AI Energy Statistics 2026: Power, Carbon and Nuclear Deals

Last Updated: August 1, 2026

Global data center electricity consumption reached 565 TWh in 2026, a 26% increase from 447 TWh in 2025, with AI-optimized servers alone consuming 175 TWh - up from 95 TWh in 2025 - and projected to out-consume conventional servers for the first time in 2027 per Gartner's June 10, 2026 forecast. The International Energy Agency projects data center consumption to nearly double by 2030 to 945 TWh, approximately 3% of all global electricity demand. AI-focused data centers grew 50% in electricity consumption in 2025 alone - more than 16 times the rate of overall global electricity demand growth of 3% per IEA's April 2026 Key Questions on Energy and AI report.

A single ChatGPT query uses approximately 2.9 Wh of electricity - roughly 10 times a Google search. Image generation consumes thousands of times more than a standard text search. Reasoning models like OpenAI's o3 consume 10 to 70 times more than standard queries. These figures are falling per query as efficiency improves - Google reports a 33x improvement in energy efficiency per query in 12 months. But total AI energy consumption keeps rising because usage grows faster than efficiency gains. This is the central tension in AI energy statistics for 2026.

This guide covers every significant AI energy statistic for August 2026, including global electricity consumption, per-query energy costs, carbon emissions, water usage, hyperscaler nuclear deals, and what it all means for businesses running AI workloads, with every figure linked to a named primary source.

🎯 Before you read on - we put together a free 2026 AI Tools Cheat Sheet covering the tools business leaders are actually using right now. Get it instantly when you subscribe to AI Business Weekly.

Get the Free Cheat Sheet

Sources: Axis Intelligence AI data center energy statistics July 2026, Gartner via AboutChromebooks July 2026, IEA Key Questions on Energy and AI April 2026, Zylos AI research February 2026

Global data center electricity consumption reached 565 TWh in 2026 per Gartner's June 10, 2026 forecast, a 26% increase from 447 TWh in 2025, driven overwhelmingly by AI-optimized servers which alone account for 31% of all data center power - while AI-focused data centers grew 50% in electricity in 2025, more than 16 times the rate of global electricity demand growth of 3% per IEA's April 2026 Key Questions on Energy and AI.

Data center electricity consumption trajectory:

Source: Gartner via AboutChromebooks, IEA Energy and AI report

The AI hardware crossover in 2027:

Conventional servers will consume approximately 195-200 TWh in 2026 and 2027, barely moving. AI-optimized servers will hit 258 TWh in 2027, out-consuming conventional servers for the first time in history per Gartner data cited by AboutChromebooks. This crossover represents a fundamental shift in the data center power profile - infrastructure designed around AI workloads rather than traditional compute.

AI server power density is exploding:

Between 2020 and 2025, AI server power density increased 11 times. The IEA projects a further fourfold increase by 2027 - meaning a single refrigerator-sized AI server rack could draw power equivalent to 65 households per Axis Intelligence's July 2026 data center energy report. Standard data center cooling systems designed for traditional server density are increasingly inadequate for AI workloads, requiring liquid cooling infrastructure that represents significant additional capital investment.

Cooling adds to the load:

Gartner forecasts electricity used by cooling systems alone will climb 22.6% in 2026 to 195 TWh, a direct result of denser AI racks requiring more aggressive thermal management per AboutChromebooks' July 2026 AI energy analysis. Cooling electricity in 2026 is roughly equal to the total electricity consumed by all AI-optimized servers just one year earlier.

Geographic concentration:

The United States accounts for approximately 36% of worldwide data center consumption at around 204 TWh in 2026, with dedicated AI data centers taking roughly 68 TWh of that figure per Gartner. Northern Virginia alone is under severe grid pressure. Virginia data centers consume 26% of state electricity per AIMultiple's energy consumption analysis. Ireland's data centers already consume 21% of national electricity, potentially rising to 32% in 2026 per AIMultiple. The IEA projects US data center energy demand will increase by 130% by 2030 per Brookings Institution's April 2026 AI energy briefing.

For how AI energy spending connects to the broader enterprise AI infrastructure investment picture, our AI spending statistics guide covers the full capex landscape.

A single ChatGPT query uses approximately 2.9 Wh of electricity during inference - equivalent to running a lightbulb for 20 minutes and roughly 10 times the energy of a Google search - while image generation consumes thousands of times more energy than a standard text query and reasoning models like o3 consume 10 to 70 times a standard query per Devera's AI environmental impact analysis.

Per-query energy consumption comparison:

Source: Devera AI environmental impact February 2026, WorldMetrics AI energy statistics June 2026, Zylos AI energy research

The reasoning model energy spike:

Standard ChatGPT queries at 2.9 Wh are already 10x a Google search. Reasoning models like o3 and DeepSeek-R1 have introduced a new tier of energy consumption that rivals video streaming on a per-interaction basis per Devera's analysis. A complex o3 reasoning chain can consume 10-70x a standard GPT-5 query. As enterprises deploy reasoning models for complex analytical workflows - legal analysis, financial modeling, code review - the energy cost per business task increases significantly versus standard chat usage.

Efficiency is improving but usage is growing faster:

Google reports a 33x improvement in energy efficiency per query in 12 months per Devera. The IEA's April 2026 update confirmed AI energy efficiency improved significantly in 2024-2025 per MayhemCode's June 2026 AI energy analysis. But total energy consumption keeps rising because the number of users and the complexity of queries grow faster than efficiency gains. Jevons Paradox - when something gets more efficient and cheaper, people use more of it - applies directly to AI energy consumption. The same efficiency gains that make AI more accessible make the total energy footprint larger.

The model selection decision as an energy decision:

There is a 200x efficiency gap between the lightest and heaviest available AI models per Devera's analysis. Using Claude Haiku instead of Claude Opus when Haiku is sufficient reduces energy consumption by an order of magnitude. Model compression techniques like quantization cut energy use by approximately 50% with minimal performance loss per MayhemCode June 2026. Knowledge distillation delivers roughly 60% faster inference with about 40% fewer parameters while keeping 97% of baseline performance. For enterprises running high-volume AI workloads, model selection is not just a performance decision - it is an infrastructure cost and sustainability decision.

The same model running on Chinese grid infrastructure (DeepSeek) produces roughly 70% more CO2 than the same model on US hyperscaler grids, and region selection within US cloud providers varies significantly - AWS eu-west-3 in Paris runs on 70%+ nuclear versus us-east-1 in Virginia with a mixed coal-and-gas grid per SolidAITech's AI power analysis.

For a complete guide to how different AI platforms compare on efficiency and capability, our AI chatbots comparison guide covers every platform.

AI systems may carry a carbon footprint equivalent to that of all of New York City in 2025 per a December 2025 ScienceDirect paper, with ChatGPT's annual CO2 emissions equivalent to 33,000 US households, training GPT-3 emitting 552 metric tons of CO2 equivalent, and global AI carbon footprint projected to reach 1.8-2.5% of global electricity emissions by 2030 per Gitnux's verified AI environmental statistics.

AI carbon footprint statistics:

Source: Gitnux AI environmental impact statistics June 2026, Food and Water Watch February 2026

The corporate emissions trajectory:

Every major hyperscaler has seen AI-driven emissions increases that are moving them away from rather than toward their stated sustainability goals. Microsoft's total emissions are approximately 30% higher than in 2020, making its 2030 carbon-negative plan harder to attain per Food and Water Watch's February 2026 report. Google's emissions surged nearly 50% from 2019 to present. Meta's 2024 emissions were 70% above 2019 levels. Google acknowledged its 2030 carbon-neutral goal is "a moonshot" in early 2026. Microsoft characterized its carbon-negative goal as "a marathon, not a sprint" per MayhemCode's June 2026 analysis.

The underreporting concern:

A Guardian investigation found that from 2020 to 2022, real emissions from company-owned data centers of Apple, Google, Meta, and Microsoft were over seven times higher than self-reported figures per Food and Water Watch. Most hyperscalers buy renewable energy certificates against operational electricity - these certificates reduce reported Scope 2 emissions without necessarily matching the actual carbon intensity of the electricity consumed hour-by-hour. Actual carbon intensity per unit of compute is generally falling but absolute emissions are rising.

Training versus inference:

Training a frontier model represents a large one-time emissions event: GPT-3 at 552 metric tons CO2e is equivalent to approximately five car lifetimes. But inference - running the model for users - accumulates continuously. At 2.9 Wh per ChatGPT query and 2.5 billion daily prompts per OpenAI's disclosure, the daily inference energy consumption of ChatGPT alone is substantial. The question of whether to report training or inference emissions dominates sustainability disclosure discussions in 2026.

For how AI's carbon footprint connects to enterprise sustainability strategies, our AI adoption statistics guide covers how organizations are factoring sustainability into AI procurement.

Microsoft's global data centers consumed 6.4 million cubic meters of water in FY2022 alone - a 34% jump over 2021 - while hyperscaler water consumption rose 25-40% year-over-year in 2024-2025 disclosures, and a 100-query ChatGPT session translates to approximately 0.5 liters of water in worst-case high-evaporation regions per Presenc AI's data center energy analysis.

Water consumption statistics:

Source: Presenc AI data center energy 2026, SolidAITech AI power analysis, Zylos AI energy research

Water is the emerging AI infrastructure constraint:

Energy has dominated AI sustainability discussions. Water is the emerging constraint. AI-intensive data centers in Arizona, Phoenix, and Northern Virginia are already creating local water pressure per Presenc AI. In 2026, environmental clearance for hyperscale facilities is increasingly tied to Water Usage Effectiveness metrics. The European Commission expects to roll out regulations requiring data center operators to meet minimum WUE performance standards per Zylos AI's energy research.

Cooling technology choices:

Traditional air cooling is being replaced by liquid cooling in AI data centers. Liquid cooling reduces direct water use by 70-90% compared to air cooling - but increases capital cost significantly and has indirect water footprints through electricity generation. The shift to liquid cooling is one of the most significant data center infrastructure changes driven by AI server density requirements, and it is creating a capital expenditure wave in data center construction that extends beyond the GPU and server purchases most coverage focuses on.

For the complete enterprise AI infrastructure picture, our AI spending statistics guide covers the full capex and infrastructure investment data.

Microsoft signed a $16 billion, 20-year deal to restart Three Mile Island Unit 1 for 835 MW of nuclear power targeting 2028 operation, Google signed the first US corporate SMR fleet deal with Kairos Power for 500 MW, and Amazon secured a 1.92 GW nuclear power purchase agreement from the Susquehanna plant plus $500 million in SMR investment per Zylos AI's energy research.

Hyperscaler nuclear deals:

Source: Zylos AI energy research February 2026, MayhemCode June 2026

Why nuclear and not just renewables:

Solar and wind are intermittent - they generate power when the sun shines and wind blows, not necessarily when a data center needs power 24 hours a day, seven days a week. Nuclear provides carbon-free baseload: continuous power regardless of weather, time of day, or season. For AI data centers running GPU workloads that cannot simply pause when renewable generation drops, baseload power is not optional. Goldman Sachs estimates data center electricity demand could rise 160% by 2030 per Zylos AI - a scale that requires every available clean energy source simultaneously, not a choice between nuclear and renewables.

The Three Mile Island story:

Microsoft's restart of Three Mile Island - renamed the Christopher M. Crane Clean Energy Center - is the most symbolically significant energy deal in tech history. The plant that became synonymous with nuclear accident fear in 1979 is now being restarted to power AI data centers in 2028. The deal represents a fundamental reframing of the nuclear conversation in the United States from risk management to strategic energy infrastructure for the AI era.

SMRs as the long-term solution:

Small Modular Reactors offer advantages over traditional large nuclear plants: faster construction, modular deployment that scales with demand, and lower upfront capital. Google's Kairos Power deal and Amazon's SMR investment reflect bets on SMR technology reaching commercial scale in the early 2030s. Between now and then, the capacity gap is being filled with natural gas - meaning the nuclear deals improve the long-term trajectory but do not resolve the near-term emissions challenge per MayhemCode.

For how AI infrastructure investment connects to the broader AI market, our AI market share 2026 guide covers how energy infrastructure is shaping competitive dynamics.

The Big Five hyperscalers - Amazon, Alphabet, Meta, Microsoft, and Oracle - will spend approximately $725 billion on AI infrastructure in 2026 alone, a 77% increase over the record $410 billion deployed in 2025, while a structural US capacity shortfall of 9.3 GW is constraining deployment timelines regardless of capital availability per Axis Intelligence's June 2026 data center statistics.

Hyperscaler AI infrastructure spending:

Source: Axis Intelligence AI data center statistics June 2026, Zylos AI energy research

The capacity constraint is as binding as the capital:

The Big Five are willing to spend $725 billion on AI infrastructure in 2026. The constraint is not capital - it is physical capacity. The 9.3 GW structural shortfall in the US market means data center construction is backlogged, power interconnection queues at utilities stretch years, and transmission infrastructure cannot be built fast enough to meet demand per Axis Intelligence. This capacity constraint is why Microsoft and Google are signing 20-year nuclear deals for power that will not arrive until 2028-2035 - they are securing future capacity now because near-term capacity is already committed.

The single most striking data point:

The Big Five will spend more on AI infrastructure in 2026 than the GDP of Switzerland per Axis Intelligence. That is not a technology investment. It is a physical infrastructure program with implications for the global electricity grid, water systems, and semiconductor supply chains simultaneously.

For complete Microsoft financial data including its AI capex context, our Microsoft Copilot statistics guide covers Microsoft's AI infrastructure picture. For Meta's capex trajectory, our Meta AI statistics guide covers the full picture.

Per-query AI energy efficiency is improving rapidly - Google achieved a 33x improvement in 12 months and model compression techniques like quantization cut energy use by 50% with minimal performance loss - yet total AI energy consumption keeps rising because usage grows faster than efficiency gains, a pattern known as Jevons Paradox that applies directly to AI infrastructure in 2026.

The Jevons Paradox in AI:

Jevons Paradox states that when the efficiency of using a resource improves, total consumption of that resource typically increases because lower costs drive higher usage. Applied to AI energy: every improvement in per-query efficiency makes AI cheaper and more accessible, which attracts more users, enables more complex applications, and produces more queries. The net effect is rising total energy consumption despite falling per-unit consumption.

The IEA's April 2026 analysis confirms this directly: "power consumption per AI task is declining rapidly - by at least an order of magnitude annually in recent years - yet total consumption still surges because the number of users and the complexity of use cases, particularly AI agents, are growing faster than efficiency gains" per IEA Key Questions on Energy and AI.

Efficiency tools that matter:

For businesses running high-volume AI workloads, these efficiency decisions compound significantly at scale. A 50-person team running 500 daily AI queries at the wrong model tier versus the optimal tier represents the difference between minimal infrastructure cost and meaningful electricity and carbon expenditure.

Three practical implications follow directly from the AI energy statistics for any business deploying AI in 2026.

1. Model selection is an infrastructure cost decision.

The 200x efficiency gap between models means your choice of Claude Haiku versus Claude Opus, or GPT-5.5 Instant versus GPT-5.6 Sol Pro, is an infrastructure cost decision as much as a quality decision. For high-volume, lower-stakes queries - customer support responses, content classification, data extraction - a smaller model delivers comparable accuracy at a fraction of the cost and energy. Reserving frontier models for high-stakes tasks where the quality difference justifies the premium is not just budget management - it is energy management.

In conversations with executives building AI-enabled products, the teams that optimized model selection by task type consistently discovered they could achieve 60-80% cost reductions without noticeable quality degradation on the majority of their queries.

2. Cloud region selection affects your AI carbon footprint.

If your organization has sustainability reporting requirements, where you run AI workloads matters as much as how efficiently you run them. Microsoft Azure's Sweden Central region runs on nearly 100% hydro and wind. Google's us-central1 in Iowa has high renewable percentage. AWS eu-west-3 in Paris runs on 70%+ nuclear. The same workload produces dramatically different Scope 2 emissions depending on where it runs per SolidAITech's analysis.

3. The regulatory environment around AI energy is tightening.

The European Commission is requiring minimum Water Usage Effectiveness standards for data centers. The EU AI Act includes energy provisions. Environmental clearance for hyperscale facilities is increasingly tied to sustainability metrics. For enterprises that build and operate AI infrastructure rather than purely consuming cloud services, regulatory compliance around energy and water use is becoming a site-selection and engineering requirement, not just a reporting exercise.

For our complete guide on implementing AI sustainably within enterprise workflows, our how to implement AI in business guide covers the full operational framework.

AI Spending Statistics 2026
The complete AI infrastructure investment picture - where the $725B in Big Five capex fits in total AI spending.

AI Adoption Statistics 2026
Enterprise AI deployment rates and how sustainability factors are entering procurement decisions.

Microsoft Copilot Statistics 2026
Microsoft's AI business including the Three Mile Island nuclear deal context.

Meta AI Statistics 2026
Meta's $125-145B capex program and its nuclear RFP for 1-4 GW of power.

AI Market Share 2026
How energy infrastructure is shaping competitive dynamics across ChatGPT, Claude, Gemini, and Grok.

AI Chatbots Comparison Guide 2026
Platform comparison including efficiency differences relevant to energy cost decisions.

AI ROI Statistics 2026
Where energy costs fit in the complete AI return-on-investment calculation.

AI Statistics 2026: The Complete Data Guide
The master hub for all AI statistics including energy and infrastructure data.

How much electricity does AI use in 2026?
Global data center electricity consumption reached 565 TWh in 2026, a 26% increase from 447 TWh in 2025, per Gartner's June 10, 2026 forecast. AI-optimized servers alone consume 175 TWh in 2026, up from 95 TWh in 2025, and are projected to reach 258 TWh in 2027 - out-consuming conventional servers for the first time in history. The IEA projects global data center consumption to nearly double to 945 TWh by 2030, representing approximately 3% of all global electricity demand. AI-focused data centers grew 50% in electricity consumption in 2025 - more than 16 times the rate of overall global electricity demand growth of 3%. The United States accounts for approximately 36% of worldwide data center consumption at around 204 TWh. Source: Gartner via AboutChromebooks July 2026, IEA April 2026

How much energy does a ChatGPT query use?
A single ChatGPT query uses approximately 2.9 Wh of electricity during inference, equivalent to running a lightbulb for 20 minutes and roughly 10 times the energy of a Google search at approximately 0.3 Wh. Google Gemini text prompts use approximately 0.24 Wh, 0.26 ml of water, and emit 0.03 gCO2e per Google's May 2025 technical paper. Image generation consumes thousands of times more energy than standard text queries. Reasoning models like o3 and DeepSeek-R1 consume 10 to 70 times a standard query. Per-query efficiency is improving rapidly - Google reports a 33x improvement in 12 months - but total consumption keeps rising because usage grows faster than efficiency gains. A 100-query ChatGPT session translates to approximately 0.5 liters of water in worst-case high-evaporation regions. Source: Devera AI environmental analysis, Presenc AI data center energy 2026

What is AI's carbon footprint in 2026?
AI systems may carry a carbon footprint equivalent to all of New York City as of 2025 per a December 2025 ScienceDirect paper. ChatGPT's annual CO2 emissions are equivalent to 33,000 US households. Training GPT-3 emitted 552 metric tons of CO2 equivalent. Training a large NLP model at maximum scale can emit 626,000 pounds of CO2, equivalent to five car lifetimes. Global AI carbon footprint is projected to reach 1.8-2.5% of global electricity emissions by 2030. Data center GHG emissions reached 3% of global totals in 2022 with AI accelerating the trend. Microsoft's total emissions are approximately 30% higher than in 2020. Google's emissions are up nearly 50% from 2019. Meta's 2024 emissions were 70% above 2019 levels. Most hyperscalers buy renewable energy certificates against operational electricity, but a Guardian investigation found real data center emissions from major tech companies were 7x higher than reported figures from 2020-2022. Source: Gitnux AI environmental statistics June 2026, Food and Water Watch February 2026

Why is Microsoft restarting Three Mile Island?
Microsoft signed a $16 billion, 20-year deal to restart Three Mile Island Unit 1 - renamed the Christopher M. Crane Clean Energy Center - for 835 MW of nuclear power targeting 2028 operation, to secure carbon-free baseload power for its AI data centers per Zylos AI's energy research. Nuclear offers 24/7 carbon-free power regardless of weather, which solar and wind cannot provide. AI data centers run GPU workloads continuously and need baseload power rather than intermittent renewable generation. Goldman Sachs estimates data center electricity demand could rise 160% by 2030, a scale requiring every available clean energy source. Google signed a similar deal with Kairos Power for 500 MW of SMR capacity coming online 2030-2035. Amazon secured a 1.92 GW nuclear PPA from the Susquehanna plant plus $500 million in SMR investment. Meta issued an RFP for 1-4 GW of new nuclear capacity. Source: Zylos AI energy research February 2026, MayhemCode June 2026

How much are tech companies spending on AI energy infrastructure?
The Big Five hyperscalers - Amazon, Alphabet, Meta, Microsoft, and Oracle - will spend approximately $725 billion on AI infrastructure in 2026, a 77% increase over the record $410 billion deployed in 2025 per Axis Intelligence. This exceeds the GDP of Switzerland. A structural US capacity shortfall of 9.3 GW is constraining deployment timelines regardless of capital availability. Meta's full-year 2026 capex guidance is $125-145 billion, primarily for AI data centers. Microsoft has committed to $80 billion+ in AI infrastructure for 2026. Cooling systems alone will consume 195 TWh of electricity in 2026, a 22.6% increase from 2025, driven by denser AI racks requiring more aggressive thermal management. Source: Axis Intelligence June 2026

What is AI's water consumption in 2026?
Microsoft's global data centers consumed 6.4 million cubic meters of water in FY2022 alone, a 34% jump over FY2021. Hyperscaler water consumption rose 25-40% year-over-year in 2024-2025 disclosures. A 100-query ChatGPT session translates to approximately 0.5 liters of water in worst-case high-evaporation regions. Liquid cooling reduces direct water use by 70-90% compared to traditional air cooling but increases capital cost. In 2026, environmental clearance for hyperscale facilities is increasingly tied to Water Usage Effectiveness metrics. The European Commission expects to roll out regulations requiring minimum WUE performance standards for data center operators. Several hyperscalers have set water-positive goals by 2030 with zero-water cooling emerging as a key design principle for new data center construction. Source: Presenc AI data center energy 2026, Zylos AI research

Is AI's energy use sustainable?
The honest answer is not yet at current trajectory. Per-query AI energy efficiency is improving - Google reports 33x improvement in 12 months and model compression can cut energy use 50% with minimal performance loss. But Jevons Paradox ensures total consumption rises as AI becomes cheaper and more accessible. AI-focused data center electricity grew 50% in 2025 - 16 times faster than overall electricity demand growth of 3%. Microsoft, Google, and Meta have all seen AI drive emissions significantly above their pre-AI baselines, complicating or delaying their stated 2030 sustainability goals. The nuclear deals represent the most credible long-term response: carbon-free baseload power for AI data centers. But Microsoft's Three Mile Island restart does not deliver power until 2028, and Google's SMR deals come online 2030-2035. Between now and then, the gap is filled primarily with natural gas. Whether AI energy use becomes sustainable depends on whether nuclear and renewable energy can scale to meet demand before the emissions trajectory creates irreversible climate commitments. Source: IEA April 2026, MayhemCode June 2026

How does AI energy consumption compare to other industries?
Global data centers consumed 565 TWh in 2026 - approximately 2% of global electricity. For comparison, global cryptocurrency mining at its 2021 peak consumed approximately 150 TWh. The steel industry consumes approximately 8% of global electricity. The aluminum industry approximately 3%. By 2030, with data centers projected at 945 TWh and AI driving most of that growth, AI's electricity footprint will rival some of the most energy-intensive industrial sectors in history. Bloomberg Intelligence predicted growth in energy demand for AI by up to four times its current level by 2032. The IEA projects servers used for AI workloads growing at 30% annually through the decade. Ireland's data centers already consume 21% of national electricity. Virginia data centers consume 26% of state electricity. The geographic concentration of AI infrastructure is creating local grid stress that at scale reflects a global resource challenge. Source: Brookings April 2026, WorldMetrics June 2026

The AI energy statistics of August 2026 converge on a single tension that defines the entire landscape.

Per-query AI energy efficiency is improving faster than almost any technology in history. Google achieved a 33x improvement in 12 months. Quantization cuts energy 50% with minimal performance loss. The 200x efficiency gap between lightest and heaviest models means smart model selection produces dramatic efficiency gains.

Total AI energy consumption is rising faster than the efficiency improvements. 565 TWh in 2026. Nearly doubling to 945 TWh by 2030. AI-focused data centers growing 16 times faster than overall electricity demand. The Big Five spending $725 billion on infrastructure in a single year. Jevons Paradox operating at civilization scale.

The nuclear deals are the most important strategic response: Microsoft restarting Three Mile Island for 835 MW in 2028, Google signing the first corporate SMR fleet deal for 500 MW in 2030-2035, Amazon securing 1.92 GW from Susquehanna. These are not incremental sustainability measures. They are decade-scale bets on carbon-free baseload power for AI infrastructure. If SMR technology delivers at scale by the early 2030s, the long-term AI energy trajectory becomes sustainable. If it does not, the natural gas that fills the gap between now and then creates emissions commitments that compound for decades.

For business leaders running AI workloads, three decisions determine your organization's position in this landscape: model selection by task type, cloud region selection by carbon intensity, and whether to engage with AI energy costs as an infrastructure variable or ignore them until regulation forces the conversation. The businesses that treat energy efficiency as a cost optimization and sustainability priority simultaneously will have lower AI operating costs and stronger regulatory positioning regardless of how the broader energy debate resolves.

The era of treating AI compute as effectively free and effectively clean ended in 2026. The data above is why.