The Pulse
The IEA now projects global AI data center energy consumption to nearly double from 485 TWh in 2025 to around 950 TWh by 2030, making AI infrastructure one of the fastest-growing sources of electricity demand this decade. Microsoft, Google, Meta, and Amazon are now contracting, acquiring, financing, and collocating dedicated power capacity rather than relying solely on utility grids. Goldman Sachs has identified energy availability as the single biggest infrastructure constraint in AI, displacing chip supply as the binding limit.
And the country that consumes roughly the same electricity annually as all US data centres did in 2024 is Pakistan, a nation simultaneously launching a $1 billion AI fund while managing 8 to 12 hours of daily load shedding. The AI energy story is not a future risk. It is a present constraint with active policy, commercial, and geopolitical consequences.
Core Significance
Why it matters:
- The Scale Is Growing Faster Than Any Other Sector: From 2024 to 2030, data centre electricity consumption grows at around 15% per year, more than four times faster than the growth of total global electricity consumption. AI-focused data centres are growing faster still, tripling their consumption in the same period. The IEA’s April 2026 updated projections are not a high-growth scenario. They are the base case.[IEA — Key Questions on Energy and AI, April 2026]
- Big Tech Has Moved Beyond Ordinary Utility Procurement: Microsoft, Google, Meta, Amazon, xAI, Oracle, and OpenAI are all contracting, acquiring, financing, or collocating dedicated electricity generation. The shift from utility customer to energy investor is one of the least-covered but most consequential structural changes in the technology industry in 2026. Energy access is becoming a competitive moat alongside chip access.
- The AI Energy Crisis Is Also Reviving Nuclear Power: The AI energy crisis is not only accelerating renewables. It is also reviving nuclear power as the preferred source of always-on, carbon-free electricity for hyperscale AI infrastructure. Microsoft, Google, Meta, and Amazon have contracted over 10 gigawatts of new nuclear capacity in the past 12 months. That is a structural commitment, not a sustainability marketing exercise.
Deep Context: How AI Changed the Data Centre Energy Equation
The energy story of AI has two distinct chapters.
The first ran from 2017 to 2022. Data centre electricity demand grew faster than the rest of the economy but remained manageable. Efficiency improvements in chip design and cooling largely offset growth in compute volume. The International Energy Agency noted in 2022 that despite large increases in the number of data centres, energy demand had remained roughly level because of efficiency improvements. The grid problem looked contained.
The second chapter opened with ChatGPT. Generative AI changed the equation fundamentally. A generative AI training cluster consumes 7 to 8 times more energy than a typical computing workload. Inference at scale, running models for millions of users simultaneously, consumes electricity continuously and intensively. Efficiency improvements no longer offset volume growth because the volume is growing exponentially.
In the US, data centres already consumed more than 4% of national electricity in 2024, and their demand is projected to more than double by 2030. As Pew Research confirmed from IEA estimates, US data centres consumed 183 TWh in 2024 alone, roughly equivalent to the entire annual electricity consumption of Pakistan. A typical AI-focused hyperscaler annually consumes as much electricity as 100,000 households. The larger clusters currently under construction are expected to use 20 times as much.
As covered in our Nvidia chip competition analysis, the companies that control compute infrastructure set the rules of the AI economy. Energy is now the second dimension of that control. Whoever secures the electricity supply controls how fast the AI economy can grow.
Data Insights
By the numbers:
All data points below are sourced. Primary sources: IEA, Pew Research, Alphabet IR, Constellation Energy, Google Blog, Reuters, Goldman Sachs.
- 485 TWh: Global data centre electricity consumption in 2025, per the IEA’s April 2026 updated projections. Growing at 15% annually.[IEA — Energy Demand from AI]
- 950 TWh: IEA base case projection for global data centre electricity consumption by 2030, representing just under 3% of total global electricity demand. AI-focused data centres triple in this period.[IEA — Key Questions on Energy and AI]
- 1,200 TWh: IEA base case for global data centre electricity consumption by 2035. By then, the US will consume more electricity for data centres than for all energy-intensive manufacturing combined.[IEA — Energy and AI Executive Summary]
- 183 TWh: US data centre electricity consumption in 2024 alone, equivalent to Pakistan’s entire national annual electricity consumption, per IEA estimates cited by Pew Research.[Pew Research — US Data Centre Energy 2025]
- 4%+: Share of total US national electricity consumed by data centres in 2024. Projected to more than double to 8%+ by 2030.[Pew Research — US Data Centre Energy]
- 160%: Goldman Sachs projected growth in data centre power demand by 2030, driven primarily by AI workloads.[Brookings — Global Energy Demands and AI Landscape]
- 150-200 MW: Continuous power consumption of a single large-scale AI data centre running 50,000+ Nvidia H100 GPUs, equivalent to powering 150,000 average US homes.[Vucense — Big Tech Nuclear Energy AI 2026]
- 34.7 GW: Total clean power contracted by Microsoft as of Q1 2026, surpassing Amazon as the largest corporate clean energy buyer globally.[Defiance ETFs — The AI Trade Is Now an Energy Trade]
- 2,000 MW: Electricity allocated by Pakistan’s government specifically for bitcoin mining and AI data centres, per a separate announcement from the Pakistan Crypto Council and finance ministry. This is distinct from the $1 billion National AI Fund.[Reuters — Pakistan 2,000 MW allocation]
Table 1: Big Tech Nuclear and Energy Deals : The 2025-2026 Build-Out
| Company | Energy Deal | Capacity | Timeline | Source |
| Microsoft | Three Mile Island restart (Constellation — Crane Clean Energy Center) | 835 MW | 2028 | Constellation official |
| Meta | Clinton Clean Energy Center, 20-year deal (Constellation) | Plant ~1.1GW class, +30MW uprate | 2027+ | Constellation official |
| Kairos Power SMR fleet (first corporate SMR agreement) | Up to 500 MW | 2030-2035 | Google Blog official | |
| Intersect Power acquisition (Alphabet) | Multiple GW | Ongoing | Alphabet IR official | |
| Amazon | Susquehanna nuclear campus (Talen Energy) | 1.9 GW through 2042 | 2042 | Talen Energy |
| Amazon | X-energy SMR investment ($500M) | 5 GW target by 2039 | 2039 | IEEE Spectrum |
| Microsoft | Total clean power contracted | 34.7 GW | Ongoing | Defiance ETFs |
Table 2: The AI Energy Equation: Scale in Context
| Comparison | Energy Required | Equivalent To |
| Training GPT-3 (one run) | 1,287 MWh | Annual electricity for 120 US homes |
| Training GPT-4 (estimated) | ~50 GWh | Annual electricity for 5,000 US homes |
| Single AI hyperscaler annually | As much as 100,000 households | A medium-sized US city |
| Large AI cluster (50,000 H100s) | 150-200 MW continuous | 150,000 average US homes |
| US data centres 2024 | 183 TWh | Pakistan’s entire national electricity consumption |
| Global data centres 2030 (IEA) | ~950 TWh | More than UK + France combined |
The tables frame the AI data center energy consumption challenge in human terms. The scale is not abstract. It is comparable to national electricity budgets of major economies.
The Business Case: Three Energy Strategies Big Tech Is Pursuing
The AI data centre energy challenge has produced three distinct corporate strategies. Each reflects a different theory about where the constraint will bite hardest and on what timeline.
Strategy 1: Vertical Integration into Generation
As Alphabet’s investor release confirmed, Google agreed to acquire Intersect Power for $4.75 billion in cash plus assumed debt. This is vertical integration into generation rather than offtake. Rather than signing power purchase agreements and waiting for utilities to deliver electricity, Google acquired the company that builds the solar and storage projects outright.
Microsoft has taken a parallel approach through its 34.7 gigawatts of contracted clean power. The strategy is not philanthropy. It is infrastructure control. A company that has locked up 34.7 gigawatts of clean energy capacity has secured a competitive moat that a new AI competitor cannot replicate through financial investment alone. Building nuclear plants and solar farms takes years regardless of budget.
Strategy 2: The Nuclear Commitment
The AI energy crisis is not only accelerating renewables. It is also reviving nuclear power as the preferred source of always-on, carbon-free electricity for hyperscale AI infrastructure. The commitment is already written in decades-long contracts.
As Constellation Energy confirmed, Microsoft signed a 20-year power purchase agreement for 835 MW from Three Mile Island Unit 1, rebranded as the Crane Clean Energy Center, with restart targeted for 2028. Google’s deal with Kairos Power, as Google’s own blog confirmed, covers up to 500 megawatts of new 24/7 carbon-free power with the first small modular reactor targeted by 2030 and further deployments through 2035.
The challenge is timeline. Goldman Sachs has identified energy availability as the single biggest infrastructure constraint in AI. In the near term, the gap between current demand and clean power availability will be filled by natural gas. That is not a choice the industry is celebrating. It is a constraint being managed while the nuclear pipeline develops through the late 2020s and early 2030s.
Strategy 3: Geographic Arbitrage
The new AI infrastructure is highly concentrated in North America, Western Europe, and Asia-Pacific, which together account for more than 90% of projected compute capacity. That concentration is driven partly by energy availability and cost. Iceland, Norway, Canada, and Texas are preferred data centre locations because each has surplus power, favourable grid conditions, and cold climates that reduce cooling costs.
The geographic arbitrage strategy has a direct implication for regions like South Asia. Countries with chronic electricity deficits, insufficient grid infrastructure, and high power costs will not attract significant AI data centre investment through policy ambition alone. As covered in our agentic AI enterprise analysis, the infrastructure layer is the prerequisite that no governance framework can substitute for.
Between the lines:
Long-term nuclear contracts typically 20 to 30 years create a structural advantage for incumbents that cannot be replicated by new entrants. Energy access is becoming a moat alongside chip access. The companies signing 20-year nuclear contracts today are not just securing power for current operations. They are locking out competitors from the energy supply that future AI infrastructure will require. That dynamic compounds. Each year of delay in securing energy capacity is a year of structural disadvantage that financial investment alone cannot reverse.
Regional Spotlight: Pakistan’s AI Ambitions and the Energy Reality
No country illustrates the AI energy tension more sharply than Pakistan. Two government initiatives point in the same ambitious direction while the underlying energy infrastructure points in a different one.
Two Distinct Initiatives, One Shared Challenge:
Pakistan has announced a $1 billion investment in AI by 2030, as Dawn reported from the National AI Policy. Separately, as Reuters confirmed from the Pakistan Crypto Council and finance ministry, Pakistan has allocated 2,000 megawatts of electricity for bitcoin mining and AI data centres. These are two distinct government initiatives. They should not be presented as a single fund architecture unless official government documentation directly connects them. Both point toward the same challenge: Pakistan’s energy infrastructure must keep pace with its AI infrastructure ambitions or neither initiative delivers its intended outcomes.
The Crisis:
Pakistan faces 8 to 12 hours of daily load shedding in 2026, down from 10 to 14 hours the previous year but still economically damaging. The Friday Times’ April 2026 analysis noted Pakistan’s power crisis is intensified by a severe geographical imbalance, with generation assets concentrated in the south and demand concentrated in Punjab’s industrial north. The structural transmission problem means installed capacity is a nominal figure rather than a functional resource during peak demand periods.
The 183 TWh that US data centres consumed in 2024 equalled Pakistan’s entire national electricity budget. A single large-scale AI cluster consuming 150 to 200 megawatts continuously would represent approximately 1 to 2% of Pakistan’s total current national electricity consumption. Hosting that cluster reliably on a grid that already cannot meet residential demand requires dedicated power infrastructure, not a connection to the national grid.
The Opportunity:
For Pakistan, the constraint is not AI talent alone. Unless dedicated power infrastructure is built alongside compute infrastructure, large-scale AI data centres will remain difficult to operate reliably on the national grid. The specific opportunity is co-located renewable energy. Pakistan’s Balochistan and Sindh regions receive some of the highest solar irradiance levels in the world. Utility-scale solar costs have fallen dramatically. A data centre powered by a dedicated solar and storage installation in southern Pakistan is a fundamentally different proposition from a data centre connected to the national grid in Punjab.
Saudi Arabia’s HUMAIN data centres, as covered in our Islamic Military Alliance analysis, are being built with dedicated power infrastructure rather than grid dependency. That model, applied in Pakistan using co-located solar rather than nuclear, is the most credible path to reliable AI data centre operations. The $1 billion National AI Fund’s architecture needs an explicit energy independence strand before the first data centre procurement tender is issued.
Expert Nuance: The Inference Problem Nobody Is Solving
The energy conversation about AI consistently focuses on training. GPT-4 consumed roughly 50 GWh to train. These are large numbers but they are one-time costs. The more important and less discussed energy problem is inference.
Every time a user asks ChatGPT a question, every time Siri 2.0 routes a complex query to Google Gemini as covered in our Siri 2.0 WWDC analysis, and every time an enterprise agent executes a workflow, inference compute runs continuously. Inference is not a one-time training cost. It is an operational cost that scales directly with user adoption.
The difficult reality is that inference demand is growing faster than clean, always-on power can be added before 2030. The IEA’s projection that AI-focused data centres triple their electricity consumption between 2025 and 2030 is driven primarily by inference growth, not training. The efficiency improvements that kept data centre energy flat from 2017 to 2022 are not keeping pace with inference volume growth because the volume is growing at a rate that no efficiency improvement can match at current model scales.
The nuclear deals being signed today will not produce power until 2028 at the earliest. The SMRs will not be operational until the early 2030s. The gap between now and then will be filled by natural gas, generating the carbon emissions the technology industry’s sustainability commitments were supposed to prevent. The tension between AI ambition and climate commitments is not being resolved in 2026. It is being actively deferred.
Strategic Outlook: What’s Next
Three developments will define how AI data center energy consumption evolves over the next three years.
- The 2028 Nuclear Bridge Moment: Microsoft’s Three Mile Island restart in 2028 will be the first large-scale nuclear capacity addition specifically contracted for AI data centre demand. If it delivers on schedule, it establishes the viability of the nuclear AI power model and unlocks the next wave of nuclear contracting. If it is delayed, the natural gas dependency extends further and the carbon accounting for AI infrastructure gets significantly worse. The 2028 date is the critical checkpoint for the entire energy-AI transition.
- AI Efficiency as a Regulatory Requirement: The EU is already developing mandatory AI energy efficiency standards as part of its AI Act implementation. When AI energy consumption becomes a line item in ESG reporting, the pressure to optimise inference efficiency will intensify significantly. That regulatory pressure will accelerate the development of more energy-efficient model architectures, which is the only solution that works in the short term without adding supply. Watch for EU AI energy disclosure requirements in Q3 2026.
- Pakistan’s Solar Window: Pakistan has a specific and time-limited opportunity in the next 18 months. The falling cost of utility-scale solar, combined with Pakistan’s exceptional solar resources in Balochistan and Sindh, creates a credible path to dedicated AI infrastructure power at competitive cost. If the $1 billion National AI Fund and the 2,000 MW allocation are coordinated to explicitly fund co-located renewable energy for each data centre initiative, the energy constraint becomes manageable. That coordination is not currently documented. It needs to be formalised before the first data centre procurement tender is issued.
Key Question Answered
How much energy does AI use and why is it becoming a problem in 2026?
Global AI data center energy consumption reached 415 TWh in 2024 and is projected by the IEA to reach 485 TWh in 2025, growing at 15% annually more than four times the growth rate of total global electricity consumption. The IEA base case projects this doubles to 945 TWh by 2030, with AI-focused data centres tripling their consumption in the same period. In the US alone, data centres consumed 183 TWh in 2024, equivalent to Pakistan’s entire national electricity consumption. A single large AI data centre running 50,000 Nvidia H100 GPUs consumes 150 to 200 megawatts continuously.
In response, Microsoft, Google, Meta, and Amazon have contracted, acquired, and financed over 10 gigawatts of new nuclear capacity in the past 12 months, with Microsoft’s Three Mile Island restart targeted for 2028, Google’s Kairos Power SMRs for 2030, and Meta’s 20-year deal with Constellation supporting the Clinton plant’s continued operation. Goldman Sachs has identified energy availability, not chip supply, as the primary constraint on AI infrastructure growth. For developing countries with grid reliability challenges, the AI energy requirement creates a structural barrier to domestic AI data centre development that policy ambition alone cannot overcome.
The Takeaway
The AI energy story is now the most consequential and least covered infrastructure story in technology. The chip shortage that dominated headlines in 2023 and 2024 has been partially resolved by supply chain expansion. The energy constraint that has replaced it cannot be resolved by supply chain decisions alone.
Building a nuclear power plant takes a decade. Building the transmission infrastructure to deliver that power reliably takes years beyond that. The technology companies that understood this earliest are now signing 20-year nuclear contracts that function as structural competitive advantages. The companies that did not are waiting for grid connections that utilities cannot approve because infrastructure capacity is already strained.
For Pakistan, the energy reality is the most honest constraint on the $1 billion AI fund’s ambitions. The $1 billion AI investment plan and the 2,000 MW allocation for AI data centres are both serious signals of intent. But intent without dedicated power infrastructure is a plan without an engine. The solar resources exist. The falling costs make co-located renewable power viable. The coordination between the AI fund and the power allocation needs to happen before procurement, not after. The technology is not the bottleneck. The electricity is.
