- Executive Summary
- The 100× Thesis
- Definition & Measurement
- CHR Trajectory: 2026–2035
- AI Demand: Granular Assessment
- When AI Demand Reshapes the Supply Stack
- The CONE-to-CHR Spectrum
- Market Implications
- Counterarguments & Stress Tests
- Scenario Modeling
- Conclusions & Measurement Standard
- CHR vs. PUE
- The SLA Lockdown
- The Peaker Paradox
- The Convergence Theorem
- Comprehensive Q&A
- Sources & References
1. Executive Summary
This report introduces the Compute Heat Rate (CHR™): a new metric for the demand side of electricity price formation. The CHR measures the maximum electricity price at which a given AI compute workload remains economically viable, expressed in $/MWh and directly comparable to the gas heat rate that has governed the supply side of wholesale markets for three decades. The gas heat rate converts fuel cost into generation cost. The CHR converts compute revenue into a maximum tolerable electricity input cost. Together, they define the boundaries of wholesale price formation in any market where AI load is material.
The CHR matters regardless of whether wholesale markets ultimately reprice to the levels some analysts project. Even if prices remain moderate, the CHR provides the analytical infrastructure that regulators, market designers, procurement professionals, and hedging desks need to understand how this structurally different demand class will behave during scarcity events, capacity shortfalls, and price spikes. No other metric captures this. The CHR is the missing half of the price formation equation.
The CHR for Q1 2026, defined as the economic tolerance ceiling, ranges from approximately $250 to $6,350/MWh depending on workload type. This is 5× to 127× higher than the gas heat rate ($25–$60/MWh). Whether or not these tolerance levels translate into wholesale price increases depends on market-specific supply and demand conditions, but the measurement itself is analytically important in every scenario: it quantifies, for the first time, the price elasticity gap between AI demand and all prior demand classes on the grid. The CHR Q1 2026 Index is published and live at computeheatrate.com/chr-index.html.
The demand signal is real and unprecedented. U.S. data centers consumed approximately 183 TWh in 2024, over 4% of national consumption (IEA, "Energy and AI," 2025). The IEA projects this will more than double to over 400 TWh by 2030. The five largest hyperscalers are projected to spend approximately $750 billion on capex in 2026 (CreditSights, February 2026), approximately 75% on AI infrastructure.
Markets are already repricing in some regions. PJM capacity auction prices surged from $28.92/MW-day (2024/25) to $333.44/MW-day (2027/28), hitting the FERC-approved cap for three consecutive auctions (PJM 2027/2028 BRA Results). Data center load accounted for 40% ($6.5B) of $16.4B in PJM’s December 2025 auction costs (Monitoring Analytics via Utility Dive, January 2026).
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.
The CHR effect is a penetration-threshold phenomenon. Where data center concentration is already high, the CHR dynamic is visible in market data. Where concentration is still low, the CHR provides a forward-looking measurement tool. In both cases, the metric serves an analytical function that no existing tool provides: quantifying demand-side price tolerance in $/MWh terms directly comparable to supply-side generation costs.
2. The 100× Thesis: Why This Revolution Is Different
In February 2026, Google DeepMind CEO Demis Hassabis described AGI as potentially delivering “ten times the impact of the Industrial Revolution, but happening at ten times the speed.” This framing has a profound and under-examined implication for energy markets. The Industrial Revolution transformed energy demand gradually across dozens of industries over more than a century. AI is compressing a comparable or larger transformation into a single decade.
Unlike the Industrial Revolution, which diffused its energy demands across textiles, steel, rail, chemicals, and agriculture, AI concentrates virtually its entire economic impact through one physical process: converting electricity into computation. The innermost loop.
The Single-Input Revolution
| Revolution | Primary Energy Inputs | Energy Concentration |
|---|---|---|
| First Industrial (1760–1840) | Coal, water, wood, animal labor | Highly diffuse across inputs and sectors |
| Second Industrial (1870–1920) | Coal, oil, electricity, natural gas | Diffuse; electricity one of several new inputs |
| Digital (1970–2010) | Electricity (modest share) | Moderate; total demand modest |
| AI Revolution (2024–) | Electricity (overwhelmingly dominant) | Extreme: >95% is electricity → compute |
Exhibit 2.0: Energy concentration by technological revolution.
3. The Compute Heat Rate: Definition & Measurement
The Compute Heat Rate (CHR) is defined as the maximum price per megawatt-hour of electricity at which a given AI workload remains economically viable. It is conceptually analogous to the gas heat rate but operates on the demand side of the market rather than the supply side.
Computed Tolerance Ceiling by Workload (Q1 2026)
Values are consistent with the SSRN paper (Royal, 2026) and the published CHR Q1 2026 Index.
| Workload Type | Rw ($/MWh) | Cnon-elec | CHR ($/MWh) | vs. Gas HR |
|---|---|---|---|---|
| Frontier Inference | $49,300 | $4,250 | $34,650 | ~690× |
| Mid-Tier Inference | $14,800 | $4,250 | $8,120 | ~162× |
| Enterprise Contracted | $9,800 | $4,250 | $4,270 | ~85× |
| Enterprise Agentic AI | $15,000 | $4,250 | $8,270 | ~165× |
| Commodity Inference | $1,850 | $4,250 | ~$800* | ~16× |
| Frontier Model Training | $2,000† | $4,250 | ~$500† | ~10× |
| Blended Average (2026) | $12,500 | $4,250 | $6,350 | ~127× |
*Cross-subsidized. †Amortized over model lifetime. 30% required margin; gas heat rate benchmark ~$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 tolerance ceiling 10–20× above the gas heat rate. This removes the historical demand-side brake and allows supply-side economics (CONE) to become the binding constraint on long-run pricing in supply-constrained markets.
4. CHR Trajectory: 2026–2035
Hardware & Algorithmic Efficiency Gains
NVIDIA Vera Rubin, unveiled at GTC 2026 (March 16–19, 2026), delivers up to 10× inference throughput per watt versus Blackwell and a claimed 10× reduction in cost per token. The Rubin GPU features 336 billion transistors, 288GB HBM4 per chip, and 50 petaflops of NVFP4 inference compute at 2.3 kW TDP. Paired with Groq 3 LPX, NVIDIA claims 35× inference throughput per megawatt. Jensen Huang raised the revenue outlook to $1 trillion in orders through 2027.
Jevons Paradox & Demand Elasticity
Jevons Paradox, first observed in 1865 regarding coal consumption, states that when technology makes a resource more efficient to use, total consumption increases because efficiency makes the resource useful for a wider range of applications. AI compute exhibits this with extraordinary clarity: per-token costs fell ~1,000× between 2022 and 2025, yet total electricity consumption grew far faster. The IEA projects global DC electricity doubling from 415 TWh (2024) to 945 TWh (2030).
Vera Rubin reinforces the pattern. The 10× efficiency improvement makes previously uneconomic workloads viable (trillion-parameter inference, agentic systems, real-time scientific simulation), expanding the total addressable compute market. Huang explicitly framed the economics around enabling new workload categories: “Finally, AI is able to do productive work, and therefore the inflection point of inference has arrived.”
For Jevons not to hold, demand elasticity must be <1.0. Current evidence suggests 2.0–4.0. At elasticity 2.5, a 10× efficiency gain expands demand 31.6×, for a net 3.16× increase in electricity consumption.
Two Trajectory Scenarios
The observed trend since 2022 shows CHR values rising over time, not falling. Each hardware generation produces more revenue per MWh faster than it reduces cost per MWh. However, a bearish scenario remains possible, so both are presented.
Scenario A: Bearish (Efficiency Outpaces Demand)
| Year | Efficiency | Demand Growth | Blended CHR |
|---|---|---|---|
| 2026 | 1.0× | 1.0× | $6,350 |
| 2028 | 3× | 2–3× | $4,000–$5,000 |
| 2030 | 10× | 8–12× | $2,500–$3,500 |
| 2035 | 60× | 40–80× | $600–$1,200 |
Scenario A: CHR declines but remains well above CONE ($80–$130/MWh) through 2035.
Scenario B: Bullish (Observed Trend Continues)
| Year | Efficiency | Demand Growth | Blended CHR |
|---|---|---|---|
| 2026 | 1.0× | 1.0× | $6,350 |
| 2028 | 3× | 8–15× | $6,000–$8,000 |
| 2030 | 10× | 30–60× | $7,000–$12,000 |
| 2035 | 60× | 300–800× | $10,000–$20,000+ |
Scenario B: Observed trend. Each generation unlocks new workload categories that consume more total electricity. Vera Rubin’s agentic positioning supports this trajectory.
Under both scenarios, the CHR remains far above CONE-based equilibrium prices ($80–$130/MWh) through 2035. The CHR serves its primary function regardless of which scenario prevails: measuring the demand-side tolerance ceiling for regulators, market designers, and procurement professionals.
5. AI Demand: Granular Assessment
| Development Stage | Est. Capacity (GW) | Conversion Rate | Realized (GW) |
|---|---|---|---|
| Operational | ~25–28 | ~100% | 25–28 |
| Under Construction | ~18–22 | ~85% | 15–19 |
| Financed/Contracted | ~12–18 | ~65% | 8–12 |
| Queue/Proposed | ~38–62+ | ~20–50% | 10–16 |
| Realized by 2030 | ~45–55 GW |
Conversion rates reflect historical PJM/ERCOT queue attrition. Even at a 50% haircut, AI demand growth is 6–10× historical rates.
6. When AI Demand Reshapes the Supply Stack
| ISO/RTO | 2024 DC (GW) | 2030E DC (GW) | 2024 DC % Peak | 2030E DC % Peak | Severity |
|---|---|---|---|---|---|
| PJM | ~8 | 15–25 | ~5% | ~10–15% | Critical |
| ERCOT | ~4 | 10–18 | ~3% | ~8–12% | Critical |
| CAISO | ~2 | 4–7 | ~2% | ~5–8% | Elevated |
| MISO | ~2 | 4–8 | ~1% | ~3–5% | Elevated |
| NYISO | ~1.5 | 2.5–4.5 | ~2% | ~5–7% | Elevated |
DC % of Peak shown for both 2024 (current) and 2030 (estimated).
The Installed Base Effect
The CHR effect is a penetration-threshold phenomenon, analogous to the duck curve, the gas heat rate emergence, and West Texas wind covariance pricing. With ~25–28 GW operational today, the effect is visible where concentration is highest. As the pipeline delivers 45–55 GW by 2030, it will propagate to every market with significant data center activity.
7. The CONE-to-CHR Spectrum
| Price Regime | Wholesale $/MWh | Market Condition |
|---|---|---|
| Below CONE | $30–$75 | AI demand fails; supply catches up |
| At CONE | $80–$130 | New supply barely keeps pace |
| Mild shortage | $100–$160 | Supply lags; baseload prices elevated |
| Severe shortage | $150–$250+ | Wide gap; scarcity hours expand |
| Supercycle | $200–$400+ | Extreme tightness; CHR-adjacent pricing |
The CHR does not predict which regime will prevail; it provides the framework for understanding demand-side behavior under each.
8. Market Implications
Impact on energy-intensive industry. Traditional industrial consumers cannot absorb $100–$160/MWh electricity. Their tolerance ceilings are 3–10× lower than AI workloads. The CHR quantifies this gap.
Hourly dynamics. The correct primary price display for baseload consumers is the 24/7 cost-to-serve, not a blended annual average that dilutes expensive peak/scarcity hours with near-zero renewable hours.
9. Counterarguments & Stress Tests
Any analytical framework must be tested against the strongest objections. In each case, we examine how the counterargument affects both the market outcome and the continued utility of the CHR as a metric.
9.1 Efficiency Gains Destroy Demand
If a sufficiently large efficiency breakthrough causes total AI electricity consumption to decline, the CHR becomes more important as a metric: it provides the measurement tool for tracking whether the demand-side brake is repairing. Historical evidence: per-token costs fell ~1,000× yet consumption grew faster. Vera Rubin’s positioning around new workload categories suggests the 10× gain will expand the addressable market. The CHR tracks the answer in real time.
9.2 Regulatory Intervention
The most likely counterargument to materially affect outcomes. The Suppression Paradox: price caps reduce incentives for new generation precisely when the grid is short. In traditionally regulated markets (MISO, Southeast), regulators can require data centers to finance their own generation, partially shielding existing ratepayers. This is a real offset in those specific markets. But regardless of regulatory path, the CHR measures the underlying demand-side dynamic that regulation is responding to.
9.3 Massive New Supply
SMRs: early-to-mid 2030s. Gas turbine backlogs through 2028–2030. PJM queue: 170K+ MW, 4–5 year timelines. Supply responds eventually, but the structural lag is 8–12+ years.
9.4 AI Revenue Commoditizes
Per-token prices down 90%+. Total revenue growing faster. OpenAI ARR tripled to ~$20B in 2025. Commodity pricing at the low end coexists with premium pricing at the frontier, keeping blended CHR elevated.
9.5 Demand Overstated
Even after a 30–40% haircut, 45–55 GW by 2030 exceeds supply additions. Hyperscaler capex (~$750B in 2026) is contractual. DC vacancy: 1.6% record low. 75% under construction is preleased.
9.6 Geographic Diversification & Off-Grid
Inference is latency-sensitive. Training is mobile but needs massive power. BTM provides zero grid capacity benefit. Mobility is a redistribution mechanism, not a dissipation mechanism.
10. Scenario Modeling
| Scenario | 2030 Wholesale | Conditions Required | CHR Utility |
|---|---|---|---|
| S1: No AI Growth | $30–$45 | Investment collapses; adoption reverses | Low |
| S2: Moderate | $55–$80 | AI grows slowly; supply responds | Moderate |
| S3: Base Case | $100–$160 | Pipeline delivers; supply lags | High |
| S4: Accelerated | $150–$250 | Pipeline exceeds forecasts; severe shortage | Critical |
| S5: Supercycle | $200–$400+ | AI demand + supply failure + policy lag | Essential |
Under every scenario except S1, the CHR serves an important analytical function.
11. Conclusions & Measurement Standard
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. Together, they provide the complete analytical framework for wholesale price formation in markets where AI load is material. See the companion SSRN paper (Royal, 2026) for the peer-reviewed formulation.
12. CHR vs. PUE: The Business Metric vs. the Building Metric
Every energy-intensive industry has a spread metric. Oil refining has the crack spread. Steel has the scrap-to-melt ratio. Aluminum has the LME-to-power ratio. These are the instruments by which billions of dollars of production decisions are made daily.
The data center industry has PUE (Power Usage Effectiveness): the ratio of total facility energy to IT equipment energy. PUE measures building efficiency. It does not tell you what the compute equipment earns per MWh consumed, how much electricity price a data center can absorb, or how AI workloads compare to aluminum smelters in their tolerance for price spikes. These are business questions, not engineering questions.
PUE is the building metric. CHR is the business metric. PUE measures how efficiently the building delivers power to the servers. CHR measures how much the servers can afford to pay for that power. The data center industry has had PUE for nearly two decades. It has never had a CHR.
The gas heat rate sets the floor. The CHR sets the hypothetical ceiling. Together, they define the boundaries of wholesale price formation in markets where AI load is material.
13. The SLA Lockdown: Why Some Data Centers Cannot Curtail
Two Locks on the Same Door
The CHR demonstrates that AI data centers will not curtail because the economics do not justify it (the economic lock). SLAs add a contractual lock: operators cannot curtail because 99.95–99.999% uptime guarantees carry penalties that dwarf any electricity savings.
| SLA Tier | Uptime | Max Downtime/Year | Penalty Structure |
|---|---|---|---|
| Standard | 99.9% | 8.76 hours | 10–25% service credit |
| High Availability | 99.95% | 4.38 hours | 25–50% service credit |
| Mission Critical | 99.99% | 52.6 minutes | 50–100% credit + damages |
| Ultra-High | 99.999% | 5.26 minutes | Full credit + consequential damages |
The Colocation Paradox
The thinnest-margin operators (Equinix, Digital Realty) are the most contractually locked. Multi-tenant leases include liquidated damages of $50–$200/kW/hr. The operator with the least economic cushion has the most rigid obligation to continue consuming.
The Unprecedented Intersection
AI data centers are the first demand class in U.S. electricity market history simultaneously large enough to move the supply stack (15–25+ GW) and non-curtailable at any price the market can produce. Every prior non-curtailable load was too small to matter. Every prior large load was curtailable enough to self-correct.
14. The Peaker Paradox
As data center penetration rises, gas peakers dispatch more hours per year. A peaker’s per-MWh cost does decline as capacity factor rises. But the first-order effect is the expanding number of hours during which peakers set the wholesale clearing price.
At 0% DC penetration: peakers set prices ~600 hours/year.
At 20% DC penetration: peakers set prices ~4,200 hours/year.
The declining per-unit cost (~35% reduction) is overwhelmed by the expanding dispatch envelope (~140–200% increase in average wholesale price). Heat rate penalty, maintenance escalation, and a cost floor of ~$52–$56/MWh ensure peakers never reach baseload costs.
“More capacity factor means cheaper power” is true for the peaker but false for the market. The declining unit cost is a second-order effect. The expanding price-setting hours are the first-order effect. The lines cross in opposite directions.
15. The Convergence Theorem
All data center market participation structures (direct QSE, REP-served, utility tariff, behind-the-meter) converge on the same CHR outcome. The speed at which wholesale price signals reach the operator differs. The magnitude does not.
The Regulated Market Nuance
In traditionally regulated markets (MISO, parts of SPP, Southeast), regulators can require data centers to pay for their own generation. This changes the distributional impact (existing ratepayers are partially shielded) but does not change the CHR outcome: the data center still consumes at CHR-level tolerance, and the generation built to serve it must still be financed within the same constrained supply chain.
The Mobility Paradox
Hyperscaler geographic flexibility strengthens the CHR thesis for immobile industrial clients. Mobility is a redistribution mechanism, not a dissipation mechanism. And inference workloads are latency-sensitive and must remain near users.
16. Comprehensive Q&A
The most common questions, clarifications, and challenges raised across academic review, peer review, and stress-testing since the CHR framework was first published in February 2026.
Q: If CHR is so high, why haven’t wholesale prices already repriced everywhere?
Because the CHR effect is a penetration-threshold phenomenon. It requires critical mass at a specific location. This is identical to how California’s duck curve did not exist before sufficient solar was installed. Where concentration is already high (PJM/Northern Virginia), prices have repriced: capacity prices up 833%, wholesale near DC clusters up 267%.
Q: Don’t SLAs make CHR irrelevant?
Five reasons CHR remains essential: (1) CHR explains why SLAs exist. (2) CHR governs behavior when SLAs break. (3) CHR explains the scale dimension. (4) CHR provides continuous measurability versus SLA’s binary yes/no. (5) CHR tracks trajectory over time.
Q: Won’t efficiency gains reduce AI energy demand?
Jevons Paradox. Per-token costs fell ~1,000×; consumption grew far faster. Vera Rubin (GTC 2026) delivers 10× throughput per watt but targets new workload categories, expanding total demand. The CHR tracks this in real time: if efficiency wins, CHR declines; if Jevons wins, CHR rises. The observed trend through Q1 2026 shows CHR rising.
Q: What about the AI bubble thesis?
25–28 GW of operational capacity does not disappear. ~$750B in 2026 capex is contractual. 75% of construction is preleased. The bubble thesis requires a reversal of enterprise AI adoption, which contradicts every observable trend.
Q: Why don’t forward curves already reflect this risk?
For further discussion.
Q: How is CHR different from PUE?
PUE is a building efficiency metric. CHR is a business economics metric. PUE tells you 20% of facility electricity is wasted on cooling. CHR tells you the facility can profitably consume electricity at 127× the wholesale average.
Q: Can data centers just go off-grid?
BTM generation is an escape valve, not a systemic bypass. Most hyperscaler load is going on-grid. BTM worsens the problem for everyone else on the grid.
Q: Won’t solar and wind overbuild crush prices?
Solar cannibalization intensifies hourly bifurcation, not resolves it. New gas entry needs 4,000–6,000 hours to recover CONE. Net effect: prices bifurcate, not collapse. Data center 24/7 flat load profiles consume through every hour, including the expensive ones.
Q: Won’t hyperscalers just move to cheaper regions?
The Mobility Paradox: hyperscaler mobility strengthens the CHR thesis for immobile clients. Mobility redistributes demand, it does not dissipate it. Inference is latency-sensitive.
Q: Is CHR patented?
No. CHR is an original economic theory. The appropriate protection is authorship, not patent prosecution. The SSRN working paper (Royal, 2026; Abstract ID 6322318) establishes scientific priority.
Q: What about regulatory intervention?
Intervention is certain; form and efficacy are not. The Suppression Paradox: price caps worsen the underlying problem. In regulated markets, some jurisdictions require DCs to finance own generation, partially shielding ratepayers. The repricing window remains open under all plausible paths.
Q: What about data centers in space?
The Orbital CHR Ceiling Principle: the effective electricity cost of orbital compute establishes a theoretical upper bound on terrestrial pricing. In practice, orbital costs are extremely high, reinforcing that terrestrial CHR dynamics have enormous room to run.
Q: Where can I find the academic paper and the published CHR Index?
SSRN working paper: Abstract ID 6322318. CHR Q1 2026 Index: computeheatrate.com/chr-index.html. Updated quarterly.
17. Sources & References
- Royal, H. (2026). “The Compute Heat Rate.” SSRN Abstract ID: 6322318. Link
- IEA (2025). “Energy and AI.” April 2025. Link
- CreditSights (2026). “Raising Hyperscaler Capex 2026 Estimates.” February 2026. Link
- PJM Interconnection (2025). 2027/2028 Base Residual Auction Results. December 17, 2025. Link
- Monitoring Analytics (2026). Data center impact on PJM capacity auction costs. Via Utility Dive, January 7, 2026. Link
- NVIDIA (2026). Vera Rubin platform announcement, GTC 2026. Link
- CNBC (2026). “Nvidia GTC 2026: CEO Jensen Huang sees $1 trillion in orders.” March 16, 2026. Link
- Epoch AI (2024–2025). Inference price and training cost research. Link
- Goldman Sachs (2025). “AI to Drive 165% Increase in DC Power Demand.” February 2025.
- Wood Mackenzie (2025). US Power Market Outlooks. October & December 2025.
- LevelTen Energy (2025). PPA Price Index, 2025.
- Lazard (2025). LCOE+ 2025.
- BNEF (2025). IRA rollback impact analysis.
- IEEFA (2025). “Projected Data Center Growth Spurs PJM Capacity Prices.” Link
- Amazon, Alphabet, Microsoft, Meta, Oracle earnings through Q4 2025 / Q1 2026.
- Constellation, Vistra, Talen SEC filings post-PJM December 2025 auction.
- OpenAI, Anthropic, Google published API pricing (Q1 2026).
Suggested Citation
Available at: https://computeheatrate.com/research.html