This report tests a central hypothesis: that AI-driven electricity demand will permanently reprice U.S. wholesale power markets over the next 5–15 years. We introduce the concept of the Compute Heat Rate (CHR): the effective $/MWh at which it remains economically rational for AI operators to electricity, and evaluate its implications for wholesale price formation, capacity markets, and the broader energy economy. More broadly, we propose the CHR as a durable measurement standard: a new metric for the demand side of electricity price formation that will remain analytically important regardless of which specific price scenario materializes.
The report is motivated by a structural observation: AI may represent the most energy-concentrated technological revolution in history. As Google DeepMind CEO Demis Hassabis stated in February 2026, the advent of AGI could deliver “ten times the impact of the Industrial Revolution at ten times the speed.” Unlike prior revolutions that spread energy demands across multiple inputs and sectors, AI concentrates virtually all of its economic activity through a single physical process: converting electricity into computation. Section 2 examines why this concentration creates unprecedented pressure on electricity markets specifically.
The CHR for 2026, defined as the economic tolerance ceiling, the maximum electricity price at which AI workloads remain profitable, ranges from approximately $250–$6,350/MWh depending on workload type. This is 5–127× higher than the gas heat rate ($25–$60/MWh). Wholesale prices are determined by a spectrum between the Cost of New Entry (CONE) and the CHR, depending on the severity of the supply-demand gap. CONE ($80–$130/MWh) sets the price floor: the minimum needed to attract new generation. The CHR sets the ceiling: where AI demand finally curtails. Under the base case, prices settle in the $100–$160/MWh range; under accelerated scenarios, they push toward $150–$250+/MWh.
The demand signal is real and unprecedented. U.S. data centers consumed approximately 183 TWh in 2024, over 4% of national consumption.IEA The IEA projects this will more than double to over 400 TWh by 2030. Of the ~75 GW pipeline, we estimate 35–40 GW is under active construction or has secured financing (see Section 5 for granular breakdown).Modeled The five largest hyperscalers are projected to spend $660–$690 billion on capex in 2026, ~75% on AI infrastructure.CreditSights
Markets are already repricing. PJM capacity auction prices surged 833% from 2024/25 to 2025/26, hitting the FERC-approved cap of $333/MW-day for three consecutive auctions.PJM BRA Data center load accounted for 40% ($6.5B) of $16.4B in PJM’s December 2025 auction costs.Monitoring Analytics PJM fell 6,623 MW short of its reliability target for the first time in history. Bloomberg shows wholesale costs near data centers up as much as 267% over five years.Bloomberg
Supply cannot keep pace in the critical window. SMRs will not reach meaningful scale until the early-to-mid 2030s. Gas turbine lead times extend through 2028–2030. PJM processed 170,000+ MW of interconnection requests since 2023 but only 956 MW of new supply cleared its latest auction.PJM BRA
The CHR effect is a penetration-threshold phenomenon, and we are still early. A natural objection: "if the CHR is so powerful, why haven’t wholesale prices already repriced everywhere?" The answer is the same reason California’s duck curve didn’t exist before enough solar was installed, the gas heat rate didn’t set prices before gas peakers were ubiquitous, and West Texas wind covariance pricing didn’t appear before sufficient wind capacity was built. With ~25–28 GW of operational data center capacity today, the CHR effect is already visible where concentration is highest (PJM: +833% capacity prices, +267% wholesale near DC clusters). As the pipeline delivers 45–55 GW by 2030, the effect will propagate to every market with available capacity. See Section 6.5.
In February 2026, Google DeepMind CEO Demis Hassabis offered what may be the most consequential economic prediction of the decade. Speaking at the India AI Impact Summit, Hassabis described artificial general intelligence as potentially delivering “ten times the impact of the Industrial Revolution, but happening at ten times the speed, probably unfolding in a decade rather than a century.” He framed the moment as comparable to the discovery of fire or the invention of electricity, not an incremental technological advance, but a civilizational phase transition.
This report does not take a position on whether Hassabis’s prediction is precisely correct. But the framing, 10× the magnitude at 10× the velocity, has a profound and under-examined implication for energy markets. The Industrial Revolution transformed energy demand gradually across dozens of industries over more than a century. AI is compressing a comparable or larger transformation into a single decade. And unlike the Industrial Revolution, which diffused its energy demands across textiles, steel, rail, chemicals, and agriculture, AI concentrates virtually its entire economic impact through one physical process: converting electricity into computation.
Every prior general-purpose technology spread its economic impact across multiple energy inputs and physical processes. Consider the breadth of the Industrial Revolution’s energy demands:
| Revolution / Technology | Primary Energy Inputs | Physical Processes | Energy Concentration |
|---|---|---|---|
| First Industrial Revolution (1760–1840) | Coal, water, wood, animal labor | Textile milling, iron smelting, steam transport, agriculture | Highly diffuse across inputs and sectors |
| Second Industrial Revolution (1870–1920) | Coal, oil, electricity, natural gas | Steel, chemicals, electrical generation, internal combustion, telephony | Diffuse; electricity one of several new inputs |
| Digital Revolution (1970–2010) | Electricity (modest share of total) | Semiconductors, networking, software, communications | Moderate; electricity important but total demand modest |
| AI Revolution (2024–) | Electricity (overwhelmingly dominant) | GPU computation, cooling, data storage | Extreme: >95% of energy demand is electricity → compute |
Exhibit 2.0: Energy concentration by technological revolution. AI is historically unique in channeling virtually all of its economic value creation through a single energy conversion pathway.
The steam engine needed coal. The automobile needed oil. The telephone needed copper. The factory needed steel, which needed coke, which needed coal and limestone. Each technology in the Industrial Revolution created demand for different energy inputs processed through different physical pathways. This diffusion meant that no single energy commodity or infrastructure system bore the full weight of the transformation.
AI is structurally different. A large language model does not need oil. It does not need steel (beyond the one-time construction of the data center). It does not need chemicals, wood, animal labor, or natural gas directly. It needs one thing: electricity delivered to GPUs. Every dollar of AI revenue, every token of inference, every training run, every agentic workflow: all of it ultimately resolves to electrons flowing through semiconductor circuits. The economic value is created at the point of electrical-to-computational conversion, and nowhere else.
If Hassabis is directionally correct, 10× the impact at 10× the speed, the compounding effect is not merely 10× or even 20×. It is the product: a transformation 100× more concentrated in time than the Industrial Revolution. But the Industrial Revolution spread its energy demands across at least four major energy inputs (coal, oil, gas, hydroelectric) and dozens of industrial sectors over more than a century. AI concentrates its entire demand footprint through electricity alone, in a decade.
If AI’s economic transformation is 10× larger and 10× faster than the Industrial Revolution, and if it channels that transformation through a single energy input rather than diffusing it across many, then the pressure on electricity markets specifically is not merely 100× greater than any prior technological shift. It is 100× greater concentrated on a single commodity. The grid has never experienced demand growth of this magnitude, at this velocity, from a single use case. This is why the Compute Heat Rate, the metric that quantifies this demand’s price tolerance, becomes an essential analytical tool.
The historical precedent closest to AI’s energy concentration is electrification itself: the Second Industrial Revolution’s adoption of electricity as a general-purpose energy carrier from the 1880s through the 1920s. But even electrification was gradual: in the U.S., electricity provided less than 5% of manufacturing horsepower in 1899, rising to 50% by 1919 and 75% by 1929, a three-decade transition. AI data center demand is projected to more than double U.S. data center electricity consumption (from ~183 TWh to 400+ TWh) in six years.
The policy, investment, and market implications are severe. Grid infrastructure designed for 1–2% annual load growth is confronting 15–25% annual growth in its fastest-expanding demand category. Capacity planning models that assume load diversity across industrial sectors cannot account for a single sector that may represent 8–15% of national consumption by 2030. And electricity procurement frameworks calibrated for price-sensitive industrial demand must contend with a buyer whose willingness-to-pay exceeds all historical precedent. The Compute Heat Rate is the metric that captures this unprecedented dynamic.
The CHR asks: “What is the maximum price a data center operator will rationally pay for a MWh of electricity before the computation becomes uneconomic?” This is analogous to the gas heat rate’s role in generation economics: it defines an boundary, but serves a fundamentally different analytical function.
The CHR is a demand-side tolerance ceiling: it measures the electricity price at which AI workloads become unprofitable, establishing the upper bound of price-inelastic demand. It does not directly set wholesale market prices. Wholesale prices are determined by supply-side economics: the cost of the marginal generator dispatched (short-run) and the cost of new entry (long-run). The CHR’s analytical value is demonstrating AI demand will not curtail at any plausible wholesale price level, making AI load structurally price-inelastic and fundamentally different from traditional industrial demand.
This distinction matters. Traditional price-sensitive industrial loads (aluminum smelters, steel mills) curtail when prices exceed $80–$120/MWh, providing a natural demand-side brake. AI demand, with tolerance ceilings of $250–$6,350/MWh, provides no such brake. When supply is constrained, prices can rise to levels that would previously have triggered demand destruction, but AI demand persists. Supply economics (CONE), not demand willingness-to-pay, becomes the binding constraint (Section 7).
In practice, hyperscalers make procurement decisions based on long-run infrastructure IRR targets rather than short-run workload revenue. A hyperscaler building a $1–$2B data center campus targets project-level unlevered IRRs of ~15–25%, with electricity modeled as a fixed operating cost over a 15–25 year facility life.Industry They negotiate PPAs and utility agreements at $40–$80/MWh, well below the CHR ceiling, because they optimize for long-run infrastructure economics, not marginal workload profitability.
The CHR ceiling matters because as wholesale prices rise to $80, $100, or $150/MWh, AI operators continue to build and operate data centers. They absorb higher electricity costs into infrastructure economics rather than curtailing, because project-level returns remain attractive. This is the mechanism through which the CHR supports the repricing thesis: not by setting prices directly, but by ensuring demand persistence at price levels that would historically have triggered industrial curtailment.
The H100 SXM draws up to 700W per GPU, the H200 is similar, and the B200 is specified at 1,000W though real-world draws ~600W under typical workloads.NVIDIA A standard 8-GPU DGX-class server draws 5.6–8.0 kW GPU power. With CPUs, networking, memory, storage: 8–12 kW total. At PUE 1.2–1.4, facility power per server: ~10–17 kW.
An 8-GPU H100 DGX: $200K–$400K. H200 DGX: $400K–$500K.GPU surveys Amortized over 3–5 years. At 3-year amortization with 70% utilization: ~$13–$19/hour GPU capex per 8×H100. Facility construction: $8–$15M per MW, amortizing to ~$1.5–$3/hr per rack over 15–20 years.
At $40/MWh and PUE 1.3, electricity costs an H100 server ~$0.40/hour. Against cloud rental of $24–$32/server-hour, electricity is 1.2–1.7% of total cost. Even at $200/MWh: only 6–8%. This is why AI demand is so price-inelastic: electricity is a small fraction of producing AI output.
| Cost Component | H100 (8-GPU) | B200 (8-GPU) | % of Total | Source |
|---|---|---|---|---|
| GPU Amortization (3 yr) | $15.20/hr | $19.40/hr | 55% | Empirical |
| Facility Amortization | $2.50/hr | $3.00/hr | 9% | Empirical |
| Networking & Storage | $3.10/hr | $3.50/hr | 11% | Modeled |
| Operations & Labor | $2.00/hr | $2.20/hr | 7% | Modeled |
| Cooling & Power Dist. | $1.20/hr | $1.40/hr | 4% | Modeled |
| Electricity (@ $50/MWh) | $0.46/hr | $0.52/hr | 1.7% | Empirical |
| Operator Margin (30%) | $7.30/hr | $9.00/hr | - | Modeled |
| Total (Cloud Rental) | $31.76/hr | $39.02/hr | 100% |
Exhibit 3.1: Full-stack hourly cost of an 8-GPU AI server, 2026 estimates. Sources: GPU pricing (IntuitionLabs); cloud rates (AWS/GCP/Azure); facility costs (Cushman & Wakefield, JLL). API pricing: Anthropic (Claude Opus 4.5: $5/$25 per M tokens), OpenAI (GPT-5: $1.25/$10 per M tokens), Google (Gemini 2.5 Pro: $1.25/$10 per M tokens).
Regime 1: Frontier API Pricing. 8×H100 at 70% utilization, 7.3 kW facility power (0.0073 MWh/hr). At ~10,000 tokens/secMLPerf and current frontier model blended pricing (~$15/M output tokens for Claude Opus 4.5 / GPT-5 class models; ~$3/M for mid-tier Sonnet/GPT-4.1 classAnthropic, OpenAI): ~$540/hr revenue at frontier pricing, yielding ~$74,000/MWh. Even at mid-tier Sonnet/GPT-4.1 pricing (~$3/$15 per M tokens), revenue is ~$108/hr, yielding ~$14,800/MWh. Note: per-token prices have declined substantially since early frontier models (GPT-4o era: ~$10/M blended), but throughput gains and total API volume have grown faster, consistent with the Jevons dynamic described in Section 4.
Regime 2: Enterprise Contracted. Volume discounts of 60–80%. Blended $1–$3/M tokens at volume discounts of 60–80%: ~$36–$108/hr, yielding ~$4,900–$14,800/MWh. This represents the bulk of current commercial AI revenue.
Regime 3: Internal Enterprise Deployment. Relevant metric is economic value created. An AI system automating $200/hr professional work at $20/hr compute cost creates $180/hr value. McKinsey estimates enterprise AI value creation at $2.6–$4.4T annually by 2030McKinsey, implying thousands of dollars per MWh even for non-API workloads.
Formula: CHR = (Revenue/MWh − Non-Elec Costs/MWh) / (1 + Required Margin)
| Workload Type | Revenue/MWh | Non-Elec Costs | CHR Ceiling | vs. Gas HR | Provenance |
|---|---|---|---|---|---|
| Frontier Inference (Opus/GPT-5 class) | $74,000 | $4,250 | $53,650 | ~1,070× | Empirical |
| Mid-Tier Inference (Sonnet/GPT-4.1 class) | $14,800 | $4,250 | $8,120 | ~162× | Empirical |
| Enterprise Contracted | $5,900 | $4,250 | $1,270 | ~25× | Modeled |
| Commodity Inference (mini) | $1,850 | $4,250 | ~$800* | ~16× | Modeled |
| Enterprise Agentic AI | $15,000 | $4,250 | $8,270 | ~165× | Uncertain |
| Frontier Model Training | $2,000† | $4,250 | ~$500† | ~10× | Modeled |
| Blended Average (2026) | $12,500 | $4,250 | $6,350 | ~127× | Modeled |
Exhibit 3.2: CHR (tolerance ceiling) by workload, 2026. Updated for current frontier pricing (Claude Opus 4.5, GPT-5 class at $15–$25/M output tokens; mid-tier Sonnet/GPT-4.1 class at $3–$15/M). *Commodity inference cross-subsidized by higher-margin workloads. †Training revenue amortized over model lifetime. 30% required margin; gas heat rate ~$50/MWh.
The CHR is not a prediction of where wholesale prices will go. It measures demand-side price inelasticity. Even the lowest-margin workloads imply a ceiling 10–20× above the gas heat rate. Blended CHR of $6,350/MWh (~127× wholesale). AI demand will not curtail below ~$250/MWh. This removes the historical demand-side brake and allows supply-side economics (CONE) to become the binding constraint on long-run pricing.
Aluminum smelting: ~$40–$80 value per MWh consumed, curtails at $80+/MWh. Steel EAF: $60–$120/MWh, moderately price-sensitive. Even the most commoditized AI inference (mini/flash-tier models): over $800/MWh revenue. Mid-tier models (Claude Sonnet, GPT-4.1 class): over $14,000/MWh. PJM and ERCOT data already show data center load persisting through high-price events that curtailed traditional industrial demand.Empirical
The central question for long-term energy market economics is whether the CHR tolerance ceiling remains elevated or declines toward levels where AI demand would become price-sensitive.
NVIDIA Blackwell (B200) delivers ~4× inference performance/watt vs. Hopper (H100).NVIDIA Each generation roughly doubles FLOPS/watt every 2–3 years. By 2030: 8–16× better; by 2035: 30–60×. Algorithmic efficiency has improved even faster: Epoch AI shows GPT-3.5-level performance became 280× cheaper between Nov 2022 and Oct 2024.Epoch AI MoE architectures (DeepSeek-V3: 37B active of 671B parameters) and quantization (FP8, FP4) further reduce per-token costs.
AI per-token costs fell ~1,000× over three years, yet total consumption grew far faster. OpenAI API usage grew ~10% month-over-month through 2024.Industry Hyperscaler AI capex doubled year-over-year for three consecutive years.CreditSights IEA projects global DC electricity doubling from 415 TWh (2024) to 945 TWh (2030).IEA
For Jevons to not hold, demand elasticity must be <1.0. Current evidence suggests 2.0–4.0.Uncertain At elasticity 2.5, a 10× efficiency gain expands demand 31.6×, net 3.16× increase in electricity consumption. Precise elasticity calibration remains uncertain (see Section 11 for elasticity <1.0 scenarios).
| Year | Efficiency (cumul.) | Demand Growth | Net Elec Impact | Blended CHR |
|---|---|---|---|---|
| 2026 | 1.0× | 1.0× | Baseline | $6,350 |
| 2028 | 3× | 5–8× | 1.7–2.7× increase | $2,800–$4,200 |
| 2030 | 10× | 15–30× | 1.5–3.0× increase | $1,500–$2,700 |
| 2033 | 30× | 50–120× | 1.7–4.0× increase | $750–$1,700 |
| 2035 | 60× | 100–300× | 1.7–5.0× increase | $400–$1,000 |
Exhibit 3.1: Projected CHR trajectory. Even in 2035 with 60× efficiency gains, CHR remains 5–12× above CONE-based equilibrium prices.Modeled
Even under aggressive efficiency assumptions, the CHR does not fall below $250/MWh through 2035. Since CONE-based prices project at $80–$130/MWh, the CHR remains 2–3× above equilibrium even in the most bearish case. AI demand stays structurally price-inelastic for the entire analysis horizon.
V1 cited aggregate pipeline figures without distinguishing development stages. This section provides a granular breakdown and tests the thesis under conservative demand realization.
| Development Stage | Est. Capacity (GW) | Confidence | Conversion Rate | Realized (GW) | Source |
|---|---|---|---|---|---|
| Operational | ~25–28 | Very High (>95%) | ~100% | 25–28 | Empirical |
| Under Construction | ~18–22 | High (80–90%) | ~85% | 15–19 | Empirical |
| Financed/Contracted | ~12–18 | Moderate (60–75%) | ~65% | 8–12 | Modeled |
| Permitted | ~8–12 | Lower (40–60%) | ~50% | 4–6 | Modeled |
| Queue/Proposed | ~30–50+ | Low (15–30%) | ~20% | 6–10 | Uncertain |
| Total Pipeline | ~93–130 | 58–75 | |||
| Realized by 2030 | ~45–55 GW | Modeled |
Exhibit 4.1: U.S. data center demand pipeline by development stage. Conversion rates reflect historical PJM/ERCOT queue attrition. Sources: PJM/ERCOT queues; McKinsey, JLL, CBRE; IEEFA; internal modeling.
The headline "75+ GW" in V1 included the full pipeline. The analytically relevant figure is realized demand by 2030: ~45–55 GW (including operational), a 30–40% haircut from raw pipeline. The repricing thesis holds under this more conservative estimate.
At 50% realization (~35–40 GW total by 2030): This still represents ~12–15 GW of new demand, equivalent to ~90–110 TWh of incremental annual consumption. For comparison, PJM’s typical annual peak load growth: 1–2 GW. Even at 50%, AI demand growth is 6–10× historical growth rates in PJM and 4–7× in ERCOT. Supply-side constraints remain identical regardless of demand realization rate.
Conclusion: At 50% haircut, the repricing thesis holds with moderate-to-high confidence at reduced magnitude. Only below ~25% realization (S1 scenario, 8% probability) does the thesis weaken materially.
| ISO/RTO | 2024 DC (GW) | 2030E DC (GW) | DC % of Peak | DC % of Growth | Tipping Point | Severity |
|---|---|---|---|---|---|---|
| PJM | ~8 | 15–25 | ~10% | 70–90% | Already occurring | Critical |
| ERCOT | ~4 | 10–18 | ~5% | 50–70% | 2026–2028 | Critical |
| CAISO | ~2 | 4–7 | ~4% | 30–40% | 2028–2030 | Elevated |
| MISO | ~2 | 4–8 | ~2% | 25–40% | 2029–2032 | Elevated |
| SPP | ~1 | 2–5 | ~2% | 20–35% | 2029–2032 | Moderate |
| NYISO | ~1.5 | 2.5–4.5 | ~4% | 30–45% | 2028–2031 | Elevated |
| ISO-NE | ~1 | 1.5–3 | ~3% | 25–35% | 2029–2032 | Moderate |
Exhibit 5.1: ISO/RTO assessment. 2030 estimates revised downward from V1 to reflect conversion-rate-adjusted demand. Sources: PJM BRA (Dec. 2025); ERCOT CDR (Dec. 2025); EIA; Wood Mackenzie; E3/JLARC.
Northern Virginia handles ~70% of global internet traffic; data centers are the largest driver of PJM load growth.PJM Capacity prices hit $333/MW-day cap for three consecutive auctions. DC load: $6.5B (40%) of $16.4B total auction costs; $6.2B for DCs not yet built.Mon. Analytics Peak load forecast: 5,250 MW higher YoY, ~5,100 MW from DCs. 170,000+ MW generation requests, 956 MW cleared (<1% increase). 6,623 MW short of 20% reserve target, first time in history. FERC Chairman Swett called the situation "very concerning."FERC
December 2025 CDR: reserve margins cross negative territory beginning 2028 (base scenario). TSP-Provided forecast: -6.2% for summer 2026.ERCOT CDR EIA projects 45% wholesale price increase at ERCOT-North in 2026.EIA Forwards: above $50/MWh, summer on-peak $110–$165/MWh in some hubs.ICE
The tipping point occurs when DC demand pushes the supply stack onto its steep portion. In most ISOs, the stack is flat through gas CC ($25–$50/MWh), then steepens through peakers ($80–$150/MWh) and scarcity pricing ($200–$5,000/MWh). DCs don’t need to be a majority of total load; they need to make the system chronically short at the margin. In PJM: already occurred. ERCOT: imminent. Others: 3–6 years under current trajectories.
A skeptic reading this report might reasonably ask: if the CHR implies such extraordinary willingness-to-pay, and if AI demand is as price-inelastic as we claim, why haven’t wholesale markets already repriced to dramatically higher levels? This is the most important timing question the report must address, and the answer illuminates why the current moment represents an opportunity rather than a refutation of the thesis.
The CHR is a penetration-threshold effect. It will reshape market pricing progressively as the installed base of data centers grows, not all at once, and not before sufficient physical capacity is built and energized. This is not a novel market dynamic. Every major technology-driven repricing event in electricity markets has followed the same pattern.
In 2008, CAISO published the now-famous "duck curve" projection. At the time, California had roughly 1 GW of solar capacity and the effect was negligible. Critics dismissed it. By 2015, with ~10 GW installed, the curve was unmistakable. By 2023, with ~35 GW, it was the defining feature of California’s price formation. The underlying physics was always true. But it only became a market phenomenon once enough panels were physically installed.
The gas heat rate became the dominant price-setting mechanism only after decades of gas turbine buildout. In the 1980s, coal set marginal prices in most hours across most ISOs. It was the sustained construction boom of combined-cycle and peaker turbines through the 1990s and 2000s that eventually made gas the marginal fuel in most hours in most markets.
In ERCOT’s West Zone, wholesale prices became negatively correlated with wind output once sufficient wind capacity (~15–20 GW) was installed in a transmission-constrained corridor. Before that threshold, wind was a marginal contributor with no observable effect on zonal pricing.
| Market Signal | Where Visible | Installed Base Context |
|---|---|---|
| Capacity prices at regulatory cap | PJM (~8 GW DC load, ~10% of peak) | Highest U.S. DC concentration |
| Wholesale +267% near DC clusters | PJM nodes within 50 mi of DCs | N. Virginia: ~70% of global internet traffic |
| Reserve margins turning negative | ERCOT (~4 GW, ~5% of peak) | Fastest-growing DC market in U.S. |
| Forward curves in persistent contango | PJM and ERCOT | Markets where DC pipeline is largest |
| Minimal price effect observed | SPP, ISO-NE, MISO interior | Low DC penetration (<2% of peak) |
Exhibit 5.2: The CHR effect is proportional to local data center penetration. Markets with high DC concentration already show repricing; markets with low penetration do not, yet.
The fact that the CHR effect has not yet fully materialized in every market is not a weakness of the thesis: it is the thesis itself. Markets have not yet priced in the effect because the installed base has not yet crossed the critical threshold in most ISOs. By the time the effect is undeniable in every market, the repricing will already be reflected in forward curves and capacity markets. The window exists precisely because the market is in the early-to-middle phase of the penetration curve.
The strongest version of the installed base thesis: any pocket of available electricity capacity in the U.S. will eventually have a data center sited on it. When a data center operator generates $1,850–$49,300 of revenue per MWh consumed, and the prevailing electricity price is $30–$60/MWh, the economics of siting a data center adjacent to any available power source are overwhelmingly favorable. This is precisely what we observe: hyperscalers are restarting retired nuclear plants (Constellation’s Three Mile Island restart for Microsoft), negotiating behind-the-meter arrangements, siting in previously undesirable locations (rural Mississippi, West Texas), and even purchasing their own generation assets.
This section synthesizes the analytical framework across three report iterations. V1 argued the CHR directly sets prices. V2 corrected to a CONE-based framework. V3 synthesizes both: the actual wholesale price falls on a spectrum between CONE and the CHR, depending on the degree of supply-demand imbalance.
| Technology | Overnight ($/kW) | CONE ($/MWh) | Timeline | Notes |
|---|---|---|---|---|
| Gas CC (new) | $1,100–$1,400 | $60–$90 | 3–5 yr | Turbine backlogs extend to 2028–2030 |
| Gas Peaker (CT) | $800–$1,100 | $100–$150 | 2–4 yr | High operating cost; peak hours only |
| Solar + Storage (4hr) | $1,600–$2,200 | $55–$85 | 2–4 yr | Tariff risk (OBBBA, AD/CVD) |
| Onshore Wind | $1,300–$1,700 | $40–$65 | 3–5 yr | Declining IRA credits |
| SMR | $6,000–$12,000 | $80–$150 | 8–15 yr | Not material before early-mid 2030s |
| Enhanced Geothermal | $4,000–$8,000 | $60–$120 | 5–10 yr | Pre-commercial scale |
| Blended CONE | $80–$130 | Weighted by likely technology mix |
Exhibit 6.1: CONE by technology. Sources: Lazard LCOE+ 2025; EIA AEO; NREL ATB 2025; Wood Mackenzie. Post-OBBBA policy assumptions.
The CONE boundary ($80–$130/MWh) represents the minimum sustainable long-run price. Below CONE, no one builds new generation, supply deteriorates, and scarcity pushes prices back up. CONE is the gravitational floor.
The CHR boundary ($250–$6,350/MWh) represents the maximum sustainable price: the level at which the marginal AI buyer finally curtails. In a market where supply never fully catches demand, prices migrate upward from CONE toward the CHR.
| Supply-Demand Balance | Price-Setting Mechanism | Expected Price Range | Scenario Mapping |
|---|---|---|---|
| Surplus | Generator competition drives price to marginal fuel cost | $30–$60/MWh | S1 (8% prob) |
| Mild shortage | CONE attracts new investment; modest scarcity premium | $80–$110/MWh | S2 (15% prob) |
| Moderate shortage | Scarcity pricing; generators have market power; AI demand absorbs spikes | $100–$160/MWh avg | S3 (40% prob) |
| Severe shortage | Persistent scarcity; marginal price set by highest-value buyer | $150–$250+/MWh avg | S4/S5 (37% prob) |
| Structural crisis | CHR becomes the binding constraint | $250+/MWh avg | S5 extreme tail |
Exhibit 6.2: The CONE-to-CHR pricing spectrum. 77% of probability weight falls in mild-to-severe shortage.Modeled
Lead time asymmetry: A hyperscaler can build a new data center in 18–24 months. A new gas combined-cycle takes 3–5 years. Transmission upgrades take 7–12 years. SMRs are 8–15 years away. Demand grows on a 2-year cycle; supply responds on a 5–15-year cycle.
Jevons Paradox compounds the problem: New generation that does get built enables more AI demand. A new 1 GW gas plant doesn’t "solve" the problem; it enables more data centers to interconnect, which eventually requires another 1 GW.
Investment constraints beyond price signals: Even at elevated prices, the build-out is constrained by gas turbine manufacturing capacity (backlogs through 2028–2030), interconnection queue processing (4–5 year average in PJM, 170,000+ MW pending), transmission siting (7–12 years), and supply chain constraints on transformers and switchgear.
| Market Condition | Base Hours ($/MWh) | Elevated Hours | Scarcity Hours | Annual Hub Avg |
|---|---|---|---|---|
| Historical (balanced grid) | ~7,500 hrs @ $35 | ~1,000 hrs @ $80 | ~260 hrs @ $300 | $48 |
| Moderate shortage (S3) | ~6,000 hrs @ $55 | ~2,000 hrs @ $120 | ~760 hrs @ $600 | $117 |
| Severe shortage (S4/S5) | ~5,500 hrs @ $65 | ~2,200 hrs @ $150 | ~1,060 hrs @ $900 | $189 |
Exhibit 6.3: Illustrative price duration curve analysis for ERCOT. During scarcity hours, the CHR directly influences clearing prices as AI load does not curtail.Modeled
The CONE-to-CHR spectrum framework has broad implications across the energy economy.
Traditional industrial electricity consumers, including aluminum smelters, steel mills, chemical plants, and data processing facilities, face a structural repricing of their primary input cost. Unlike AI operators, these industries cannot absorb $100–$160/MWh electricity. Their economic tolerance ceilings are 3–10× lower. This creates a two-tier electricity economy: sectors with high revenue-per-MWh (AI, cloud computing) will persist through any plausible price level, while traditional industry faces margin compression or geographic relocation.
PJM forwards: CY2028 trading at $8–$10/MWh premium to 2025. ERCOT shows similar separation since mid-2023.ICE Bloomberg nodal analysis: >70% of nodes with largest price increases since 2020 are within 50 miles of significant DC activity.Bloomberg Solar/wind zero-marginal-cost generation compresses some hours, but AI 24/7 flat-load demand pulls up averages and peaks.
Under CONE-based price projections, a fixed-price instrument executed at current wholesale levels ($45–$75/MWh) generates positive NPV in over 80% of probability-weighted scenarios across a 15-year horizon.Modeled The downside case requires near-zero AI demand growth (<10% probability). Third-party curves that assume mean-reversion are underpricing future wholesale electricity by $20–$60/MWh.
| Fixed Price | S1: No Growth | S2: Moderate | S3: Base | S4: Accelerated | S5: Supercycle | Prob-Wtd NPV |
|---|---|---|---|---|---|---|
| 15-Year Solar Equivalent: ERCOT (NPV/MWh) | ||||||
| $45/MWh | −$3 | +$12 | +$28 | +$52 | +$85 | +$29 |
| $55/MWh | −$13 | +$2 | +$18 | +$42 | +$75 | +$19 |
| $65/MWh | −$23 | −$8 | +$8 | +$32 | +$65 | +$9 |
| $75/MWh | −$33 | −$18 | −$2 | +$22 | +$55 | +$1 |
| 15-Year Wind Equivalent: PJM (NPV/MWh) | ||||||
| $45/MWh | +$2 | +$18 | +$35 | +$60 | +$95 | +$36 |
| $55/MWh | −$8 | +$8 | +$25 | +$50 | +$85 | +$26 |
| $65/MWh | −$18 | −$2 | +$15 | +$40 | +$75 | +$16 |
| $75/MWh | −$28 | −$12 | +$5 | +$30 | +$65 | +$7 |
Exhibit 7.1: NPV analysis for fixed-price procurement. Weights: S1=8%, S2=15%, S3=40%, S4=25%, S5=12%. 7% discount rate, 15-yr term, hub settlement.Modeled
OBBBA accelerates IRA credit expiration; BNEF estimates 23% reduction in renewable deployment by 2035.BNEF AD/CVD on solar modules adds 15–30% to costs.LevelTen These headwinds tighten markets further, reinforcing the repricing dynamic.
The most likely regulatory interventions, including differentiated tariffs, curtailment authority, and capacity price caps, suppress visible prices for residential customers but do not fundamentally alter wholesale hub pricing. The only intervention that addresses the root cause (supply-side acceleration) has a 3–7 year implementation lag. See Section 9.2 for the full regulatory deep dive.
Per-token costs fell 1,000× yet consumption grows exponentially. Demand elasticity est. 2.0–4.0 (calibration uncertain; see Section 10). Addressable market expanding. For Jevons to fail, AI must reach saturation, unlikely at earliest enterprise adoption stages.
Verdict: Efficiency slows growth but doesn’t reverse it through 2035.
This is the most likely counterargument to materially affect outcomes. Politicians will not keep their jobs at $150/MWh residential power prices. Regulatory response is not a question of if but what form and how effective.
See deep dive below.
| Intervention Type | Precedent | Price Impact | Duration | Probability |
|---|---|---|---|---|
| Capacity Price Caps | PJM Shapiro cap ($333/MW-day) | Suppresses capacity prices 20–40% | 2–5 years | Already occurring (PJM) |
| Energy Price Caps | EU energy crisis caps (2022–23) | Caps real-time prices at set level | 1–3 years | Moderate in crisis |
| Differentiated Tariffs | Virginia proposed DC surcharges (2025) | Shields residential; shifts costs to DCs | Potentially permanent | High (30–50%) |
| Data Center Moratoria | Singapore moratorium (2019–2022) | Reduces demand growth locally | 2–5 years | Moderate (20–30%) |
| Mandatory Curtailment | ERCOT SB6 (2025) | Suppresses scarcity spikes | Ongoing | High in ERCOT; spreading |
| Expedited Permitting | EU REPowerEU; U.S. FAST-41 | Increases supply, moderates long-run prices | 3–7 year lag | Moderate (20–35%) |
| Market Redesign | UK Capacity Market; PJM CIFP | Restructures price formation | 3–5 years to implement | Moderate (25–40%) |
| Generator Windfall Taxes | EU inframarginal revenue caps (2022) | Caps generator profits, not market price | 1–3 years | Low–Moderate |
Exhibit 8.1: Taxonomy of regulatory interventions, precedents, and market impacts.
The central irony of most price-capping interventions is that they worsen the underlying problem. The Shapiro cap in PJM: by capping capacity prices below the uncapped clearing price (~$530/MW-day), the cap reduces revenue to incentivize new generation, precisely when PJM is 6,623 MW short of its reliability target. The EU energy crisis of 2022–2023 provides a real-world stress test: caps provided short-term relief but delayed investment. Within 18 months, most caps were relaxed as policymakers recognized they were prolonging shortages.
Near-term (2026–2028): Differentiated tariffs, expanded curtailment authority, continued capacity price caps. These suppress visible residential price increases but do not alter wholesale hub dynamics.
Medium-term (2028–2032): Political pressure shifts to supply-side acceleration: expedited permitting, streamlined interconnection. This is the intervention that actually works, but takes 3–7 years to translate into operating generation. The repricing window remains open.
Long-term (2032+): New supply arrives at scale. Regulatory posture shifts to market design reform.
SMRs: early-mid 2030s. Gas turbine backlogs: 2028–2030. PJM queue: 170K+ MW, 4–5 yr timelines. OBBBA reduces renewables ~23% by 2035 (BNEF).
Verdict: Supply responds eventually but 8–12+ year structural lag.
Per-token prices down 90%+. But total revenue growing faster: OpenAI ARR tripled to ~$20B in 2025. Volume outpaces price decline. Enterprise agents command premiums.
Verdict: CHR may compress but stays above CONE levels.
IEEFA: PJM forecasts may be inflated. ERCOT: load forecast "unrealistic." But: Section 4 granular breakdown with stage-appropriate conversion rates. Even after 30–40% haircut, 45–55 GW by 2030 exceeds supply additions. Hyperscaler capex ($660–$690B) not speculative. DC vacancy: 1.6% record low. 75% under construction preleased.
Verdict: Overstated 30–50% but thesis holds at reduced magnitude.
Inference: latency-sensitive, near users. Training: mobile but needs massive power. U.S.: ~50% of global DC compute, 75% of giga-scale AI DCs under construction.
Verdict: Real but insufficient to relieve U.S. grid pressure.
OBBBA IRA phase-downs and AD/CVD on solar modules. If credits expire faster, renewable deployment slows. But: This reinforces the repricing thesis: constrained supply tightens markets.
Verdict: Threat to renewables but tailwind for repricing.
Major efficiency breakthroughs (DeepSeek-class) could create 12–18 month windows where gains temporarily outpace demand.
Verdict: Temporary pauses, long-run trajectory intact.
AI bubble bursts. Demand stalls. Supply catches up. Wholesale reverts to $30–$45/MWh by 2030.
Efficiency partially offsets demand. Load growth 3–5% p.a. Supply responds. Wholesale $50–$70/MWh by 2030.
Demand outpaces supply 2–3× through 2032. Prices converge to CONE: $80–$100/MWh by 2030, $90–$130 by 2035.
Demand exceeds forecasts. Supply bottlenecks worsen. Prices overshoot CONE to $100–$150/MWh by 2030.
Full demand acceleration. 5–10× by 2030. Wholesale exceeds $150/MWh with frequent $500+ scarcity events.
ERCOT hub: $75–$95/MWh by 2030, $90–$120/MWh by 2035. PJM hub: $80–$105/MWh by 2030. These are $20–$60/MWh above third-party curves (Wood Mackenzie, ABB Horizons, LevelTen).Modeled
| Demand Realization | Realized GW by 2030 | ERCOT Wholesale 2030 | Thesis Impact |
|---|---|---|---|
| 100% (Headline) | ~45–55 | $95–$130 | Strongly confirmed |
| 75% (Moderate haircut) | ~35–42 | $80–$110 | Confirmed |
| 50% (Severe haircut) | ~25–30 | $65–$90 | Holds, reduced magnitude |
| 30% (Extreme haircut) | ~15–18 | $50–$70 | Marginal |
| 15% (Near-collapse) | ~8–10 | $35–$50 | Thesis fails |
Exhibit 10.1: Demand realization sensitivity.Modeled
| Elasticity | Effect of 10× Efficiency | Net Elec Change | CHR 2030 | Thesis Impact |
|---|---|---|---|---|
| 4.0 (Strong Jevons) | Demand 100× | 10× increase | $1,200–$1,800 | Strongly confirmed |
| 2.5 (Base case) | Demand 31.6× | 3.16× increase | $1,500–$2,700 | Confirmed |
| 1.5 (Moderate) | Demand 10× | Flat (1.0×) | $500–$900 | Holds; demand plateau |
| 1.0 (Unit elastic) | Demand 10× | Flat (1.0×) | $400–$700 | Holds; elec constant |
| 0.5 (Anti-Jevons) | Demand 3.16× | 0.32× decrease | $200–$400 | Weakens; CHR nears CONE |
Exhibit 10.2: Jevons elasticity sensitivity. Even at 0.5 elasticity, CHR stays above CONE.Modeled
| Supply Scenario | New GW by 2030 | Reserve Impact | Price Effect | Thesis Impact |
|---|---|---|---|---|
| Constrained (base) | 15–25 | Tighten 5–10% | +$30–$70/MWh | Confirmed |
| Moderate response | 30–40 | Stabilize | +$15–$40/MWh | Reduced magnitude |
| Aggressive build-out | 50–65 | Slightly improve | +$5–$20/MWh | Marginal |
| Unprecedented surge | 75+ | Restored | $0–$10/MWh | Thesis fails |
Exhibit 10.3: Supply sensitivity. "Unprecedented surge" requires gas turbine production tripling, queue timelines halving, and SMR fleet deployment by 2030, simultaneously. Est. <5% probability.Modeled
Thesis failure requires simultaneously: (a) demand realization <30%, (b) Jevons elasticity <1.0, AND (c) unprecedented supply surge. Joint probability: <3%.
The thesis is robust across a wide range of assumptions. Fails only under combined extreme pessimism (<3% joint probability). The repricing dynamic exhibits positive expected value in >85% of probability-weighted scenarios.
The CHR tolerance ceiling is real and enormous. At $250–$6,350/MWh (blended: ~$6,350 in 2026, declining to ~$400–$1,000 by 2035), AI operators’ economic tolerance exceeds projected equilibrium by an order of magnitude. AI demand is structurally price-inelastic.
Markets will reprice along the CONE-to-CHR spectrum. CONE ($80–$130/MWh) is the price floor: the minimum needed to attract new generation. Because supply cannot catch demand within the analysis horizon, prices migrate up the spectrum toward the CHR. Under the base case: $100–$160/MWh. Under accelerated scenarios: $150–$250+/MWh. During scarcity hours specifically, the CHR directly influences clearing prices as AI data centers bid well above generator costs rather than curtail.
AI demand is already repricing markets. PJM capacity +833% in two years. Wholesale near DCs +267% over five years. ERCOT forwards imply doubling over eight years. Observed market data, not projections.
The CHR effect is a penetration-threshold phenomenon, and we are in the early-to-middle phase. The repricing is visible today where data center density is highest (PJM, Northern Virginia) and emerging where density is growing fastest (ERCOT). The installed base is growing at 15–20 GW per year. Any remaining pocket of available capacity will attract a data center, and the CHR dynamics will follow. The fact that markets have not yet fully priced this effect is the opportunity, not the counterargument.
Supply cannot respond fast enough. Even at 50% demand haircut, structural lag ensures tight markets through early-mid 2030s. Thesis holds under all but the most extreme combined stress tests (<3% joint probability).
1. A new era of wholesale electricity pricing. The gas heat rate has been the dominant price-formation mechanism in U.S. wholesale markets for three decades. AI demand introduces a second, fundamentally different price signal: one driven not by fuel costs but by the economic value of computation. The interaction between these two regimes will define market dynamics for the next decade.
2. The emergence of a two-tier electricity economy. Energy-intensive industries with low revenue-per-MWh face structural margin compression. Industries that can generate high value from each MWh consumed (AI, cloud computing, advanced manufacturing) will outcompete traditional load for access to constrained electricity supply.
3. Geographic arbitrage becomes the dominant strategy. Markets with varying levels of data center penetration will exhibit dramatically different pricing dynamics. Settlement points adjacent to major DC buildout will reprice first and most aggressively. Electricity procurement strategies that account for hub-level demand intelligence will dramatically outperform those based on aggregate forecasts.
4. Fixed-price procurement instruments executed at current levels carry radically asymmetric risk-reward. Current wholesale pricing reflects a world before the installed base crosses the critical threshold. Any mechanism that locks in sub-$75/MWh electricity represents a structural hedge against a repricing event that is more likely than not to materialize. The cost of hedging is minimal; the cost of not hedging is potentially enormous.
5. Consensus forecasts are structurally mispriced. Third-party forward curves embed an implicit assumption that AI demand is either a bubble or that supply will catch demand quickly. Both assumptions are contradicted by observed data. The $20–$60/MWh gap between consensus projections and CONE-based demand scenarios is not a rounding error: it is a structural analytical failure that creates opportunity for those who recognize it.
6. The Compute Heat Rate itself is the foundational contribution. Independent of any specific price projection, the CHR establishes a new measurement standard for the demand side of electricity price formation. The gas heat rate has defined the supply side for thirty years. The CHR defines the demand side for the era of AI-driven electricity consumption. As formalized in Section 13, this metric provides the analytical infrastructure for ongoing measurement, academic research, and market design, ensuring that even as specific projections are revised with new data, the framework for understanding AI’s relationship to electricity markets remains rigorous and durable.
The preceding sections apply the CHR to specific market projections and scenario analyses. This section makes a different and more fundamental argument: the Compute Heat Rate is a valuable metric regardless of whether any particular price projection proves correct. Even if wholesale markets do not reprice to the levels modeled in Sections 7–10, the CHR captures a structural economic relationship that did not exist before AI and that will persist for as long as computation generates economic value from electricity consumption.
The Compute Heat Rate (CHR) is defined as the maximum price per megawatt-hour of electricity at which a given AI workload remains economically viable, calculated as:
CHRw = (Rw − Cnon-elec) / (1 + m)
Where:
CHRw = Compute Heat Rate for workload type w ($/MWh)
Rw = Revenue or economic value generated per MWh of electricity consumed by workload w
Cnon-elec = All non-electricity costs per MWh (GPU amortization, facility, labor, networking, cooling)
m = Required operator margin (typically 0.20–0.40)
The aggregate market-level CHR is computed as a revenue-weighted average across workload types:
CHRblended = Σ (wi × CHRi) where wi = share of total AI electricity consumed by workload i
| Input Variable | Data Source(s) | Update Frequency | Confidence |
|---|---|---|---|
| GPU power consumption (watts) | NVIDIA specs, MLPerf benchmarks, independent testing | Per hardware generation (12–18 months) | High |
| Facility PUE | Uptime Institute surveys, operator disclosures | Annual | High |
| Cloud rental / API pricing | AWS, GCP, Azure, OpenAI, Anthropic published pricing | Monthly (or more frequent) | High |
| GPU capital cost | OEM pricing, secondary market surveys, earnings disclosures | Quarterly | Moderate |
| Inference throughput (tokens/sec) | MLPerf, NVIDIA TensorRT-LLM benchmarks, independent benchmarks | Per hardware generation | Moderate–High |
| Workload revenue mix | Hyperscaler earnings, industry surveys, API traffic estimates | Quarterly | Moderate |
| Enterprise value per MWh (non-API) | McKinsey/BCG enterprise AI adoption surveys, case studies | Annual | Low–Moderate |
Exhibit 13.1: CHR input data requirements and sources. All inputs are derived from publicly available or commercially obtainable data, enabling independent replication.
The CHR’s value as a metric is separable from any specific claim about future wholesale prices. Three distinct analytical uses are independent of the price projections in this report:
1. Demand-side price elasticity measurement. The CHR quantifies, for the first time, the price tolerance of AI-driven electricity demand. Whether wholesale prices rise to $100/MWh or remain at $50/MWh, the fact that AI demand tolerates $250–$6,350/MWh tells market participants something fundamental about how this load class will behave during scarcity events, capacity shortfalls, and price spikes. No other metric captures this.
2. Structural demand classification. Electricity markets have long classified demand by sector (residential, commercial, industrial) and by price sensitivity (baseload, interruptible, curtailable). The CHR introduces a new classification axis: revenue per MWh consumed. This enables analysts, ISOs, regulators, and market designers to distinguish between load classes that will curtail at $80/MWh and load classes that will persist through $500/MWh: a distinction that becomes operationally critical as AI load grows from 4% to potentially 10–15% of national consumption.
3. Technology transition tracking. As GPU architectures evolve, algorithmic efficiency improves, and AI workload mixes shift, the CHR trajectory provides a real-time signal of whether AI demand is becoming more or less price-elastic. A declining CHR over time (as modeled in Section 4) would indicate that efficiency gains are outpacing demand growth, a leading indicator that the repricing dynamic is weakening. A stable or rising CHR would confirm the Jevons Paradox thesis. The metric provides an objective, calculable answer to what is currently a speculative debate.
Prior to this work, no formal metric existed to quantify the electricity price tolerance of computation-driven demand. The gas heat rate, the dominant analytical tool for wholesale price formation for three decades, captures only the supply side: the cost of converting fuel to electricity. The Compute Heat Rate captures the demand side: the value of converting electricity to economic output. Together, these two metrics define the upper and lower boundaries of wholesale price formation in markets where AI load is material. The CHR is the missing half of the price formation equation.
The CHR is not without precedent as a type of measurement. Energy markets have a history of standardized metrics that quantify economic boundaries and became foundational to how markets operate:
| Metric | What It Measures | Market Function | Year Established |
|---|---|---|---|
| Gas Heat Rate | Fuel cost per MWh of generation | Sets marginal clearing price in wholesale electricity | ~1990s (as dominant price signal) |
| CONE (Cost of New Entry) | All-in cost to build new generation | Sets capacity market clearing prices, informs long-run equilibrium | ~2000s (PJM, ISO-NE) |
| LCOE (Levelized Cost of Energy) | Lifetime cost per MWh of a generation asset | Investment benchmarking, technology comparison | ~2008 (Lazard annual publication) |
| VIX (Volatility Index) | Implied volatility of S&P 500 options | Risk measurement, derivatives trading, portfolio hedging | 1993 (CBOE) |
| Compute Heat Rate (CHR) | Demand-side price tolerance of AI electricity consumption | Demand classification, elasticity measurement, price formation boundary | 2026 (this report) |
Exhibit 13.2: The CHR in context of established market metrics. Each of these metrics became foundational to how its respective market operates. The CHR addresses a measurement gap that will become increasingly important as AI electricity demand grows.
The CHR as specified in this report is a first-generation formulation. Several extensions are natural candidates for further development and formal academic study:
Hub-level CHR. The blended national CHR masks significant geographic variation. Data center clusters in Northern Virginia, Central Texas, and the Phoenix metro area face different supply constraints, price dynamics, and regulatory environments. A hub-level or ISO-level CHR would provide higher-resolution signal for capacity planning and price formation analysis. Input requirements include hub-level workload mix estimates and local PUE/facility cost data.
Real-time CHR index. With the appropriate data infrastructure, the CHR could be calculated and published at regular intervals (monthly or quarterly), similar to how Lazard publishes LCOE annually, or how ICE publishes forward curves daily. A real-time CHR index would serve as an early-warning system for demand-side pressure on electricity markets, flagging when the CHR-to-wholesale price spread is widening (indicating increasing disequilibrium) or narrowing (indicating market adjustment).
International CHR. AI data center demand is global, but electricity market structures vary enormously across jurisdictions. A comparative CHR analysis across the U.S., Europe, Asia-Pacific, and the Middle East would inform both hyperscaler siting decisions and national energy policy. European markets, with higher baseline electricity prices and different regulatory frameworks, may experience CHR dynamics earlier or differently than U.S. markets.
Empirical validation and calibration. The current CHR calculation relies on publicly available data on GPU economics, API pricing, and cloud rental rates. Future work should incorporate proprietary data: actual hyperscaler procurement costs, internal workload economics, and realized curtailment behavior during price events, to calibrate and validate the theoretical framework. An academic study with ISOs or hyperscaler cooperation could establish the CHR’s empirical foundation at the level of rigor required for regulatory and market design applications.
Workload disaggregation. The workload taxonomy in Section 3 (frontier inference, enterprise contracted, commodity, training, agentic) is a first approximation. As the AI industry matures, more granular workload classification, including emerging categories such as embodied AI, scientific simulation, and autonomous systems, will be necessary to maintain the metric’s analytical precision.
The Compute Heat Rate, as formally specified in this report, represents an original contribution to energy market analysis. The methodology: bottom-up GPU economics yielding a demand-side price tolerance metric, measured in $/MWh and directly comparable to the gas heat rate, has no published precedent. This report establishes the first formal definition, calculation methodology, and application framework for the CHR, and proposes its ongoing development as a standardized, published benchmark for the intersection of AI economics and electricity market dynamics.
PJM capacity auction results, clearing prices, reliability metrics (PJM official results, December 2025). ERCOT CDR reports, reserve margins, load forecasts (ERCOT official publications). GPU specifications, power consumption (NVIDIA documentation, MLPerf). Cloud rental rates (AWS/GCP published pricing). API pricing (OpenAI, Anthropic, Google published pricing). Hyperscaler capex guidance (SEC filings, earnings). Forward power prices (ICE strips, EIA). LevelTen PPA indices. DC vacancy and construction (CBRE, JLL, Cushman & Wakefield). Bloomberg nodal analysis.
CHR calculations (bottom-up GPU economics + API revenue/MWh). CONE-based equilibrium projections (technology costs from Lazard, NREL, EIA). Demand pipeline conversion rates (historical PJM/ERCOT queue attrition). Wholesale price scenarios (CONE-based paths, probability-weighted, 7% discount). Jevons demand growth (observed 2022–2025 elasticity, extrapolated).
Precise AI compute demand elasticity (est. 2.0–4.0, limited data). Agentic workload economics (nascent market). Jevons timing (long-run direction supported; precise timing uncertain). Regulatory trajectory (inherently unpredictable). OBBBA implementation details. SMR deployment timeline and costs. Geographic distribution of future AI demand.
CHR relies on current AI pricing models, which evolve rapidly. CONE model assumes competitive dynamics; regulation, vertical integration, or market design changes could alter transmission mechanism. Scenario probabilities reflect informed subjective assessment, not actuarial precision. Analysis focuses on U.S.; international dynamics could affect domestic thesis through demand migration.
EIA, STEO and Electricity Monthly Update, Nov 2025 / Jan 2026. PJM, 2027/2028 BRA Results, Dec 17 2025. PJM, Critical Issue Fast Path, 2025. ERCOT, CDR Reports, May & Dec 2025. ERCOT, Long-Term Load Forecast, Apr 2025. FERC, Open Meeting (Chairman Swett), Dec 18 2025. Monitoring Analytics, capacity auction analysis, Oct 2025 & Jan 2026. NREL, ATB 2025.
Goldman Sachs, “AI to Drive 165% Increase in DC Power Demand,” Feb 2025. Goldman Sachs, “How AI Is Transforming Data Centers,” Aug 2025. Wood Mackenzie, US Power Market Outlooks, Oct & Dec 2025. IEA, “Energy and AI,” Apr 2025. CreditSights, “Hyperscaler Capex 2026,” Nov 2025. LevelTen, PPA Price Index, Q1–Q4 2025. Lazard, LCOE+ 2025. McKinsey, DC capacity forecasts & “Economic Potential of Generative AI” (2023, updated 2025). Deloitte, “As Generative AI Asks for More Power,” 2024/2025. Bloomberg Intelligence, DC nodal price analysis, Sep 2025. Ascend Analytics, ERCOT Outlook 2025. BNEF, IRA rollback impact, 2025. CBRE, JLL, Cushman & Wakefield, H2 2025. Wells Fargo, AI power demand research, 2025.
WRI, “Powering the US Data Center Boom,” 2025. National Center for Energy Analytics, Nov 2025. Epoch AI, inference price & training cost research, 2024–2025. SIGARCH, Jevons Paradox analysis, Jul 2025. Luccioni et al., arXiv:2501.16548, 2025. Energy for Growth Hub, SMR overview, Sep 2025. E3/JLARC analysis, 2025. IEEFA, demand forecast analysis, 2025.
Amazon, Alphabet, Microsoft, Meta, Oracle earnings through Q4 2025 / Q1 2026. Constellation, Vistra, Talen SEC filings post-PJM Dec 2025 auction. ICE ERCOT forwards. PJM forwards (EIA, Energy by 5). NYMEX Henry Hub. GPU pricing: IntuitionLabs, Clarifai, GMI Cloud. Cloud rates: AWS, GCP. MLPerf Inference v4.0. NVIDIA TensorRT-LLM benchmarks. OpenAI, Anthropic, Google API pricing.
Wissner-Gross & Diamandis, “Solve Everything: Achieving Abundance by 2035” (2026).