reads

recycled AI-generated articles

published by xavier.grehant on 2026-05-15

The Power Ceiling: AI's Electricity Bottleneck, Grid Constraints, and the Economics of Dedicated Silicon

Research Paper | May 2026

Abstract

Artificial intelligence is triggering the first sustained surge in US electricity demand in nearly two decades — and the grid is not ready for it. AI data center electricity consumption is compounding at 12–22% per year, while transmission infrastructure grows at roughly 1–3% annually and high-voltage line construction hit near-historic lows in 2024. The resulting supply-demand imbalance is already visible in capacity market prices that have risen tenfold in the PJM region, in power approval timelines stretching to 24–36 months for new facilities, and in residential electricity bills climbing well above inflation across data center clusters. This paper analyzes the scale and pace of AI electricity demand growth, the structural limits of the US grid, the cost consequences for both AI operators and ordinary consumers, and the role of dedicated silicon in modifying the equation. The central finding is that hardware efficiency gains — real and dramatic at the chip level — are unlikely to reduce aggregate electricity consumption because lower cost per inference reliably generates more inference: a dynamic economists recognize as Jevons' Paradox. The paper closes by examining the most credible structural responses, including the nuclear power purchase agreements now being signed at multi-gigawatt scale by the major hyperscalers, and by identifying the second-order questions that current evidence cannot yet resolve.

1. The Demand Side: How Fast Is AI Eating Electricity?

1.1 From Flat to Explosive

For more than a decade, US electricity consumption was broadly flat. Efficiency gains from LED lighting, better appliances, and industrial process improvements absorbed most of the growth from population and GDP expansion. That era is now over.

The IEA's landmark Energy and AI report (2025) establishes that global data center electricity consumption stood at approximately 415 TWh in 2024, or about 1.5% of global electricity demand, after growing at roughly 12% per year since 2017 — more than four times faster than total electricity consumption growth across all other sectors combined.[^1]

The US sits at the center of this surge. The IEA estimates that in 2024, the United States accounted for 45% of global data center electricity consumption, with China at 25% and Europe at 15%.[^2] The Lawrence Berkeley National Laboratory projects that US data center demand will grow from 176 TWh in 2023 (4.4% of US electricity) to between 325 and 580 TWh — i.e., 6.7% to 12% of total US consumption — by 2028.[^3]

S&P Global's 451 Research found that US data center grid power demand rose 22% in 2025 alone, reaching approximately 61.8 GW, and forecasts that it will nearly triple by 2030.[^4] Virginia's data center demand alone hit approximately 12.1 GW in 2025, up from 9.3 GW in 2024.[^5]

1.2 Why AI Workloads Are Categorically Different

This is not simply more of the same compute demand. The step-change is architectural. AI training and inference workloads run on GPU racks whose power density is up to six times higher than conventional server racks.[^6] Training GPT-3 consumed approximately 1.29 GWh; training GPT-4 is estimated to have consumed over 50 GWh — a roughly 40-fold increase in a single generation.[^7] AI-focused data centers can draw as much electricity as power-intensive factories such as aluminum smelters, but they are far more geographically concentrated: nearly half of US data center capacity is in just five regional clusters.[^8]

The IEA projects that global data center electricity consumption will approximately double by 2030, reaching around 945 TWh in the base case, or close to 3% of total global electricity demand.[^9] A more aggressive scenario projects 1,200 TWh by 2035.[^10] Meanwhile, a January 2026 Bloom Energy report forecasts that US data centers' combined energy demand will nearly double from 80 to 150 GW between 2025 and 2028 alone — equivalent to adding Spain's entire electricity consumption in three years.[^11]

Key takeaway: AI electricity demand has been compounding at 12–22% per year, while the grid has historically expanded by roughly 1–3% per year and is currently struggling to sustain even that pace for bulk supply. Chip production scales exponentially; electricity infrastructure does not.

2. The Supply Side: Can the Grid Keep Up?

2.1 The Physical Constraints

The US electricity grid comprises approximately 600,000 miles of transmission lines (240,000 of which are high-voltage) and over 5.5 million miles of local distribution lines.[^12] Its capacity to absorb massive new loads is constrained by three distinct bottlenecks:

Transmission. In 2024, only 888 miles of high-voltage transmission (≥345 kV) were constructed — a historically low figure. By comparison, approximately 4,000 miles were built in 2013 alone.[^13] Interconnection queues — the backlog of generation and load projects waiting to connect to the grid — have grown to extraordinary lengths, often stretching 3–5 years even for projects with committed financing.

Transformers. The US faces an acute shortage of large distribution transformers due to supply chain bottlenecks and limited domestic manufacturing. Lead times for new transformers stretched to 80 to 210 weeks as of June 2024 (120 weeks on average), up from 50 weeks in 2021, and transformer costs have risen 60–80% since 2020.[^14] The ASCE estimates that distribution transformer capacity will need to rise 160–250% by 2050 to meet anticipated needs — with the current production infrastructure badly unprepared to deliver that.

Permitting and planning cycles. Consultancy Grid Strategies found that utility five-year load-growth forecasts jumped nearly fivefold from 2022 to late 2024, reaching 128 GW of anticipated new demand — a pace that would require an unprecedented 16% increase in the country's capacity to generate and deliver electricity by 2029.[^15] The US regulatory and permitting apparatus, historically calibrated for slow, steady growth, has no precedent for this rate of change.

2.2 Regional Stress Hotspots

The problem is not uniform. The IEA's study using a Power Stress Index (PSI) shows that regions like Oregon, Virginia, and Ireland may experience PSI values exceeding 0.25, indicating severe local grid vulnerability, while more diversified systems like Texas and Japan have greater absorption capacity.[^16]

Northern Virginia — "Data Center Alley" in Loudoun County — already hosts 25 million square feet of data centers consuming a quarter of the state's energy.[^17] In July 2024, a single voltage fluctuation in northern Virginia simultaneously disconnected 60 data centers, triggering a 1,500 MW power surplus that required emergency grid adjustments to prevent cascading outages.[^18]

The consequences are already showing up in wholesale markets. In the PJM regional electricity market (spanning Illinois to North Carolina), capacity market clearing prices for the 2026–2027 delivery year rose to $329.17/MW — over ten times the price of $28.92/MW for the 2024–2025 delivery year, with rapid data center growth identified as a primary driver.[^19] Similar pressures are emerging in NYISO and ERCOT.

2.3 An Important Distinction: Generation vs. Transmission

It is important to note that the grid supply constraint has two distinct dimensions that are often conflated. Generation capacity is actually growing fast — the US added 62.8 GW in 2024, led by solar (which accounted for 58% of new capacity), with battery storage nearly doubling to 30 GW.[^20] The more acute bottleneck is transmission and distribution: the power can be generated, but getting it to where the data centers are, at the reliability and density required, is the binding constraint. This is why AI operators are increasingly building behind-the-meter power generation (gas generators, onsite solar, and soon SMRs) rather than depending on the public grid.

3. AI Infrastructure Costs: Are They Rising?

3.1 Power, Not Chips, Is Now the Scarce Resource

For most of 2022–2024, the dominant constraint for AI operators was GPU availability. H100 chip shortages meant months-long procurement queues, and the economics of AI inference were modeled around GPU $/hr. That has changed decisively. As of 2025, data center power capacity is the more pressing constraint for new AI infrastructure deployments. Major markets including Northern Virginia, Silicon Valley, and Northern Europe have seen power approval timelines stretch to 24–36 months for new facilities, regardless of hardware availability.[^21]

Power purchase agreement (PPA) prices rose by an average of 35% in 2024, driven almost entirely by hyperscalers' surge in procurement. Big Tech companies — Google, Meta, Amazon, Microsoft — accounted for 43% of all global clean energy PPAs signed in 2024.[^22]

The four hyperscalers collectively spent over $200 billion in capital expenditure in 2024, representing a 62% year-over-year increase, with a substantial portion going to securing power infrastructure rather than compute hardware alone.[^23]

3.2 The Cost Structure of AI Tokens

Electricity is now a first-order cost item in AI inference economics. An H100 GPU draws approximately 700W and delivers roughly 35 tokens per second per watt at current software optimization levels.[^24] At $0.12/kWh — a competitive commercial electricity rate — running a 1,000-GPU H100 cluster for a month costs approximately $72,000 in electricity alone, not including cooling overhead (Power Usage Effectiveness or PUE), which typically adds 15–45% on top of the IT load.

In markets where the power constraint is binding, the effective cost per token is rising. This is not uniformly visible in consumer AI pricing yet, because hyperscalers have been absorbing cost through scale economies and competitive pricing pressure, but it is compressing margins and influencing where inference workloads are routed geographically.

3.3 Will AI Get More Expensive?

The directional answer is yes — but with important caveats around time horizon and market structure.

In the short term (2025–2028), electricity supply constraints are real and worsening in key markets. The supply-demand imbalance in regional grids is already manifesting as higher wholesale prices, higher capacity market prices, and longer interconnection queues. AI operators with existing long-term power contracts are insulated; new entrants face structurally higher costs.

In the medium term (2028–2032), the picture is more contested. New generation — solar, gas peakers, and potentially restarted or new nuclear — will come online and partially relieve the constraint. But given the 3–7 year lead time for major generation and transmission projects, the overhang will persist throughout this window.

In the long term (post-2032), the constraint could ease materially if nuclear SMRs deploy at scale, transmission reform proceeds, and algorithmic/hardware efficiency gains compound. But this is not a given.

4. Spillover: What Happens to Everyone Else's Electricity Bill?

4.1 The Burden Is Already Being Passed to Consumers

This is perhaps the most politically significant dimension of the problem — and one that is already generating measurable economic and electoral consequences.

US average residential electricity prices rose by approximately 25% between 2020 and 2024, reaching roughly 19 cents/kWh by end-2025 — up from approximately 13 cents/kWh, a level that had been stable for over a decade.[^25] In 2025 alone, residential electricity prices rose 11.5%, more than double the headline inflation rate of 2.9% that year.[^26]

Goldman Sachs published a research note in February 2026 stating that electricity prices jumped 6.9% in 2025 year-over-year, and that data centers account for 40% of incremental US electricity demand growth through the end of the decade.[^27] Goldman warned that rising electricity bills will lower household disposable income, drag consumer spending, and marginally slow GDP growth.

A Carnegie Mellon University study estimates that data centers and cryptocurrency mining could lead to an 8% increase in the average US electricity bill by 2030, potentially exceeding 25% in highest-demand markets such as northern and central Virginia.[^28]

The distribution is deeply uneven:

  • Northern Virginia: In the PJM market, data centers accounted for an estimated $9.3 billion price increase in the 2025–26 capacity market, translating to $18/month extra for western Maryland residential customers and $16/month in Ohio.[^29]
  • Virginia rate cases: Dominion Energy filed its first base-rate increase since 1992 in February 2025, adding approximately $8.51/month per typical household starting in 2026.[^30]
  • Wholesale spot markets: Wholesale electricity costs as much as 267% more than five years ago in areas near major data center clusters. More than 70% of nodes recording price increases are within 50 miles of significant data center activity.[^31]

Looking ahead, the US Energy Information Administration forecasts electricity prices could increase by up to 40% by 2030 compared to 2025, with the most aggressive third-party estimates (Jack Kemp Foundation) projecting 70% by 2029 in affected markets, potentially adding over $1,200/year to average household bills.[^32]

4.2 A Structural Inequity in Who Pays

A disturbing asymmetry has emerged in the price distribution: residential consumers are bearing most of the cost increase while large commercial and industrial users — including data centers — are paying relatively stable or even lower rates. A January 2026 Yale Climate Connections analysis found that data centers are consuming more electricity than ever but their average price has risen only marginally, while residential prices rose 25% between 2020 and 2024.[^33]

This reflects the structure of electricity regulation: large industrial customers typically negotiate long-term contracts at wholesale rates with favorable terms, while regulatory cost allocation frequently assigns grid infrastructure upgrade costs to the broader residential and small-commercial ratepayer base. There is a growing political movement — including governor's races in New Jersey and Virginia in 2025 — built partly around the question of who should bear the cost of grid upgrades driven by hyperscaler demand.

5. The Dedicated Chip Question: Does Silicon Efficiency Change the Equation?

5.1 The Efficiency Gains Are Real and Dramatic

The dedicated chip question is the most technically sophisticated part of the analysis, and the answer is genuinely impressive at the chip level.

NVIDIA claims that across six architecture generations, inference throughput per megawatt has improved by 1,000,000x (one million times).[^34] More modestly but concretely: NVIDIA's latest data shows that Blackwell Ultra GB300 NVL72 systems deliver up to 50x higher throughput per megawatt and 35x lower token cost than the Hopper generation (H100) for DeepSeek-R1 inference.[^35]

Google's TPU v6 (Trillium) has a 300W thermal design power versus 700W for the H100 and 1,000W for the B200 — a 2.3–3.3x power advantage at the chip level for comparable inference throughput.[^36] At scale across 100,000 chips, that power difference represents the annual electricity consumption of a small country.

Purpose-built ASICs (Application-Specific Integrated Circuits) go further. Google's TPUs, Amazon's Trainium/Inferentia, Groq's LPUs, and Cerebras' wafer-scale engines are each optimized for specific AI workload profiles — most commonly, the transformer architecture that underpins virtually all modern large language models. They achieve efficiency gains by eliminating the general-purpose logic that GPUs carry for non-AI workloads, dramatically reducing memory bandwidth pressure (the primary bottleneck in transformer inference), and tightly co-designing the memory, interconnect, and compute fabric for specific data flows.

Google's TPU v6e is available at committed-use discounts as low as $0.39 per chip-hour, cheaper than spot H100s in most regions once egress and interconnect costs are included.[^37]

5.2 Jevons' Paradox: Why Efficiency Doesn't Reduce Total Consumption

Here lies the central structural contradiction, and it is worth stating precisely.

Jevons' Paradox (named for 19th-century economist William Stanley Jevons, who observed that more efficient steam engines led to more coal consumption, not less) holds that when a resource is used more efficiently, the reduction in cost per unit drives so much additional demand that total resource consumption increases.[^38]

In the context of AI and electricity, the mechanism operates at multiple levels:

  1. Direct demand effect: Lower inference cost per token makes AI economically viable for use cases that were previously too expensive — autonomous agents, video generation, real-time translation, always-on AI assistants, AI-in-every-application. Each new use case adds absolute load to the grid, even if each token is far cheaper.

  2. Market expansion effect: Cheaper AI unlocks new markets. Startups, schools, hospitals, and small businesses that could not previously afford AI-powered tools become buyers. The total addressable market expands faster than per-unit efficiency reduces per-unit demand.

  3. Infrastructure investment validation: Paradoxically, efficiency breakthroughs validate continued hyperscaler investment. When DeepSeek demonstrated dramatically lower training costs in early 2025, the market initially reacted as though AI infrastructure spending would slow. Within days, Satya Nadella wrote on X: "Jevons' paradox strikes again! As AI gets more efficient and accessible, we will see its use skyrocket." Within weeks, Meta raised its 2025 AI spending to $60–65 billion (+50% year-over-year).[^39]

  4. Capability expansion effect: More efficient inference per dollar makes it economically rational to run larger models, run more inferences, and build more ambitious applications. The capability frontier expands, not just the deployment of existing capabilities.

Academic work directly examining Jevons' Paradox in AI (Luccioni et al., arXiv 2501.16548, 2025) confirms that "efficiency gains may paradoxically spur increased consumption" and that "cost savings achieved by more efficient AI hardware can spur increased demand for new AI functionalities, which in turn drive further hardware upgrades."[^40]

The historical record in cloud computing strongly supports this mechanism. AWS dramatically reduced compute costs per unit throughout the 2010s; global cloud electricity consumption rose several hundred percent over the same period.

Key takeaway on chips: Dedicated silicon makes individual AI inference dramatically cheaper and more power-efficient. But the evidence strongly suggests this will expand AI adoption faster than it reduces absolute electricity consumption. Chips change the efficiency of AI, not the aggregate demand trajectory. They are a necessary but insufficient condition for resolving the grid constraint.

5.3 Where Chips Do Help: Edge Inference and Load Redistribution

There is one important dimension where chip efficiency genuinely moderates the grid constraint rather than merely deferring it: edge inference. As frontier AI performance that once required data-center-scale infrastructure becomes accessible on consumer hardware within approximately one year of initial release, a meaningful fraction of inference workload migrates to laptops, phones, and on-premise enterprise servers.[^41]

This matters because edge inference draws on distributed power infrastructure — laptop batteries, building circuits, cellular base stations — rather than concentrating massive loads in regional grid hotspots. A query answered locally by a Qualcomm Snapdragon or Apple M-series chip does not appear in northern Virginia's grid demand. Research in the Intelligence per Watt paper (arXiv, 2025) shows that local LM coverage on single-turn chat and reasoning queries rose from 23.2% in 2023 to 48.7% in 2024 to 71.3% by August 2025 — indicating rapid migration of routine inference to edge devices.[^42]

This trend does not eliminate the grid problem — training, complex reasoning, agentic multi-step tasks, and real-time video/audio generation will remain centralized for years — but it meaningfully moderates the growth trajectory for a significant share of routine inference.

6. Supply-Side Responses: How the Market Is Adapting

6.1 The Nuclear Renaissance

The most dramatic structural response to the power bottleneck is the renewed interest — and concrete investment — in nuclear energy. Big Tech companies signed contracts for more than 10 GW of new or restored US nuclear capacity in the 12 months preceding December 2025 alone:[^43]

  • Microsoft signed a 20-year PPA with Constellation Energy to restart Three Mile Island Unit 1 (835 MW) by 2028, in a deal valued at approximately $16 billion.[^44]
  • Google ordered up to 500 MW of small modular reactors (SMRs) from Kairos Power (expected 2030+), and in May 2025 committed early-stage capital for three additional reactor sites totaling 1.8 GW with Elementl Power.[^45]
  • Amazon expanded its nuclear offtake agreement with Talen Energy to 1,920 MW through 2042 from the Susquehanna plant in Pennsylvania, and backed 5 GW of X-energy SMR projects.[^46]
  • Meta signed a 20-year PPA with Constellation for the 1.1 GW Clinton Clean Energy Center in Illinois (previously scheduled for retirement in 2027), plus issued an RFP for 1–4 GW of new nuclear capacity.[^47]

Why nuclear? Four reasons: (1) 24/7 baseload supply — unlike solar and wind, nuclear runs continuously, matching the constant-load profile of AI data centers perfectly; (2) carbon-free — necessary for corporate sustainability commitments and increasingly for regulatory compliance; (3) long-term price certainty — 20-year PPAs lock in predictable rates, insulating operators from electricity price volatility; (4) colocation opportunity — placing a reactor adjacent to a data center campus (as Amazon is doing at Susquehanna) eliminates transmission bottlenecks entirely.

The SMR value proposition is particularly relevant: SMRs reduce capital cost per MW by approximately 50%, shorten construction time from 10+ years to approximately 3 years in target scenarios, and can be scaled modularly to match data center expansion. President Trump signed four executive orders in May 2025 to accelerate SMR deployment and ease NRC licensing.[^48] The first commercial SMR-powered data centers are expected by 2029–2030.

6.2 Other Responses

  • Direct gas generation: Companies unable to wait for grid connections are installing natural gas reciprocating generators on-site — inefficient and polluting, but available. Microsoft reportedly has deployed significant behind-the-meter gas generation at several campuses.
  • Demand response and load shifting: Some data center operators are negotiating interruptible service contracts — agreeing to curtail load during grid stress events in exchange for lower rates. AI inference (but not training) is partially amenable to this.
  • Geographic arbitrage: New data center investment is routing toward markets with surplus grid capacity — Texas (ERCOT, deregulated, large solar/wind surplus), Iowa, Ohio, and internationally to Scandinavia and Canada — reducing pressure on constrained markets but not reducing aggregate demand.
  • Renewable PPAs: Big Tech companies collectively accounted for 43% of all global clean energy PPAs in 2024.[^49] However, renewable energy — absent storage — does not provide the 24/7 reliability AI data centers require.

7. Synthesis

7.1 Is electricity the binding constraint on AI growth?

In specific markets and at specific time horizons, yes. Nationally, generation capacity is growing faster than the overall grid's transmission and distribution infrastructure can absorb. The binding constraint is not generation per se but transmission, transformer supply, and permitting timelines. In the top 5–10 US data center markets (northern Virginia, Silicon Valley, Phoenix, Columbus, Dallas), the constraint is already binding: power approval timelines stretch 24–36 months and are gating data center construction timelines.

7.2 Will AI infrastructure costs increase as a result?

Directionally yes, but the effect is uneven. Hyperscalers with existing long-term power contracts and behind-the-meter generation are insulated. New entrants and second-tier operators face materially higher power costs. The cost pressure is more visible in capacity market prices and PPA rates than in consumer-facing AI product pricing, which remains competitive due to market structure (intense hyperscaler competition). But margin compression is real and will eventually manifest in pricing or capital allocation decisions.

7.3 Will rising AI electricity demand push up prices for all consumers?

It already is. Average US residential electricity prices rose approximately 25% from 2020 to 2024, and 11.5% in 2025 alone — well above inflation. Goldman Sachs identifies data centers as the driver of 40% of incremental US demand growth. The impact is geographically concentrated near data center clusters but spreading as utilities recover grid upgrade costs across broader ratepayer bases. Forecasts suggest cumulative increases of 40–70% in the highest-demand markets by 2029–2030. This is no longer a speculative risk; it is a documented, ongoing transfer of cost from AI operators to households.

7.4 Will dedicated chips reverse the grid pressure?

Chips will make AI cheaper per token, but almost certainly not cheaper in aggregate energy terms. The efficiency gains are genuine and historically large — NVIDIA claims a 1,000,000x improvement over six generations; the Blackwell-to-Hopper leap alone represents 35–50x better economics per token. But Jevons' Paradox appears to be strongly operative: lower cost per inference has consistently led to more inference, not less total compute. The market's reaction to every efficiency breakthrough (DeepSeek, Blackwell, TPUv6) has been to increase AI spending, not reduce it. Edge inference migration to consumer devices is the most meaningful exception to this pattern, but it is a moderating influence rather than a structural reversal.

8. Questions This Analysis Raises

The foregoing analysis generates a set of second-order questions that deserve independent investigation:

  1. Who should pay for grid upgrades? Should hyperscalers bear the full cost of transmission and distribution upgrades their demand requires, or should these be socialized across ratepayers? This is increasingly a regulatory and legislative battleground.

  2. Will SMRs actually deploy on schedule? The nuclear thesis depends on construction timelines of 3 years that no SMR vendor has yet demonstrated in practice. If they slip to 7–10 years (the historical norm for new reactor designs), the 2030s relief does not materialize.

  3. Is there an AI equivalent of the DRAM glut? Semiconductor history includes cycles where massive capacity investment creates supply surpluses and price collapses. Could AI compute — and by extension, AI electricity demand — follow a similar bust cycle if AI adoption growth disappoints relative to infrastructure buildout?

  4. Can demand response work at AI data center scale? If AI inference workloads can be made genuinely interruptible (tolerating latency increases during grid stress), they could function as a massive demand response resource — effectively a 50–100 GW buffer for grid operators. Is this technically and commercially achievable?

  5. What is the regulatory risk? Proposals to require data centers to contribute to grid infrastructure costs, or to restrict data center siting in constrained grid regions, are gaining traction in state legislatures. How much regulatory risk is priced into hyperscaler capital allocation?

  6. Does distributed/federated AI inference change the calculus? If open-weight models continue closing the capability gap with closed frontier models (local LM coverage hit 71% in August 2025), and consumer hardware continues improving, a significant fraction of AI inference could be permanently redistributed to the edge — removing it from the grid stress equation entirely.

9. Conclusion

The electricity grid is becoming AI's most important physical constraint — more binding, in the near term, than chip supply. The scale of the mismatch is not trivial: AI data center electricity demand is compounding at 12–22% annually while grid infrastructure grows at 1–3% annually, with transmission construction near historic lows and transformer supply chains under acute stress.

The effects are already observable: capacity market prices in PJM have risen tenfold, residential electricity bills are rising well above inflation in affected regions, and data center construction timelines are increasingly gated by power availability rather than hardware delivery or capital. The grid constraint does not immediately threaten AI's growth — hyperscalers are resourceful, and the combination of behind-the-meter generation, nuclear PPAs, and geographic diversification will extend the runway — but it does impose real and growing costs.

Dedicated chips are the most powerful technical lever available to AI operators, and the efficiency gains they deliver are genuine. But history — from coal to automobiles to cloud computing — consistently shows that efficiency gains in transformative technologies drive adoption expansions that dwarf the efficiency savings. The IEA's base case projects AI data center electricity consumption at nearly 3% of global demand by 2030 despite continuous hardware improvement. Jevons' Paradox is not an academic curiosity; it is the most probable outcome of the current trajectory.

The resolution — partial and time-lagged — will come from a combination of SMR deployment, transmission reform, regulatory rebalancing of cost allocation, and edge inference migration. But through 2028 at minimum, the electricity grid remains a genuine bottleneck, AI infrastructure costs will reflect that bottleneck, and ordinary electricity consumers will continue paying a significant portion of the bill.

References

[^1]: IEA, Energy and AI, 2025. https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai [^2]: IEA, Energy and AI: Executive Summary, 2025. https://www.iea.org/reports/energy-and-ai/executive-summary [^3]: Lawrence Berkeley National Laboratory (cited in Belfer Center, 2026). https://www.belfercenter.org/research-analysis/ai-data-centers-us-electric-grid [^4]: S&P Global / 451 Research, October 2025. https://www.spglobal.com/energy/en/news-research/latest-news/electric-power/101425-data-center-grid-power-demand-to-rise-22-in-2025-nearly-triple-by-2030 [^5]: Ibid. [^6]: IEA, Energy and AI, 2025; Arxiv 2604.06198. [^7]: Patterson et al. (GPT-3); Cohen (GPT-4), cited in arXiv:2509.07218v4. [^8]: IEA, Energy and AI: Executive Summary, 2025. [^9]: IEA, Energy and AI, 2025. https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai [^10]: Brookings Institution, Global energy demands within the AI regulatory landscape, 2026. https://www.brookings.edu/articles/global-energy-demands-within-the-ai-regulatory-landscape/ [^11]: Bloom Energy / Consumer Reports, January 2026. https://www.consumerreports.org/data-centers/ai-data-centers-impact-on-electric-bills-water-and-more-a1040338678/ [^12]: ASCE Infrastructure Report Card 2025. https://infrastructurereportcard.org/cat-item/energy-infrastructure/ [^13]: Americans for a Clean Energy Grid / Grid Strategies, July 2025. https://cleanenergygrid.org/new-report-reveals-u-s-transmission-buildout-lagging-far-behind-national-needs/ [^14]: ASCE Infrastructure Report Card 2025. [^15]: Grid Strategies / Canary Media, December 2024. https://www.canarymedia.com/articles/transmission/lots-of-demand-too-little-grid-the-us-power-sector-in-2024 [^16]: arXiv:2604.06198 (AI–energy coupling framework, 2025). [^17]: arXiv:2501.16548v1, Luccioni et al., 2025. [^18]: Belfer Center, AI, Data Centers, and the U.S. Electric Grid, February 2026. [^19]: arXiv:2509.07218v4; PJM capacity auction data. [^20]: US EIA, Preliminary Monthly Electric Generator Inventory, 2024. [^21]: Spheron Network, AI Inference Power Consumption and GPU Electricity Costs: 2026 Guide. [^22]: Brookings Institution, 2026. [^23]: Belfer Center, 2026. [^24]: ScienceDirect / MLPerf Inference v4.1 data, 2025. [^25]: EESI, Data Center Power Demands Are Contributing to Higher Energy Bills, February 2026. https://www.eesi.org/articles/view/data-center-power-demands-are-contributing-to-higher-energy-bills [^26]: Ibid. [^27]: Goldman Sachs research note, cited in CNBC, February 2026. https://www.cnbc.com/2026/02/12/electricity-price-data-center-ai-inflation-goldman.html [^28]: Carnegie Mellon University study, cited in Pew Research Center, October 2025. https://www.pewresearch.org/short-reads/2025/10/24/what-we-know-about-energy-use-at-us-data-centers-amid-the-ai-boom/ [^29]: Pew Research Center, October 2025. [^30]: Belfer Center, 2026; Dominion Energy rate case filing. [^31]: Bloomberg, How AI Data Centers Are Sending Your Power Bill Soaring, September 2025. [^32]: EIA forecast cited in EESI, 2026; Jack Kemp Foundation / SSRN, November 2024. [^33]: Yale Climate Connections, January 2026. https://yaleclimateconnections.org/2026/01/home-electricity-bills-are-skyrocketing-for-data-centers-not-so-much/ [^34]: NVIDIA Technical Blog, Scaling Token Factory Revenue, April 2026. https://developer.nvidia.com/blog/scaling-token-factory-revenue-and-ai-efficiency-by-maximizing-performance-per-watt/ [^35]: Ibid. [^36]: ainewshub.org, Nvidia to Google TPU Migration 2025, November 2025. [^37]: Ibid. [^38]: Jevons, W.S., The Coal Question, 1865; Wikipedia, Jevons Paradox. [^39]: aiproem.substack.com, The Jevons Paradox in AI Infrastructure, February 2025. [^40]: Luccioni et al., arXiv:2501.16548, January 2025. [^41]: arXiv:2511.07885, Intelligence per Watt, 2025. [^42]: Ibid. [^43]: Introl.com, Nuclear power for AI, January 2026. [^44]: Ibid; Microsoft / Constellation Energy announcement, September 2024. [^45]: Introl.com, May 2025; nuclearbusiness-platform.com, June 2025. [^46]: Talen Energy Form 8-K, SEC filing, June 2025; Amazon / X-energy, October 2024. [^47]: Meta / Constellation Energy PPA announcement, June 2025. [^48]: Introl.com SMR implementation analysis, March 2026. [^49]: Brookings Institution, 2026.