The maximum electricity price AI workloads can economically sustain, presented as long-run CHR and dispatch-focused CHR-D reference values.
Even after the Q3 decline, AI workload electricity tolerance remains orders of magnitude above conventional thermal-generation economics.
Q3 CHR is roughly 30% below the published Q2 reference value. The quarter incorporates a corrected full-system power boundary alongside updated serving assumptions, workload taxonomy and weights, and refreshed market inputs. Because Q1 and Q2 remain preserved as originally published rather than restated on the Q3 basis, the quarter-over-quarter change should not be interpreted as a pure market-price movement.
| Metric | Public Headline | Canonical v3.2 Output | Primary Application |
|---|---|---|---|
| Blended CHR | ~$5,630/MWh | $5,631.40/MWh | Long-run build and siting economics |
| Blended Dispatch CHR (CHR-D) | ~$9,070/MWh | $9,070.82/MWh | Dispatch and curtailment decisions |
Primary headlines are rounded for readability. Exact values are the canonical outputs of CHR_Master_Model_Q3 v3.2 under the Q3 reference case.
Compute Heat Rate (CHR) is the long-run, full-cost electricity price threshold associated with workload economics, including non-electricity costs and a required return treatment.
Dispatch Compute Heat Rate (CHR-D) is the dispatch-focused electricity price threshold at which an already-deployed workload becomes economically indifferent between continuing to operate and curtailing. CHR-D excludes sunk capital costs and required return and deducts only costs avoidable upon curtailment.
| Q3 Workload Type | Weight | Long-Run CHR | CHR-D |
|---|---|---|---|
| Frontier inference | 11% | $12,790.58/MWh | $18,377.75/MWh |
| General-purpose inference | 36% | $7,602.97/MWh | $11,633.86/MWh |
| Internal / self-hosted inference | 10% | −$402.60/MWh | $1,226.62/MWh |
| Enterprise contracted | 29% | $4,888.04/MWh | $8,104.45/MWh |
| Sovereign Open Source | 6% | $886.65/MWh | $2,902.64/MWh |
| Commodity / mini models | 8% | $711.22/MWh | $2,674.59/MWh |
| Blended Q3 reference | 100% | $5,631.40/MWh | $9,070.82/MWh |
Training is excluded from the active deployed-commercial-inference blend.
Training measurement boundary: Training is excluded from the deployed commercial inference blend because CHR requires an observable or defensibly imputable current economic value of compute. Commercial inference can generally be tied to current monetization observables such as API pricing, contracted capacity, or market-based compute pricing. Model providers plainly have an internal willingness to pay for training, but a training run produces a model whose value is uncertain, realized later, and dependent on commercialization. A public $/MWh training value would therefore require speculative assumptions about model success, utilization, pricing, and lifetime token volume. Training may be analyzed separately for context, but CHR cannot rigorously measure that willingness to pay from current public market observables. Prior treatment: Q1 and Q2 included training in the broader workload taxonomy; Q3 excludes it from the deployed commercial inference blend and treats it separately.
Why is self-hosted long-run CHR negative? Internal workloads have no observable token selling price, so Q3 uses prevailing H100 rental parity of approximately $3 per GPU-hour as a transparent opportunity-cost proxy, not a token price. The proxy is conservative and may understate the full internal value organizations derive from self-hosted GPUs. Actual self-hosted CHR may therefore be higher, and refining the proxy remains on the CHR research agenda. A negative long-run CHR means rental-parity value does not fully clear new-build costs and required return; it does not mean the GPUs create no value or should be shut down. Positive CHR-D indicates that continued operation remains economical once sunk capital costs are excluded. Prior treatment: Q1 and Q2 grouped this activity within Enterprise / internal workloads using a direct value proxy. Q3 identifies Internal / self-hosted inference separately and values it using the H100 rental opportunity-cost proxy.
The Q3 blended values use an energy/consumption-weighted basis. Each weight represents the workload tier’s estimated share of electricity consumption within the defined Index scope.
| Q3 Workload Type | Energy Weight | Basis |
|---|---|---|
| Frontier inference | 11% | Reasoning and agentic activity increased; lower serving throughput makes the tier more electricity-intensive than its token share alone suggests. |
| General-purpose inference | 36% | Core metered inference activity, separated from contracted enterprise routing. |
| Internal / self-hosted inference | 10% | Inference on enterprise-controlled capacity without an observable external inference price. |
| Enterprise contracted | 29% | Provisioned, committed, reserved, and negotiated third-party enterprise inference. |
| Sovereign Open Source | 6% | Paid third-party hosted open-weight inference, kept mutually exclusive from self-hosted and contracted deployments. |
| Commodity / mini models | 8% | Genuine small, high-throughput model activity; revised serving physics lowers estimated MWh share even as token activity grows. |
| Active blend | 100% | U.S.-relevant deployed commercial AI inference; training excluded. |
Token share is not electricity share. Q3 composition moved toward Frontier and Enterprise contracted inference, while revised serving physics materially reduced the estimated MWh intensity of genuine Commodity / mini activity. Frontier can therefore represent a relatively small share of tokens but a materially larger share of electricity consumption.
Sovereign Open Source is limited to paid third-party hosted open-weight inference. Self-hosted open-weight activity is assigned to Internal / self-hosted inference, while provisioned or contracted open-weight activity is assigned to Enterprise contracted. This routing keeps the categories mutually exclusive.
Enterprise Agentic AI is no longer carried as a standalone Q3 tier. Agentic activity is assigned across Frontier inference, General-purpose inference and Enterprise contracted according to the underlying serving and commercial relationship, preserving mutual exclusivity across the Q3 workload taxonomy.
Q1 and Q2 remain available exactly as published under their original methodology and are not silently restated. Q3 adopts the corrected full-system power boundary, updated serving assumptions, revised taxonomy and weights, refreshed pricing inputs, and the CHR-D cost boundary. Quarter-over-quarter comparisons therefore reflect both changes in underlying compute economics and improvements to the Index methodology. The historical tabs preserve the publication record rather than presenting a backcast on the Q3 basis.
The weight review originally scheduled for Q4 was accelerated into Q3 as part of the full-system power-boundary correction and workload-taxonomy rebuild. Q1 and Q2 remain preserved as published; the revised weights apply prospectively beginning with the Q3 vintage.
Master Model: CHR_Master_Model_Q3 v3.2, Q3 reference case, August 23, 2026.
Reference scenario: H100-CORR; NVIDIA DGX H100; 10.2 kW IT power; PUE 1.3; 13.26 kW facility power; 30% required return treatment.
Throughput treatment: workload-class operating points from the Q3 model source register, including MLPerf references and disclosed modeling assumptions where comparable production benchmarks are unavailable.
Pricing vintage: provider records and ratified market judgments available through August 19, 2026, as recorded in the model source register.
Cost treatment: H100-calibrated non-electric cost build; CHR-D deducts only costs classified as avoidable upon curtailment.
Workload weights: ratified August 20, 2026 Q3 MWh-consumption vintage; 11% / 36% / 10% / 29% / 6% / 8%; training excluded.
Model checks: all hard calculation and reconciliation checks pass. The source register retains nine non-blocking provenance items for future completion under the v3.2.1 publication-readiness ruling.
The Q3 Index combines sourced inputs with disclosed assumptions and adjudicated methodology judgments. It does not claim that every constituent has a current primary-source observation. Open provenance work is retained in the model source register and addressed through future vintages rather than silently changing this ratified quarterly release.
The public Index provides the core methodology and reference results. Readers seeking additional methodology detail, assumptions, applications, or supporting information are encouraged to contact us.
When referencing CHR values from this publication:
Royal, H. (2026). "CHR Reference Values, Q3 2026." Compute Heat Rate Index, September 1, 2026. https://computeheatrate.com/chr-index
Underlying methodology:
Royal, Hans, The Compute Heat Rate: Quantifying AI-Driven Electricity Price Tolerance and Its Implications for Wholesale Market Repricing (February 28, 2026). Available at SSRN: http://dx.doi.org/10.2139/ssrn.6322318
CHR measures long-run build and siting economics. CHR-D is the dispatch-focused measure used for curtailment and operating decisions.
The CHR does not constitute investment advice, a price forecast, or a recommendation. It is an analytical framework measuring demand-side electricity price tolerance.
The maximum electricity price each AI workload tier can profitably sustain, derived from GPU-level economics using exclusively public data.
The Q2 blended CHR increased approximately 27% quarter-over-quarter, driven by five developments: (1) significant upward movement in frontier API pricing, particularly OpenAI's launch of GPT-5.5 at $5/$30 per million tokens (output pricing tripled versus GPT-5); (2) mid-tier API pricing firming as GPT-5.4 replaced GPT-5 as the production workhorse at $2.50/$15, a +32% increase in blended token pricing; (3) the agentic AI tier solidifying with Anthropic's launch of Managed Agents and Claude Code reaching $2.5B in annualized revenue; (4) the commodity inference tier achieving positive standalone CHR for the first time, as GPT-5.4 Nano displaced legacy mini models at higher price points; and (5) enterprise contracted catching up to mid-tier repricing with a +36% CHR increase as Q1 contract renewals incorporated higher base rates. The training tier was unchanged.
The direction is unambiguous: the compute market is moving toward higher-value, higher-price-tolerance workloads faster than the supply side can respond. The frontier ceiling is rising, the mid-tier is thickening, and the agentic tier is emerging as the largest single category of deployed commercial inference.
| Workload Tier | R(w) $/MWh | Cne $/MWh | CHR $/MWh | vs. Q1 |
|---|---|---|---|---|
| Frontier Inference (Opus 4.7/GPT-5.5) | $83,250 | $4,250 | $60,770 | ↑ 13.3% |
| Mid-Tier Inference (Sonnet 4.6/GPT-5.4) | $19,500 | $4,250 | $11,730 | ↑ 44.5% |
| Enterprise Agentic AI | $16,500 | $4,500 | $9,230 | ↑ 14.2% |
| Enterprise Contracted | $6,500 | $4,250 | $1,730 | ↑ 36.3% |
| Commodity Inference (mini/nano)* | $5,700 | $3,800 | $1,465 | ↑ 83% |
| Frontier Model Training* | $2,000 | $5,200 | ~$500† | Flat |
| Blended Average (Q2 2026) | ~$14,800 | $4,350 | ~$8,000 | ↑ 27% |
Gas heat rate benchmark: ~$50/MWh. *Non-electricity costs differentiated by infrastructure tier. †Standalone CHR is negative; effective CHR reflects portfolio-level cross-subsidization (see Royal, 2026). Commodity inference achieves positive standalone CHR for the first time in Q2, reflecting the upward shift in budget model pricing. 30% required margin applied. Q2 R(w) values derived from first-principles re-calculation using May 2026 API pricing. Model references updated to current flagship tiers (Opus 4.7, GPT-5.5, Gemini 3.1 Pro).
The blended CHR is a consumption-weighted average across six workload tiers, where weights reflect each tier's estimated share of deployed commercial inference electricity consumption. The scope is "deployed commercial inference environment": facilities actively serving customer workloads. Training clusters, internal research, and pre-deployment testing are excluded from the primary weighting but represented in the training tier at a small allocation.
| Workload Tier | Weight | Basis |
|---|---|---|
| Frontier Inference | 1% | Small share of total deployed MWh; highest-capability models used selectively |
| Mid-Tier Inference | 30% | Workhorse production models (Sonnet, GPT-5.4, Gemini Pro) serving high-volume API traffic |
| Enterprise Agentic AI | 37% | Multi-step agentic workflows, coding agents, orchestrated tool-use sessions |
| Enterprise Contracted | 29% | Fixed-price enterprise deployments, private model hosting, dedicated inference |
| Commodity Inference | 1.5% | Budget models (Haiku, Nano, Flash-Lite); high query count, minimal per-query MWh |
| Frontier Training | 1.5% | Active training runs as share of total deployed compute at any given time |
Weights derived from hyperscaler earnings disclosures, published API traffic estimates, and industry surveys (McKinsey, Goldman Sachs, IEA). Q2 2026 uses Q1 weights for time-series comparability. Formal weight review scheduled for Q4 2026; any revision will be disclosed one quarter in advance. Sensitivity: shifting 3pp from Enterprise Contracted to Enterprise Agentic AI moves the blended CHR by approximately +$250/MWh.
The robustness of the blend is notable: computing the blended CHR using only the four tiers with positive standalone values yields a result within 2% of the full six-tier blend, because the high-revenue workloads dominate the revenue-weighted average regardless of how the low-revenue tiers are weighted.
The core regression in Royal (2026) decomposes LMP at each settlement point into component drivers (natural gas prices, renewables, weather, and a data center load proxy), estimated using OLS with heteroskedasticity-robust standard errors clustered at the settlement point level. The coefficient on the data center load proxy represents the marginal price impact ($/MWh per GW of weighted data center capacity) attributable to data center demand concentration.
PJM’s Base Residual Auction results for three consecutive delivery years provide independent corroboration: the 2025/2026 auction cleared at $269.92/MW-day (roughly 10x the prior year), the 2026/2027 auction at $329.17/MW-day (hitting the FERC-approved cap), and the 2027/2028 auction again at the cap of $333.44/MW-day. Three consecutive record-high capacity auctions, driven by a single demand class, represent a market signal consistent with the structural repricing thesis.
Full specification and results are in the SSRN working paper (Royal, 2026; revised June 4, 2026).
The CHR framework has been cited by PJM Interconnection (May 2026). Full citation record at Publications & Citations.
Significant market developments since Q1 2026, analyzed through the CHR framework:
Where Rw is the revenue per MWh of compute for workload type w, Cnon-elec is the non-electricity operating cost (cooling, networking, real estate, labor, amortized hardware), and m is the required return margin (30%), applied as a markup on electricity cost. All inputs derived from public data. Full derivation including the GPU throughput-to-MWh conversion chain is documented in Royal (2026), SSRN Abstract ID 6322318.
GPU specifications and benchmarks: NVIDIA documentation (H100, B200, GB200); MLPerf inference benchmarks.
API pricing (Q2 2026): Anthropic (Opus 4.7: $5/$25; Sonnet 4.6: $3/$15 per million tokens), OpenAI (GPT-5.5: $5/$30; GPT-5.4: $2.50/$15), Google (Gemini 3.1 Pro: $2/$12; Gemini 2.5 Pro: $1.25/$10). Verified May 2026.
Cloud compute rates: AWS, Google Cloud, Azure published on-demand GPU instance pricing.
Infrastructure costs: Cushman & Wakefield, JLL data center market reports; hyperscaler SEC filings.
Energy data: EIA (Henry Hub natural gas spot prices, wholesale electricity data).
Regression data: PJM LMP data (public), ERCOT settlement point prices (public), EIA-930, interconnection queue filings.
All inputs are exclusively public data. No proprietary or confidential data sources are used in the CHR calculation.
When referencing CHR values from this publication:
Royal, H. (2026). "CHR Reference Values, Q2 2026." Published at computeheatrate.com, June 2026.
Underlying methodology:
Royal, Hans, The Compute Heat Rate: Quantifying AI-Driven Electricity Price Tolerance and Its Implications for Wholesale Market Repricing (February 28, 2026). Available at SSRN: http://dx.doi.org/10.2139/ssrn.6322318
The Compute Heat Rate (CHR) is independent research by Hans Royal. The full methodology, including the four-link R(w) derivation chain, non-electricity cost decomposition, worked examples, and sensitivity analysis, is available at computeheatrate.com and in Royal (2026), SSRN (Abstract ID 6322318, revised June 4, 2026).
CHR values are updated quarterly as underlying data changes (GPU specifications, API pricing, cloud compute rates, facility cost benchmarks). Workload weights are held constant for time-series comparability; formal weight review is scheduled semi-annually, with any revision disclosed one quarter in advance. Material intra-quarter changes may trigger interim updates.
The CHR does not constitute investment advice, a price forecast, or a recommendation. It is an analytical framework measuring demand-side electricity price tolerance.
The maximum electricity price each AI workload tier can profitably sustain, derived from GPU-level economics using exclusively public data.
Where Rw is the revenue per MWh of compute for workload type w, Cnon-elec is the non-electricity operating cost (cooling, networking, real estate, labor, amortized hardware), and m is the required return margin (30%), applied as a markup on electricity cost.
| Workload Tier | R(w) $/MWh | Cnon-elec $/MWh | CHR $/MWh |
|---|---|---|---|
| Frontier Inference (Opus/GPT-5 class) | $74,000 | $4,250 | $53,650 |
| Mid-Tier Inference (Sonnet/GPT-4.1 class) | $14,800 | $4,250 | $8,120 |
| Enterprise Agentic AI | $15,000 | $4,500 | $8,080 |
| Enterprise Contracted | $5,900 | $4,250 | $1,270 |
| Commodity Inference (mini models)* | $1,850 | $3,800 | ~$800† |
| Frontier Model Training* | $2,000 | $5,200 | ~$500† |
| Blended Average (Q1 2026) | $12,500 | $4,350 | $6,350 |
Gas heat rate benchmark: ~$50/MWh. *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 by hyperscale operators (see Royal, 2026 for discussion). 30% required margin applied to all tiers.
The gas heat rate has governed wholesale electricity price formation for decades: fuel cost divided by thermal efficiency yields the marginal generation cost. The Compute Heat Rate introduces the demand-side analogue: economic value of compute divided by electricity consumption yields the maximum price the operator can rationally pay.
Even the lowest positive-margin workload tier (enterprise contracted, CHR ~$1,270/MWh) implies a price tolerance approximately 25 times the gas heat rate. For the four workload tiers with positive standalone CHR values, AI demand will not curtail at any price level currently observed in U.S. wholesale markets, including extreme scarcity events.
As AI data center load grows from ~4% to 15-25% of regional peak demand at major hubs, this price-inelastic demand class becomes the marginal price-setter in an increasing number of hours. The CHR quantifies the ceiling toward which prices converge. No existing forward curve model incorporates this variable.
GPU specifications and benchmarks: NVIDIA documentation (H100, B200); MLPerf inference benchmarks.
API pricing: Anthropic (Claude Opus 4.5: $5/$25 per million tokens), OpenAI (GPT-5 class: $1.25/$10), Google (Gemini 2.5 Pro: $1.25/$10). Published pricing as of Q1 2026.
Cloud compute rates: AWS, Google Cloud, Azure published on-demand GPU instance pricing.
Infrastructure costs: Cushman & Wakefield, JLL data center market reports; hyperscaler SEC filings (Meta 10-K, Microsoft 10-K, Alphabet 10-K).
Energy data: EIA (Henry Hub natural gas spot prices, wholesale electricity data).
All inputs are exclusively public data. No proprietary or confidential data sources are used in the CHR calculation.
When referencing CHR values from this publication:
Royal, H. (2026). "CHR Reference Values, Q1 2026." Published at computeheatrate.com, March 2026.
Underlying methodology:
Royal, Hans, The Compute Heat Rate: Quantifying AI-Driven Electricity Price Tolerance and Its Implications for Wholesale Market Repricing (February 28, 2026). Available at SSRN: http://dx.doi.org/10.2139/ssrn.6322318
The Compute Heat Rate (CHR) is independent research by Hans Royal. The full methodology, including workload taxonomy, input derivation, sensitivity analysis, and market implications, is published in Royal (2026), available on SSRN (Abstract ID 6322318).
CHR values are updated quarterly as underlying data changes (GPU specifications, API pricing, cloud compute rates, facility cost benchmarks). Material intra-quarter changes (e.g., significant API price shifts or new GPU releases) may trigger interim updates.
The CHR does not constitute investment advice, a price forecast, or a recommendation. It is an analytical framework measuring demand-side electricity price tolerance.