electricity markets have not yet developed the analytical tools to understand what this value implies for demand-side price formation. The gas heat rate has governed the supply side of wholesale price formation for three decades: it converts fuel cost into generation cost. No equivalent metric exists for the demand side. The CHR fills this gap.

It is important to clarify what this framework is not. It is not a forecast of where wholesale electricity prices will settle. It is a measurement tool: the demand-side analogue to the gas heat rate, quantifying the price tolerance of a new and structurally different class of electricity demand.

2. Definition

The Compute Heat Rate is defined as the maximum electricity price at which a given AI compute workload remains profitable for the operator, after accounting for all non-electricity costs and a required return margin.

CHRw = (Rw − Cnon-elec) / (1 + m)
Equation 1. The Compute Heat Rate for workload type w.
CHRw
Compute Heat Rate for workload type w, expressed in $/MWh. Represents the maximum electricity price at which the workload remains economic.
Rw
Gross revenue per MWh of electricity consumed by workload type w. Derived from API pricing, cloud compute rates, or enterprise contract values, converted to revenue per MWh using GPU power consumption and throughput data.
Cnon-elec
Non-electricity operating costs per MWh, including GPU amortization, facility overhead, cooling, networking, and maintenance. Estimated from published infrastructure cost data and industry benchmarks.
m
Required return margin (expressed as a decimal). Represents the minimum profit margin the operator requires. Baseline assumption: 0.30 (30%).

The CHR is conceptually analogous to the gas heat rate but operates on the demand side of the market rather than the supply side. Where the gas heat rate converts a fuel input price into a generation cost (setting a floor for wholesale prices), the CHR converts a compute revenue output into a maximum tolerable electricity input cost (setting a ceiling on the price at which AI demand curtails).

3. Input Data Requirements

The CHR is calculated from publicly available data sources. No proprietary or confidential data is required.

InputSourceRole in Calculation
GPU power consumptionNVIDIA published specifications; MLPerf benchmarksConverts compute throughput to electricity consumption per unit of output
Power Usage Effectiveness (PUE)Industry benchmarks (Uptime Institute, DOE)Scales IT load to total facility electricity consumption
API / inference pricingOpenAI, Anthropic, Google published pricingEstablishes revenue per token/request; converted to Rw per MWh
Cloud compute rental ratesAWS, GCP, Azure published pricingAlternative revenue basis for non-API workloads
GPU throughputNVIDIA documentation; independent benchmarksLinks power consumption to revenue-generating output
Infrastructure costsHyperscaler SEC filings; industry reports (JLL, CBRE)Basis for Cnon-elec estimation
Workload mix estimatesIndustry analyst reports; hyperscaler earnings guidanceWeighting for blended CHR calculation

Table 1. Input data sources for CHR calculation. All sources are publicly available.

R(w) Derivation: The Four-Link Conversion Chain

The revenue per MWh of electricity consumed, R(w), is derived through a four-link conversion chain that transforms observable market prices into the electricity-denominated revenue metric required by the CHR formula:

Link 1 (Price): Published API or contract pricing establishes revenue per token ($/million tokens).
Link 2 (Throughput): GPU throughput benchmarks establish tokens produced per second per GPU (tokens/sec).
Link 3 (Power): GPU power specifications convert throughput into tokens per unit of electricity consumed (tokens/kWh).
Link 4 (R(w)): The product of tokens per MWh and revenue per token yields R(w) in $/MWh.

The reference hardware configuration for all workload tiers is the NVIDIA DGX H100 (8x H100 SXM5, 700W TDP per GPU, 5,600W server GPU power). Applying a Power Usage Effectiveness (PUE) factor of 1.3 (Uptime Institute 2025 Global Data Center Survey) yields total facility power of approximately 7,280W (0.00728 MWh/hr) per server.

Worked Example: Frontier Inference (Opus/GPT-5 class, Q1 2026)

Anthropic's published API pricing for Claude Opus 4.5 is $5/$25 per million tokens (input/output). At a standardized production throughput of approximately 10,000 tokens/sec for the 8xH100 server and a blended price of approximately $15/M tokens, revenue is approximately $540/hr. Dividing by facility power consumption of 0.00728 MWh/hr yields R(w) of approximately $74,000/MWh.

Worked Example: Mid-Tier Inference (Sonnet/GPT-4.1 class, Q1 2026)

At Sonnet/GPT-4.1 class pricing (~$3/M output tokens blended), the same 10,000 tokens/sec server throughput yields approximately $108/hr in revenue: R(w) = $108 / 0.00728 = approximately $14,800/MWh. This is independently corroborated by published cloud GPU rental rates: at approximately $13.50 per GPU-hour, an 8-GPU server generates $108/hr in rental revenue, reproducing the token-derived R(w) from an independent data source.

Non-Electricity Cost Decomposition

C(non-elec) captures all facility-level operating costs excluding electricity, scoped to the data center site level. Corporate-level general and administrative expenses are excluded as they are not attributable to the marginal cost of operating a given facility. Principal components, estimated from hyperscaler SEC filings and industry reports (Cushman & Wakefield; JLL): GPU amortization over a 3-year lifecycle (~55%); facility construction amortization (~9%); networking and storage infrastructure (~11%); operations and labor (~7%); cooling and power distribution (~4%); other (~14%).

Full derivation methodology for all six workload tiers, including throughput anchor sources (MLPerf verified results), utilization assumptions, and sensitivity analysis, is documented in Royal (2026), SSRN Abstract ID 6322318.

4. Initial Results (2026)

The following table presents the baseline CHR calculations across major AI workload categories using Q1 2026 pricing and infrastructure cost data, as published in Royal (2026). The gas heat rate benchmark of approximately $50/MWh is provided for comparison. For the latest quarterly reference values, see the CHR Index.

Workload TypeRw ($/MWh)Cnon-elecCHR ($/MWh)Multiple vs. Gas HR
Frontier Inference (Opus/GPT-5)$74,000$4,250$53,650~1,070x
Mid-Tier Inference (Sonnet/GPT-4.1)$14,800$4,250$8,120~162x
Enterprise Agentic AI$15,000$4,500$8,080~162x
Enterprise Contracted$5,900$4,250$1,270~25x
Commodity Inference (mini)$1,850$3,800*~$800*~16x
Frontier Model Training$2,000†$5,200*~$500†~10x
Blended Average (2026)$12,500$4,350$6,350~127x

Table 2. Computed CHR by AI workload type, Q1 2026. Assumes 30% required margin. *Non-electricity costs differentiated by infrastructure tier: commodity inference uses lower-cost air-cooled facilities ($3,800/MWh); frontier training requires liquid cooling, high-bandwidth networking, and denser power delivery ($5,200/MWh). †Standalone CHR is negative; effective CHR reflects portfolio-level cross-subsidization (see Royal, 2026). Gas heat rate benchmark: ~$50/MWh.

The blended CHR of approximately $6,350/MWh implies that AI data center demand will not curtail at electricity prices below roughly 127 times the current wholesale average. Even the lowest positive-margin workloads exhibit tolerance ceilings 25 times above the gas heat rate. This asymmetry is the central finding: the demand-side brake that has historically governed wholesale price formation does not apply to AI compute workloads at any price level relevant to current market dynamics.

5. The CONE-to-CHR Pricing Spectrum

The long-run equilibrium wholesale price is determined by the Cost of New Entry (CONE), typically $80 to $130/MWh in current U.S. markets. The CONE-to-CHR spectrum defines three pricing regimes:

Regime 1 (Equilibrium): Generation capacity is sufficient to meet all demand including AI load. Wholesale prices settle near CONE. The CHR is irrelevant because supply is not constrained.

Regime 2 (Persistent Disequilibrium): Demand growth outpaces supply response. Prices rise above CONE toward the CHR ceiling. Traditional industrial load curtails; AI load does not. Duration depends on the speed of new generation construction.

Regime 3 (Structural Repricing): AI load becomes a sufficiently large share of total demand that its price tolerance sets the marginal clearing price in a significant number of hours. Wholesale prices settle at levels between CONE and the blended CHR.

6. Implications

For Industrial Electricity Consumers

The most consequential implication is for traditional industrial electricity consumers: the manufacturers, processors, and producers for whom electricity is a significant input cost. The CHR creates a two-tier electricity economy. In one tier, industries with high revenue per MWh (AI computation, cloud services) can absorb price increases that would be ruinous for traditional industry. In the other, industries with low revenue per MWh (aluminum at ~$60 to $80, steel at ~$80 to $120, chemicals at ~$100 to $150) face a structural disadvantage compounded by geographic immobility.

For Forward Curve Construction

Current consensus forward curves at DC-concentrated settlement points do not incorporate a demand-class-specific price tolerance variable. The CHR framework suggests these curves are systematically mispriced: they underestimate the demand-side price pressure that emerges as data center concentration crosses penetration thresholds.

For Financial Instruments

The CHR identifies a hedgeable risk that has no existing market. The analogy to the gas heat rate derivative market (heat rate options, spark spread futures) suggests a natural product family. The identification of natural counterparties (generators in DC-heavy regions as sellers, industrial consumers as buyers) provides the basis for instrument design.

For Policymakers and Market Designers

The CHR provides a quantitative basis for evaluating cost allocation questions: should data centers bear a proportional share of grid upgrade costs? Should market designs be modified to account for extreme demand-side price inelasticity? PJM's May 2026 white paper explicitly used the CHR workload decomposition to frame proposed ORDC reform.

7. Analogues in Energy Economics

MetricFunctionRelationship to CHR
Gas Heat RateConverts fuel price to generation cost; sets marginal wholesale priceSupply-side counterpart. CHR is the demand-side analogue.
CONE (Cost of New Entry)Long-run equilibrium price for new generation capacitySets the price floor in supply-constrained markets. CHR sets the ceiling.
LCOELevelized cost of electricity from a generation sourceSupply-side metric. CHR measures demand-side willingness-to-pay.
VIXImplied volatility derived from options pricingDerived metric (not directly observed) that became an industry standard and tradable index.

Table 3. Positioning of CHR among established energy and financial metrics.

8. Research Agenda

The CHR framework introduced here represents an initial formulation. Several extensions are under active development:

Hub-Level CHR Calculation

Computing location-specific CHR values at individual settlement points by incorporating regional workload mix data, local infrastructure cost differentials, and geographic demand concentration patterns.

CHR Time Series & Index Construction

Developing a methodology for periodic (quarterly or monthly) CHR publication to enable tracking of the metric over time. This includes defining index weighting methodology, data update procedures, and publication standards suitable for use as a market reference.

CHR Spread Analysis

Examining the difference between the CHR tolerance ceiling and actual wholesale prices at specific settlement points as a measure of latent repricing potential, with applications to forward curve analysis and risk assessment.

International Application

Extending the CHR framework to European and Asian electricity markets, accounting for different market designs, regulatory structures, and data center demand patterns.

Publications & Citations

Academic & Institutional

Third-Party Citations & Coverage

Compute Heat Rate Research (Substack)

LinkedIn

Subscribe for ongoing CHR research and market analysis: computeheatrate.substack.com

Formal Citation

Royal, Hans, The Compute Heat Rate: Quantifying AI-Driven Electricity Price Tolerance and Its Implications for Wholesale Market Repricing (February 28, 2026; revised June 4, 2026). Available at SSRN: http://dx.doi.org/10.2139/ssrn.6322318

Data Sources

All calculations in this framework are derived from publicly available data sources, including: NVIDIA published GPU specifications and documentation; OpenAI, Anthropic, and Google published API and model pricing; AWS, GCP, and Azure published cloud compute pricing; hyperscaler capital expenditure guidance from SEC filings; U.S. Energy Information Administration (EIA) electricity market data; PJM Interconnection and ERCOT published market data; Lazard Levelized Cost of Energy reports; LevelTen Energy PPA Price Index; CBRE, JLL, and Cushman & Wakefield data center market reports; and Uptime Institute and DOE data center efficiency benchmarks.