Big Tech’s AI Ambitions Are Colliding With the Power Grid
Meta’s nuclear deals, Microsoft’s power contracts, and Eric Schmidt’s warnings reveal how AI growth is becoming an energy and infrastructure problem - with global consequences.
Eric Schmidt, former Google CEO, has spent most of his professional life inside companies that scaled without ever having to worry about where their electricity came from. In recent weeks, that has changed.
In a public post earlier this year, Eric Schmidt wrote that artificial intelligence is no longer limited by software, talent, or even chips, but by land, power, and water. Data centres, he argued, are colliding with infrastructure limits that cannot be solved through faster iteration or better models.
Within days of Eric’s post, Meta confirmed it had signed long-term agreements with three nuclear energy companies, securing more than six gigawatts of future electricity capacity. The scale of the commitment was striking, but the reasoning behind it was straightforward: without guaranteed power, expansion plans stall.
The rapid expansion of AI is transforming data centres into a significant new electricity load, with global power demand from the sector projected to rise 500% by 2040, and about three-quarters of that demand still expected to be met by hydrocarbons, executives and policymakers heard during Abu Dhabi Sustainability Week.
Taken together, the two developments point to a reality that large technology companies have been grappling with privately for more than a year. Artificial intelligence, once sold as weightless software, is increasingly constrained by physical systems that move slowly, attract political scrutiny, and are expensive to build.
Jeff Graf, Global Head of Business Development at SandboxAQ, said the shift has been unavoidable. He described how internal assumptions inside the industry have broken down.
“For a long time, energy was treated as background noise,” Graf said. “You assumed the grid would be there, that utilities would figure it out, that capacity would show up when you needed it. At AI scale, that assumption stops working. You’re no longer talking about incremental load. You’re talking about industrial demand appearing very quickly in very specific places.”
When power becomes the organising principle
For most of the past decade, energy rarely appeared in technology strategy discussions. Data centres expanded gradually, utilities absorbed additional load, and sustainability goals were managed through renewable procurement and offsets that did not interfere with growth.
AI disrupted that balance; modern AI systems require persistent, high-density compute that runs around the clock. They concentrate demand geographically, often in regions where transmission infrastructure is already strained. In many markets, new grid connections now take years to approve, not months.
This shift explains why Google entered into a deeper partnership with NextEra Energy, linking its cloud expansion directly to new generation capacity. The aim was not just cleaner power, but predictability — a guarantee that future data centres would not be stranded by grid constraints.
Microsoft’s agreement with Constellation Energy follows the same logic, but with higher stakes. By supporting the restart of a nuclear unit at Three Mile Island, Microsoft is effectively securing a block of firm, always-on power rather than waiting for renewable build-outs and transmission upgrades that may not arrive in time.
Oracle’s reported exploration of data centres powered by small modular nuclear reactors reflects how far the conversation has moved. What once sounded extreme now appears pragmatic.
Graf said these decisions are reshaping how companies plan growth.
“Once power becomes the gating factor, it stops being a line item and starts becoming the organising principle,” he said. “Where you build, how fast you build, what customers you prioritise — all of that starts with energy availability.”
Why nuclear keeps coming back, despite the risks
Nuclear energy sits uncomfortably in the AI story. It is expensive, politically sensitive, and slow to permit. Yet it keeps resurfacing because it solves a problem few alternatives can handle at scale.
AI workloads demand reliability. Wind and solar are expanding rapidly, but intermittency requires storage, backup generation, and major grid reinforcement. In regions where transmission queues are already clogged, those dependencies become binding constraints.
An energy systems analyst who advises hyperscale operators, speaking in an interview with Bloomberg last year, described the appeal succinctly.
“If you are committing billions of dollars to data centres, the last thing you want is exposure to grid volatility. Nuclear gives you certainty, even if it comes with regulatory complexity.”
Meta’s decision to sign multiple nuclear agreements rather than back a single project reflects that logic. By spreading commitments across developers and technologies, the company is buying optionality in a market where delays are common.
Graf was clear about why efficiency alone will not close the gap.
“There’s this idea that better models will magically reduce energy demand,” he said. “That’s not how it plays out in practice. Every efficiency gain gets absorbed by more usage. Models get cheaper to run, new use cases open up, and total demand keeps rising.”
The economics are still unresolved
Behind the scramble for power lies a more fragile reality: the economics of AI are not yet settled.
Training and operating large models is expensive. While enterprise demand is growing, many AI products have yet to demonstrate stable, long-term margins. Locking in energy supply offers cost visibility, but it also ties companies to infrastructure commitments that last decades.
According to analysis published by the International Energy Agency, electricity demand from data centres could more than double in parts of the world by the end of the decade, raising the risk of overbuild if AI adoption does not match current projections.
Graf said this tension is now front and centre in executive discussions.
“The hardest conversations aren’t about whether the technology works,” he said. “They’re about whether the economics still make sense once you account for power, cooling, uptime guarantees, and the fact that outages are no longer tolerable. That’s where a lot of business cases get thinner.”
Some analysts view the current wave of energy deals less as confidence and more as insurance. Securing power early reduces downside risk in a constrained market, even if it increases long-term exposure.
The Gulf’s structural advantage — and its limits
While power constraints are tightening in the US and Europe, the Gulf is approaching the AI–energy equation from a different starting point.
Countries such as Saudi Arabia and the UAE operate large, state-backed energy systems with surplus generation capacity and long planning horizons. That has made the region increasingly attractive to data centre operators and AI infrastructure investors.
Saudi Arabia’s Aramco has been clear about its ambitions to move beyond hydrocarbons into data, AI, and advanced computing. In public remarks at the LEAP technology conference and in investor briefings, Aramco executives have described data centres and AI workloads as “strategic demand centres” aligned with the kingdom’s industrial base.
Aramco has already signed partnerships around cloud infrastructure, digital twins, and AI-driven optimisation, backed by access to reliable power and state-supported infrastructure.
A Gulf energy policy adviser, speaking to the Financial Times last year, described the region’s positioning without ambiguity.
“We built energy systems for heavy industry and export. Large-scale data centres fit naturally into that model, especially when other regions are struggling with grid constraints.”
Similarly, ADNOC has positioned AI and high-performance computing as core to its digital transformation, supported by dedicated energy capacity and sovereign investment.
The social and political edge
As AI infrastructure expands, its impact is becoming visible beyond corporate balance sheets.
Grid operators in the US and Europe have warned that large, concentrated data centre loads can delay connections for other users and push up electricity prices. Communities hosting these facilities are increasingly vocal about land use, water consumption, and public subsidies.
A senior US utility executive told regulators last year that clusters of AI data centres behave more like heavy industry than traditional commercial customers, with implications for grid stability.
“People don’t protest algorithms,” Graf said. “They protest higher bills and unreliable service. That’s where the backlash will come from if this isn’t managed carefully.”
A heavier future for AI
There is a positive case for what is unfolding. AI-driven demand could accelerate investment in clean energy, revive nuclear projects, and modernise ageing grids. Long-term contracts can make infrastructure financeable in previously difficult ways.
But the deeper truth is harder to ignore. AI is no longer weightless. It is bound to the same physical, political, and economic constraints that shape every industrial system.
Schmidt’s focus on energy reflects that recognition. The next phase of AI will not be decided solely by who builds the best models, but by who can secure power without overwhelming the systems that support them.
Graf summed up the moment in practical terms.
“The bottleneck isn’t intelligence anymore,” he said. “It’s construction, permits, and power. That’s the race people are actually running, whether they admit it or not.”
The end of the easy phase
What is unfolding now is not a temporary adjustment. It is the end of the easy phase of AI.
For more than a decade, the technology industry grew inside a system that absorbed its demands without pushing back. Power was cheap enough, grids were taken for granted, and growth felt abstract. AI has changed that balance. It is forcing technology companies to confront limits they did not set and do not control.
Energy deals, nuclear contracts, land purchases, and infrastructure partnerships are not signs of confidence alone. They are signs of exposure. They reveal how dependent AI has become on systems that move slowly, invite political scrutiny, and redistribute costs unevenly.
In regions like the Gulf, where energy systems are centralised and capital is patient, AI infrastructure fits more naturally into existing industrial strategies. In the US and Europe, where grids are fragmented and permitting is contested, the same expansion creates friction — between companies and regulators, between data centres and households, between ambition and capacity.
This is the reality behind the headlines. AI is no longer just a question of innovation. It is a question of allocation: who gets power, at what price, and with whose consent.
The companies that succeed in the next phase will not simply be the ones with better models or more data. They will be the ones who navigate energy politics, infrastructure bottlenecks, and public tolerance without triggering backlash or breakdown.
That is a very different kind of competition from the one Silicon Valley is used to. It is slower, heavier, and harder to narrate.
But it is the one that now matters.