The Compute Ledger Vol. 1 · Iss. 1 · April 2026
Research Report · Environment & Political Economy

The Compute
Buildout and
Its Costs

A sober ledger of AI's ecological footprint — measured against the justifications being used to rationalize it, and the physical lock-in that neither efficiency nor AGI is on track to undo.

Executive Summary

AI infrastructure is now materially reshaping the world's electricity, water, and mineral systems — and the dominant justifications for those costs are weaker than the industry's rhetoric.

Global data-center electricity use reached roughly 415 TWh in 2024 — about 1.5% of world consumption — and on the International Energy Agency's base case will more than double to ~945 TWh by 2030, slightly exceeding Japan's total electricity use today. In the United States, Lawrence Berkeley National Laboratory projects data centers will consume 6.7–12% of national electricity by 2028, up from 4.4% in 2023.

This buildout is, in the near term, being powered primarily by new natural gas and delayed coal retirements — not renewables or nuclear — with roughly 40 GW of behind-the-meter gas announced for AI sites in 2025 alone, and at least 15 US coal plants having had their retirements pushed back since January 2025. Hyperscaler greenhouse-gas emissions have risen 23–51% since their 2019–2020 baselines even as their climate pledges remain on paper, and independent audits judge those pledges to have "lost their meaning."

Meanwhile, the case that AI itself will repay these costs — through climate solutions, scientific discovery, productivity gains, or eventual AGI — rests on genuine but narrow scientific wins, speculative macroeconomics, and an AGI-solves-it faith not supported by the 2024–2026 evidence. This report assembles what is measured, what is projected, and what remains advertisement.

The Ledger — at a glance
2.3×
Global data-center electricity by 2030 vs. 2024
(415 → ~945 TWh)
IEA, 2025
+51%
Google's GHG emissions above 2019 baseline
Google Environmental Report 2025
+23%
Microsoft's emissions above 2020 baseline
Microsoft Sustainability 2025
+162%
Amazon's Scope 1 since 2019 Climate Pledge
Stand.earth / Amazon 2024
1,050%
PJM capacity-auction price rise, 2024/25 → 2027/28
IEEFA · Utility Dive
70 GW
Virginia data-center interconnection demand (all-time peak: 25)
Dominion / Virginia Business
5.4 ML
Water consumed training a single GPT-3-scale model
Ren et al., UC Riverside
0
AI-designed drugs approved by FDA as of April 2026
Industry disclosures
Section I 01

How big the footprint actually is

Measurement, not modeling. What the best available accounting says about electricity, water, carbon, and hardware as of early 2026.

Electricity

The IEA's April 2025 Energy and AI report is the best consolidated accounting. In 2024, data centers drew ~415 TWh globally (≈1.5% of world electricity, growing at ~12%/yr), of which the US took roughly 180 TWh, China ~100 TWh, and the EU-4 ~41 TWh. IEA's Base Case puts the global figure at ~945 TWh by 2030 and ~1,200 TWh by 2035; a "Lift-Off" case reaches ~2,000 TWh.

LBNL's December 2024 US assessment — the most authoritative US-specific data — puts 2023 consumption at 176 TWh (4.4%) and 2028 at 325–580 TWh (6.7–12%). For scale, the upper bound approximates the total annual electricity consumption of Germany. BloombergNEF's December 2025 update projects 106 GW of US data-center capacity by 2035 (~8.6% of electricity), a 36% upward revision in nine months. LBNL's own 2016 projections undershot 2018 reality; all current forecasts should be read as floors.

Figure 01
Global data-center electricity demand, 2020–2035
IEA's Base Case doubles consumption by 2030. Its Lift-Off scenario implies ~2,000 TWh — more than India's current total electricity use.

Regionally, the concentration is extraordinary. Dominion Energy's contracted data-center pipeline in Virginia jumped from 21 GW (July 2024) to 47 GW (October 2025); total Virginia DP requests hit ~70,000 MW, nearly 3× the statewide all-time peak. ERCOT's large-load interconnection queue stands at 226 GW as of late 2025, roughly three-quarters data-center, against a 2023 peak of 85.5 GW.

Ireland's data centers consumed 24% of national electricity in 2024, up from 5% in 2015, forcing Dublin's four-year grid-connection moratorium and a bespoke regulatory regime. PJM's capacity auction clearing prices rose from $28.92/MW-day for 2024/25 delivery to the $333.44/MW-day cap for 2027/28 — a 1,050% increase that the PJM market monitor attributes 63% to new data-center demand, costing ratepayers an estimated $9.3 billion in a single auction.

Figure 02
Regional concentration · Virginia's pipeline & PJM's auction cap
Two views of the same phenomenon. Data-center demand has broken both the largest US interconnection pipeline and the largest US capacity market inside eighteen months.

Water

Shaolei Ren and colleagues at UC Riverside provide the most rigorous accounting. Training GPT-3 consumed roughly 700,000 liters of onsite water in Microsoft's US fleet; with electricity-generation water added, 5.4 million liters total. Inference adds a 500 mL bottle per ~10–50 queries. GPT-4 has not been separately quantified because OpenAI and Microsoft do not disclose. Scaling forward, Ren projects 4.2–6.6 billion cubic meters of AI water withdrawals annually by 2027 — more than four Denmarks. US data centers drew ~211 billion gallons of indirect (thermoelectric) water in 2023.

The siting pattern is perverse: data centers are clustering in water-stressed regions. Google's planned Cerrillos, Chile facility sought ~7 billion liters annually, equivalent to 80,000 residents' consumption, before Chile's environmental court partially reversed the permit in September 2024 and Google switched to air cooling. Google's Uruguay project sought 7.6 million L/day during that country's worst drought in 70 years. Meta's Talavera de la Reina project in Spain's drought-stressed Tagus basin was scaled back 24% after regulatory pressure. TSMC's Taiwan fabs draw 156,000 tonnes/day and required emergency water trucking during the 2021 drought; its new Arizona Fab 1 uses 4.75 million gallons/day.

Disclosure Gap
Microsoft's global WUE of 0.30 L/kWh, widely advertised, masks wide regional variation — Arizona sites reach 1.52 L/kWh — and excludes embodied water in chip manufacturing, which Ren estimates would roughly 10× the total.
Figure 03
Water at scale — from a training run to a continental forecast
Log scale, in liters. Direct and indirect draw combined where both reported. Ren's 2027 projection (low end of scenario range) is shown for aggregate AI.

Carbon emissions

The hyperscalers' own disclosures, read straight, tell the story. Google's total GHG emissions are up ~51% over its 2019 baseline as of its 2025 Environmental Report; Scope 3 rose 22% in a single year and now comprises 73% of the footprint. The company quietly relabeled its 2030 net-zero target "ambition-based."

Microsoft reports total emissions up 23.4% over its 2020 baseline, with Scope 3 up 26% and comprising 97% of the footprint; Brad Smith's public language has moved from "moonshot" (2020) to "the moon has moved" (2025) to "marathon, not sprint."

Amazon's 2024 total hit 68 Mt CO₂e (+6% YoY); Stand.earth calculates its Scope 1 has risen 162% since the 2019 Climate Pledge was launched. Meta discloses emissions on a market-based basis (1,658 tCO₂e for its data-center Scope 2 in 2024) that is roughly 3,000× lower than the location-based figure of 5.14 million tCO₂e for the same activity — a disclosure gap now the subject of a 2026 shareholder proxy resolution.

The 2025 NewClimate Institute / Carbon Market Watch Corporate Climate Responsibility Monitor concluded that the targets of Amazon, Apple, Google, Meta and Microsoft "fall short of demonstrating credible leadership."

Figure 04
Hyperscaler emissions against their own pledged baselines
Percent change from each company's self-declared climate-pledge baseline year. All four committed to net-zero or carbon-negative trajectories during 2019–2020.

Model-specific training emissions are disclosed for open-weight models (Llama 1–4: 300, 539, ~1,900, and 1,999 tCO₂e) but deliberately obscured for frontier commercial models (GPT-4 estimates range 7,000–15,000 tCO₂e; Gemini, Claude, and DeepSeek disclose nothing). Hugging Face's Sasha Luccioni has demonstrated that inference emissions can exceed training within months of deployment, yet no hyperscaler reports per-query or per-user inference emissions. This is roughly analogous to an automobile industry that refuses to publish fuel-economy numbers.

Hardware, minerals, and e-waste

The Wang et al. 2024 paper in Nature Computational Science projects cumulative 1.2–5.0 million tonnes of generative-AI e-waste by 2030, up from ~2,600 tonnes annually in 2023 — a near-thousandfold increase. GPU operational lifespans are short: Meta's Llama 3 training run on 16,384 H100s logged a ~9% annualized GPU failure rate, implying refresh cycles of 1–3 years versus the 5–6-year accounting lives hyperscalers report. NVIDIA accelerators consume 700–1,500 W each; the Blackwell generation is nearly double Hopper.

Critical minerals are a second chokepoint. China controls ~98% of global gallium and roughly 90% of rare-earth processing. Beijing imposed export licensing on gallium and germanium in August 2023, banned exports to the US in December 2024, added seven heavy rare-earths in April 2025, and then suspended parts of the restrictions after the November 2025 Xi–Trump meeting. European gallium prices rose ~365% across that cycle.

PFAS ("forever chemicals") are still required at essentially every advanced lithography node, with semiconductor manufacturing accounting for roughly 10% of EU fluoropolymer demand; substitutes exist in the lab but not at production scale.

Figure 05
Generative-AI e-waste, 2023 → 2030
Unit: thousand metric tonnes. The scenario range reflects Wang et al.'s low, mid, and high assumptions for model size, refresh cycle, and deployment intensity.
The single most important finding for the AI-and-climate debate is that, at the margin, new AI load is being served by new natural gas and deferred coal retirements — not by clean generation. — Finding of this report
Section II 02

The grid is getting dirtier where AI is being built

The mix serving AI load in 2024 is majority fossil. The marginal additions through 2028 are mostly gas, not renewables, and at least fifteen US coal plants have had their retirements pushed back.

The IEA's own data show the US data-center power mix in 2024 was roughly 40% gas, 24% renewables, 20% nuclear, and 15% coal — majority fossil. Global Energy Monitor's early-2026 tally shows the US leading the world in new gas development, with 80.6 GW in Texas alone and ~40 GW tied directly to data-center demand.

Distilled.earth and S&P Global identify ~50 GW of behind-the-meter gas projects announced for data centers, 90% of them in 2025 alone. Kinder Morgan and Energy Transfer project data-center gas demand of 6–10 Bcf/day — roughly 10% of current US gas consumption — within two years. Kinder Morgan and GE Vernova have heavy-duty turbine lead times of 5–7 years, which is itself becoming a binding constraint on the buildout.

Figure 06
US data-center electricity mix, 2024
The mix currently serving AI workloads is majority fossil. Announced marginal additions through 2028 are roughly 80% gas.

55% fossil (gas + coal) is the current mix, before any of the behind-the-meter gas buildout clears interconnection.

Hyperscaler PPAs for renewables (~84 GW contracted across the top four US buyers) are overwhelmingly annual-matching, not 24/7.

AI training runs continuously. At night, the marginal generator is still gas.

Coal is being extended rather than retired. DeSmog and Utility Dive document at least fifteen US coal units whose retirement dates have been pushed back since January 2025, including Plant Bowen, Plant Scherer Unit 3, North Omaha Unit 5, Gibson Generating Station, Dave Johnston, Columbia Energy Center, and Dominion's Clover Power Station. Southern Company CEO Chris Womack has publicly committed to "extend coal plants as long as we can" to serve data-center load in Georgia. The Trump administration's Department of Energy has repeatedly invoked Section 202(c) emergency orders to keep plants running past planned retirement.

The new nuclear deals receive disproportionate press attention but have added zero incremental gigawatts to the grid as of April 2026. Microsoft's Three Mile Island / Crane Clean Energy Center restart targets 2028; Google's Kairos Hermes 2 SMR targets 2030; Meta's Clinton PPA keeps existing capacity online but is not new generation; most SMR "deals" are MOUs without firm PPAs. Meanwhile hyperscaler PPAs for renewables are overwhelmingly annual-matching, meaning that continuous AI training loads are in practice still being served by whatever is on the grid at night — which is increasingly gas and extended coal.

Figure 07
Fifteen US coal units whose retirements were pushed back
Years of delay since January 2025, largest cases. All cited AI / data-center load in their regulatory filings or public announcements.

Two flagship sites

xAI's Colossus in South Memphis installed dozens of unpermitted methane turbines adjacent to the predominantly Black Boxtown neighborhood, where asthma and cancer rates already run roughly 4× the national average; an expansion in Southaven, Mississippi will add up to 41 permanent turbines and produce an estimated >1,700 tons/year of NOx, likely the single largest NOx source in the Memphis metro. The Southern Environmental Law Center, NAACP, and local community groups are pursuing enforcement actions and appeals through 2026.

Meta's Hyperion campus in Richland Parish, Louisiana is being served by an Entergy natural-gas buildout that expanded in March 2026 to ten new gas plants totaling ~7.5 GW — more than six times New Orleans's peak demand and roughly 30% of Louisiana's entire grid. The Union of Concerned Scientists estimates Louisiana ratepayers face $26 billion in incremental electricity system costs over 15 years and an additional $90 billion in public-health and climate damages from the data-center buildout, while Meta's actual lease runs only four years under its Blue Owl joint-venture structure.

The ratepayer-cost question is emerging as the core political fault line. Virginia's JLARC December 2024 study concluded that data centers currently pay cost-of-service but that future system costs will shift to other customers under current structures, with residential bills projected to rise from $159/month today to as much as $381 by 2045 absent policy change. The Virginia State Corporation Commission approved a new GS-5 rate class for >25 MW customers in November 2025 requiring 85% minimum demand payments — a genuine but partial correction. Eight other states introduced similar bills in their 2026 sessions. Maine passed the first statewide data-center moratorium in April 2026. Data Center Watch tallies $98 billion and 20 blocked or delayed projects in Q2 2025 alone across 11 states, with 188 active opposition groups.

Section III 03

Compute as a strategic resource — and the nuclear analogy

The US and its allies are now treating AI compute as strategic infrastructure on a spectrum from the petrodollar to fissile material. The analogy holds in five respects and breaks down in six.

Five policy inflection points frame the regime: BIS's October 2022 export controls; their October 2023 tightening; the January 2025 AI Diffusion Rule (three-tier framework); the Trump administration's May 2025 rescission of Diffusion; and the January 14, 2026 Section 232 tariff on advanced AI chips (25%, with broad US-domestic exemptions) paired with a revived H20/H200 licensing regime that now includes an unprecedented 15% revenue-share to the US Treasury on Nvidia and AMD sales to China. This is not orthodox trade policy — it is state participation in the export margin — and it has no nuclear-era precedent.

Figure 08
Policy inflection points, October 2022 → January 2026
The regime has accreted five major layers in thirty-nine months. Each was a significant departure from the previous trade-policy equilibrium.
October 2022
BIS Advanced Computing Controls
First-ever end-use controls on advanced AI chips. Prohibits export of leading-edge logic (≤14 nm) and HBM memory to China. Catches Nvidia A100/H100 and equivalent foreign-made chips.
October 2023
Tightening + A800/H800 loophole close
Performance thresholds revised to block Nvidia's China-specific A800 and H800 parts. Introduces "total processing performance" and "performance density" metrics. Expands covered countries beyond China to include any country at risk of diversion.
January 2025
AI Diffusion Rule (three-tier framework)
Biden administration's parting regulation: 18 Tier-1 allies with no restrictions, ~120 Tier-2 countries with compute caps and Validated End User pathway, Tier-3 (China, Russia, etc.) effectively embargoed. Industry backlash within weeks.
May 2025
Diffusion Rule rescinded
Trump administration replaces the three-tier framework with bilateral deals. Gulf states (UAE, Saudi Arabia) receive expanded access via HUMAIN and G42 pathways. Tier-2 category effectively dissolved.
January 2026
Section 232 tariffs + 15% revenue share
25% tariff on advanced AI chips with broad US-domestic exemptions. H20/H200 exports to China licensed on condition of a 15% revenue-share to the US Treasury — direct state participation in the export margin, unprecedented in nuclear or trade policy.
Sources: CRS R48642 · A&O Shearman · PwC · Fortune · CNBC

Parallel sovereign AI programs have crystallized in every major economy. The scale of simultaneous industrial-policy activation is unprecedented since the 1950s.

European Union
InvestAI

€200 billion mobilization by 2030, including €20 billion for four to five AI Gigafactories. The largest coordinated EU industrial policy since the Single Market.

€200 B · 5 GW target
France
Paris AI Action Summit

€109 billion package announced February 2025. Focus on sovereign compute infrastructure and dedicated nuclear allocation for data centers.

€109 B · nuclear-backed
United Kingdom
AI Growth Zones

Planning-law carve-outs for designated zones starting at Culham. Commits to 20× expansion of public compute.

Planning-law override
UAE
Stargate UAE / G42

5-GW Abu Dhabi campus announced May 2025. ~35,000 Blackwell-equivalent chip approvals under the Trump Gulf-tour bilateral framework.

5 GW · 35k chips
Saudi Arabia
HUMAIN

PIF-backed sovereign AI with US chip licensing under bilateral deal. Part of the Gulf's pivot from petrodollar to "compute-dollar."

PIF-funded
India
IndiaAI Mission

34,000 subsidized GPUs deployed through common-compute scheme. Explicit sovereign-AI framing; BharatGen foundation models as complement.

34k GPUs
China
Eastern Data, Western Computing

~246 EFLOPS of national compute mid-2024. National 55,000-km optical fabric launched late 2025. Huawei Ascend 950PR in mass production.

246 EFLOPS · 600k chips/yr
United States
Stargate + Section 232

OpenAI–Oracle–SoftBank $500B Stargate commitment. Section 232 tariffs and bilateral compute-for-alignment deals with allies.

$500 B pledge

Where the nuclear analogy holds — and where it breaks

The nuclear-munitions analogy — popularized by Leopold Aschenbrenner's Situational Awareness and embedded in BIS's end-use controls — is partly right and partly misleading. The details matter, because policy built on a flawed mental model (as the rescinded AI Diffusion Rule was) consistently lags diffusion dynamics.

Where it holds

  1. Chokepoint geography. EUV lithography at ASML, leading-edge fabrication at TSMC, HBM at SK Hynix / Samsung / Micron, and GPU design at Nvidia all resemble enrichment facilities: a handful of sites, non-fungible, controllable.
  2. Tiered-access architecture. VEU, Entity List, and model-weights controls closely track the NPT / NSG / IAEA template of differentiated access and end-use verification.
  3. Deterrence logic. Concentrated capability matters more than aggregate stockpile. A few training runs at the frontier matter more than millions of inference cycles.
  4. Dual-use duality. A training cluster is indistinguishable from a civilian data center until it isn't. Same physics problem as reactor vs. enrichment cascade.
  5. Sovereign-first response. HUMAIN, G42, IndiaAI, BharatGen, and EU Gigafactories are all proliferation-style hedges, exactly the dynamic seen in 1960s nuclear.

Where it breaks down

  1. Digital exfiltration. Model weights fit on a USB drive. DeepSeek R1's open-weighting did more to diffuse capability than any smuggled chip could. No nuclear analogue.
  2. Commercial ubiquity. Tens of millions of restricted accelerators in play against a few hundred reactors worldwide. Enforcement surface is categorically different.
  3. Moore's Law depreciation. Today's controlled frontier is tomorrow's commodity on a 2–3-year cycle. Fissile material does not depreciate.
  4. No NPT-style universality. Wassenaar is consensus-bound and was bypassed by UAE / Malaysia / Singapore cloud-access routes almost immediately.
  5. 15% revenue-share. State participation in the export margin has no nuclear counterpart. Looks more like petroleum concessions or 19th-century railroad land grants.
  6. Verification immaturity. IAEA-style audits of compute use remain technologically speculative. The regime is running without its verification arm.

A more accurate composite analogy combines rare-earths supply politics, nuclear nonproliferation architecture, 19th-century railroad land grants, and petrodollar-style bilateral financial entanglement. Virginia's data-center tax exemptions — projected originally at $1.5M annually, now running above $1.6 billion — are textbook grant-underpricing that would be familiar to anyone who has read the 1880–1920 litigation over railroad rate-setting. The JLARC report reads like a 21st-century echo of that history.

The rare-earths analogy is also more useful as precedent than caricature. Academic re-analysis of China's 2010 Senkaku embargo finds the supposed targeted cutoff less sharp than popular memory suggests — but the narrative of rare-earth leverage drove a decade of diversification, and China's 2025 cascade of minerals controls shows the playbook remains active. The oil analogy is partial: like oil, compute is a strategic input with geographic concentration and capacity to anchor a reserve-currency arrangement (Gulf AI campuses ≈ petrodollar recycling, now "compute-dollar"). Unlike oil, chips depreciate in months and are substitutable across generations; there are no proven reserves to monetize.

Under the AGI-solves-it argument, the footprint has to be recoverable. The grid asset lifetimes say it is not. — On the climate math
Section IV 04

Steel-manning the case for these costs

The strongest version of the proponent's argument rests on four pillars. Each is serious; each deserves engagement, not dismissal.

Proponents' core argument is a wager: the near-term costs of AI infrastructure are small relative to the long-term gains from transformative applications. Here is the best version of that case, expressed in the language its defenders would accept.

Pillar One — Scientific wins are real and non-trivial

AlphaFold 2 and 3 reshaped structural biology; Hassabis and Jumper shared the 2024 Nobel in Chemistry. DeepMind's GNoME (November 2023) identified 2.2 million new stable crystal structures; Microsoft's MatterGen extends the approach generatively. GraphCast, Aurora, NVIDIA Earth-2, and ECMWF's AIFS have been deployed operationally — ECMWF since February 2025, NOAA's new AI-based Global Forecast System since December 2025 — with roughly two orders of magnitude less compute than traditional physics-based models and comparable or better skill. AlphaProof and AlphaGeometry reached silver-medal performance on the 2024 International Mathematical Olympiad and gold-medal performance in 2025. These are not marketing. They are concrete, peer-reviewed contributions that would be difficult to achieve without the compute stack in question.

Pillar Two — The efficiency trajectory is real and rapid

Epoch AI's work indicates training compute efficiency has been doubling roughly every 8–9 months — faster than Moore's Law. NVIDIA's Blackwell generation delivers roughly 25–50× inference throughput per watt of Hopper at the rack level, a gain of ~1,000,000× across six GPU generations. Small models — Microsoft's Phi-3/4, Google's Gemma-2, Meta's Llama 3.2 1B/3B — now deliver in 2026 what required frontier models in 2022. DeepSeek V3/R1 (December 2024, January 2025) demonstrated that state-of-the-art training can be achieved at a fraction of the compute previously thought necessary. Satya Nadella and others invoke Jevons not as a confession but as a claim: cheaper compute expands access, democratizes the technology, and allows efficiency gains to be reinvested in higher-value work.

Pillar Three — Productivity and competitiveness

Goldman Sachs estimated a 7% boost to global GDP and 1.5 percentage points of annual productivity growth over a decade; McKinsey estimates $4.4 trillion in annual potential. Firm-level studies have shown 15–55% task-level gains — the MIT Noy-Zhang writing study, GitHub Copilot trials, and BCG's 40% performance boost on consulting tasks. Penn Wharton's September 2025 model places long-run GDP effects at 1.5% by 2035 and 3% by 2055 — more modest than Goldman but still material. If compute is the input to a general-purpose technology on the order of electricity, a multi-decade lag before macro productivity appears (Paul David's thesis about electrification) is expected, not a refutation.

On national security, Dario Amodei's January 2025 essay On DeepSeek and Export Controls makes the clearest version of the race argument: if frontier AI produces decisive military or economic advantage, then concentrating capability in democracies is an overriding priority, and the electricity cost is a rounding error against civilizational stakes.

Pillar Four — The AGI wager

Machines of Loving Grace (October 2024) and Sam Altman's The Gentle Singularity (June 2024) — with more technical hedging by Demis Hassabis — treat near-term emissions as a short-term investment against a compressed 21st-century window of scientific and medical progress. Amodei's specific claims — a decade's worth of biology progress in five to ten years post-AGI, near-elimination of most mental illness, doubling of global development rates in poor countries — are counterfactual, but the argument structure is coherent: expected value is dominated by tail outcomes if one accepts both (a) a non-trivial probability of powerful AI this decade and (b) that compute investment materially shortens the timeline. Under that structure, even several gigatonnes of incremental CO₂ look small against the present value of accelerated decarbonization, disease cures, and discovery.

Additionality — the closing move
Hyperscaler capital is funding new nuclear restarts (Three Mile Island / Crane), new SMR deployment (Kairos, X-energy), new renewables at scale (>100 GW contracted by hyperscalers globally), and grid modernization (DeepMind–Tapestry, DERMS pilots) that would not be happening at this pace without data-center demand. Google's 24/7 carbon-free-energy framework is the most rigorous corporate clean-energy metric deployed anywhere.
Section V 05

Where the justifications meet the evidence

On current evidence, the wager is losing. But the strongest arguments deserve specific responses, not dismissal.

On scientific wins

The case holds narrowly and does not yet transfer to emissions. Better weather models are genuine; the path from AlphaFold to tonnes of CO₂ avoided runs through pharmaceutical pipelines, policy adoption, and engineering deployment that AI does not itself perform. As of April 2026, zero AI-designed drugs have been FDA-approved. Insilico's flagship INS018_055 remains in Phase 2, not Phase 3. BenevolentAI's BEN-2293 failed Phase 2a in 2023 and the company has since restructured twice. Recursion and Exscientia merged under market pressure. The industry's own discourse has moved from "revolutionary" to "trough of disillusionment." Fusion's critical path is magnets and plasma physics; Commonwealth Fusion's SPARC first plasma is now slated for 2026 with net energy in 2027, and AI is a helpful but non-critical tool. No peer-reviewed study attributes a material fraction of global emissions reduction to AlphaFold, GNoME, or GraphCast; DeepMind's data-center cooling result has not diffused. The "AI solves climate" claim, honestly evaluated, is a plausible future contribution of unknown magnitude — not a ledger entry.

On the efficiency trajectory

This is the weakest link in the proponent case, and it was effectively falsified in 2025. Total AI electricity demand rose in 2025 despite DeepSeek's training efficiency breakthrough. Meta raised 2025 AI capex by roughly 50% after R1's release. The Stanford AI Index 2026 reports DeepSeek V3 uses ~23 Wh per medium prompt against ~5 Wh for Claude — cheaper training, more expensive serving, particularly for reasoning models. Luccioni's AI Energy Score v2 found that of 14 non-reasoning models tested against early-2025 references of similar size, eight used more energy than their predecessors. IEA's own modeling concludes efficiency gains are already baked into the Base Case and that even a "High Efficiency" scenario shaves only about 15% off 2035 demand. Jevons is operating as predicted. Absolute physical throughput — kWh, gallons, tonnes of gas, tonnes of concrete — is rising, and the efficiency narrative should be retired as a climate argument.

The 2025 Natural Experiment
DeepSeek R1's release in January 2025 was the cleanest efficiency shock the AI industry has experienced. Within six months, Meta had raised AI capex ~50%, training-run compute budgets had grown, and total data-center electricity demand continued rising. Every major lab treated cheaper training as license to train more. That is the Jevons paradox operating in real time.

On the macro case

Daron Acemoglu's Simple Macroeconomics of AI (2024) derives a cumulative TFP boost of about 0.5–0.7% over a decade, implying 1.1–1.6% GDP — one-quarter to one-fifth of Goldman's headline number. The 2024–2025 Bureau of Labor Statistics data so far vindicate Acemoglu: US nonfarm business labor productivity grew 2.1% in 2025 (in line with trend), and TFP growth decelerated from 1.5% in 2024 to 0.8% in 2025. Firm-level gains of 15–55% on specific tasks have not aggregated. This could be a Paul David productivity paradox, but "could" is not "has" — two-and-a-half years into the generative-AI era, the macro statistics look ordinary.

On the AGI wager

This is load-bearing optimism, and the physical lock-in is against it. The climate math is unforgiving: Morgan Stanley projects cumulative data-center emissions of roughly 2.5 Gt CO₂e by 2030; Microsoft alone is +23.4% from its 2020 baseline; on-site gas at data centers went from 5% of new demand at end-2024 to 39% in 2025; Louisiana, Virginia, Georgia, and Texas utilities are signing 20–40-year gas-plant contracts. Even if AGI arrives on Amodei's most optimistic timeline, the methane plants serving Meta Hyperion and the delayed retirement of Plant Bowen have already locked in a decade or more of incremental emissions that no AI-enabled climate model can retract. The expected-value argument implicitly assumes the footprint is recoverable; the grid asset lifetimes say it is not.

Historical precedent runs against "we'll figure it out later"

The base rate is discouraging: leaded gasoline (Midgley, 1921), CFCs (Midgley again, 1928), and asbestos all followed patterns of known harm, decades of deferral, and partial remediation after incumbents had captured the regulatory ground. The Montreal Protocol is the heartening exception — and it worked because substitutes existed and coordination was tractable. No equivalently clean substitute exists for AI's current emissions path.

For the productivity question, electrification (Paul David's 20+ year lag) is the most apt analogy. For the emissions question, the crypto-mining precedent (Greenidge, Scrubgrass, Stronghold, Marathon) is the closest structural match: computational demand reactivating fossil assets that would otherwise have retired. AI differs from crypto in producing socially valuable output, but resembles it exactly in the mechanism now unfolding. For the capex-cycle question, railroad mania (1873, 1893) offers a cautionary template, with one important disanalogy: railroad track was a durable asset that quietly enabled later productivity; GPUs are not. There is no AI equivalent of dark fiber.

Section VI 06

What correction looks like, and whether it is adequate

Real corrective mechanisms are assembling, but their binding force arrives years after the emissions are locked in.

Disclosure regimes (EU AI Act, SB 253, AI Energy Score) address information asymmetry but contain no hard emissions caps; efficiency standards don't bind until 2028+; rate-class reforms (Virginia GS-5, Texas SB 6) protect residential ratepayers at the margin but cannot eliminate the secondary grid-cost effects; community resistance is reactive and localized. No jurisdiction has imposed a binding requirement that new large-load AI capacity be matched by additional, time-and-location-matched clean generation. Absent that, the physical trajectory continues.

Figure 09
Regulatory correction · timeline
The binding instruments all arrive after 2025. The enforcement gap — the period during which emissions lock in but rules do not yet apply — is the window currently being used.
December 2024
Virginia JLARC Data Centers Report
Most comprehensive state-level audit to date. Finds data centers pay cost-of-service now but grid costs will shift to residential under current structures — up to $381/month by 2045 without reform.
August 2025
EU AI Act · Article 40 energy reporting begins
Technical-documentation and energy-reporting obligations apply to general-purpose AI providers from 2 August 2025. Fines delayed to August 2026; harmonized standards not due until August 2028.
November 2025
Virginia GS-5 rate class approved
State Corporation Commission creates new rate class for >25 MW customers with 85% minimum demand payments. A genuine but partial correction. Eight other states introduced similar bills for 2026 sessions.
December 2025
Ireland CRU decision · moratorium lifted with conditions
Dublin moratorium ends but now requires dispatchable on-site generation matching maximum import capacity and 80% annual energy from new Irish renewables on a glide path.
April 2026
Maine · first statewide moratorium
Blocks new large-load data centers pending study. Pioneering state-level use of siting authority to constrain buildout.
August 2026
California SB 253 · first enforceable disclosure
Scope 1+2 disclosure from covered companies begins. Third-party assurance from 2027. First legally-enforceable emissions report from hyperscalers, globally.
August 2028
EU AI Act · binding efficiency standards due
Harmonized standards take effect. By this date roughly 80 GW of gas generation will already be under construction for US data centers, with 20–40-year offtake contracts.
Sources: Virginia JLARC · EU AI Act · Irish CRU · California SB 253

The honest assessment: disclosure regimes address information asymmetry but contain no hard emissions caps; efficiency standards don't bind until 2028+; rate-class reforms protect residential ratepayers at the margin but cannot eliminate the secondary grid-cost effects; community resistance is reactive and localized. No jurisdiction has imposed a binding requirement that new large-load AI capacity be matched by additional, time-and-location-matched clean generation. Absent that, the physical trajectory continues.

Independent benchmarking is arriving. Hugging Face's AI Energy Score (v2 late 2025) ranks 166+ models across 10 tasks — but OpenAI, Anthropic, and Google decline to participate, leaving frontier models opaque. The ML.Energy leaderboard and Epoch AI's efficiency tracking provide complementary signals. The 24/7 carbon-free-energy movement (Google, Microsoft, UN-EP) is methodologically superior to annual-matching and should become the default corporate metric; Amazon and Meta still report only annual matching.

Community-level correction is the most active adjustment. Prince William County's Digital Gateway rezoning was voided by a Virginia court in August 2025 on procedural grounds and the voiding affirmed on appeal in April 2026. Loudoun County ended by-right data-center approval in March 2025. The Netherlands cancelled Meta's Zeewolde project in July 2022 and bans hyperscalers outside two designated zones. Chile's environmental tribunal partially reversed Google's Cerrillos permit in September 2024. Maine's statewide moratorium passed in April 2026. Federal legislation from Sanders and Ocasio-Cortez is introduced but unlikely to pass. The cumulative effect is slowing and geographically dispersing the buildout — not halting it.

Conclusion

What this adds up to

Neither the civilizational salvation its proponents describe nor the imminent environmental catastrophe its strongest critics claim. Something in between, and closer to the latter than the former.

The AI infrastructure buildout is a large, measurable increase in the energy, water, carbon, mineral, and ratepayer footprint of the global economy, concentrated in specific communities and grids, being rationalized with a blend of genuine scientific achievements, credible but unrealized productivity forecasts, and a faith-based AGI argument whose physical lock-in effects will outlast any hypothesized payoff window.

Three conclusions follow.

One · Separate the frames

The "AI for climate" framing should be separated from the "AI climate impact" framing. They operate on different timescales and through different mechanisms, and netting them against each other — which is what the AGI-solves-it argument effectively does — is bad accounting. Specific applications (weather prediction, materials screening, narrow optimization) deserve support on their merits; they are not a license for indefinite emissions growth.

Two · The nuclear analogy, used carefully

The geopolitical frame is closer to early nuclear policy than to oil, but the analogy systematically misleads on velocity, digital exfiltration, and commercial saturation. Treating compute as strictly a national-security asset justifies domestic subsidy and allied coordination, but it does not justify exempting data centers from the pollution, water, and ratepayer rules that govern every other large industrial user. Louisiana's fast-tracked gas buildout for Meta, xAI's unpermitted turbines in Memphis, and Virginia's Digital Gateway are not security measures — they are deregulatory capture.

Three · What defensible policy looks like

The most defensible policy is boringly substantive:

  • Mandatory additionality — new AI load must be paired with new, time-matched clean generation;
  • Transparency on a per-model and per-site basis — training and inference, emissions and water, supply-chain embodied carbon;
  • Cost-of-service rate classes that prevent residential ratepayers from subsidizing hyperscaler infrastructure;
  • Binding community-consent procedures around siting.

These corrections are all under construction; none is yet strong enough to match the scale of the buildout. Whether they become so, or whether "we'll figure it out later" becomes the epitaph of yet another industrial wave, is the live question of the late 2020s.

The industry's rhetoric asks to be judged by the technology's long-run potential. The physical record — gas plants built, aquifers drawn down, REC-adjusted footprints that diverge 3,000-fold from location-based reality, delayed coal retirements, blocked drugs and deferred fusion, a Jevons effect that ate efficiency gains within a single year — asks to be judged by what has actually happened. For now, the record is clearer than the rhetoric.

Principal Sources & Further Reading