In the race for artificial intelligence dominance, American tech giants have most of the pieces in place—money, chips, data centres—but one new and growing obstacle now threatens to slow them down: electric power.
The new bottleneck: compute and power
In a recent podcast conversation between Satya Nadella (CEO of Microsoft) and Sam Altman (CEO of OpenAI), Nadella was blunt: “The biggest issue we are now having is not a compute glut, but it’s the power and...the ability to get the builds done fast enough close to power.”
He added, “So if you can’t do that, you may actually have a bunch of chips sitting in inventory that I can’t plug in.”
This perfectly captures the shifting question—for AI infrastructure today the question isn’t just “Do we have enough chips?” but increasingly “Do we have enough power to use them?”
Why the scale is massive
Major tech firms such as Google LLC, Microsoft, Amazon Web Services (AWS) and Meta Platforms are pouring unprecedented billions into AI infrastructure growth—an estimated ~$400 billion in 2025 alone, with even more expected in 2026. These funds are going into building enormous data-centres, in-house processor production (to chase Nvidia Corporation) and the enormous supporting infrastructure.
But data centres aren’t just building blocks; they are power-intensive factories of computing at scale.
Key figures showing the strain
-
A recent Deloitte report emphasises that AI-driven hyperscale data centres create “large, concentrated clusters of 24/7 power demand”.
-
The International Energy Agency (IEA) projects data-centre electricity demand worldwide will more than double by 2030 to around 945 TWh. In the U.S., data centres are on course to account for almost half the growth in electricity demand between now and 2030.
-
According to BloombergNEF, U.S. data-centre power demand is forecasted to rise from almost 35 GW in 2024 to 78 GW by 2035.
-
A Pew Research short-read reports data centres consumed about 183 TWh of electricity in the U.S. in 2024 (just over 4% of U.S. electricity use). By 2030 this could rise by +133% to ~426 TWh.
These numbers reveal a stark reality: the AI industry can acquire chips and build racks—but if the power isn’t there (or reliably there), expansion stalls.
The “energy wall”
Power-grid upgrades take years
Building a large data centre in the U.S. typically takes ~2 years. By contrast, bringing new high-voltage power lines into service can take 5 to 10 years. Therein lies a mismatch: compute builds ramp up fast; grid upgrades lag far behind.
Geographical clustering
Large AI/data-centre expansions are concentrated in certain regions (for example, Virginia, Texas). Utilities are seeing massive demand. For instance, one Virginia utility reported a data-centre order book of 40 GW (equivalent to 40 nuclear reactors) and later 47 GW. Such concentrated demand stresses the local grid infrastructure.
Who pays for upgrades?
Utilities, regulators and tech companies are facing thorny questions: when a new data centre requests huge amounts of power, who picks up the cost of transmission lines, sub-stations, local distribution upgrades? Some states are seeing push-back from households and small businesses who fear paying higher rates for upgrades driven by hyperscale data-centre load.
Fossil fuels, nuclear and the climate compromise
Turning back to coal & gas
With grid upgrades slow and renewables limited by intermittency or siting, some U.S. utilities are delaying coal-plant retirements. Natural-gas generators, often faster to deploy, are gaining renewed favour: around 40% of data-centre power worldwide currently comes from natural gas.
In one U.S. state (Georgia), a utility requested authorization to install 10 GW of gas-powered generators to meet data-centre demand.
Nuclear and renewables enter the pitch
Some tech firms are quietly shifting climate pledges in favour of power reliability. For instance:
- Google removed a “net-zero 2030” emissions pledge from its website mid-2025.
- Amazon is supporting a revival of Small Modular Reactors (SMRs).
- Some big firms are signing long-term nuclear power-purchase or supply-agreements.
- At the same time, investment in solar + battery storage in key states (Texas, California) is ramping to meet the power load of data centres.
The trade-off
The challenge: accelerate power supply at a scale and pace matching AI demand without derailing climate goals. So far the urgency of compute/AI seems to be winning. In effect: Power reliability → AI growth, even if the clean-energy transition slows.
The risk: falling behind in the AI race
For the U.S., the implications are multi-fold:
- If tech firms cannot secure reliable power near their data-centres, they may shift investments or capacity overseas (to regions with more available grid capacity or cheaper power).
- Infrastructure bottlenecks (power, cooling, transmission) may slow rollout of latest-generation AI models or delay deployment of supercomputing clusters—thus eroding the U.S. competitive lead.
- Rising energy costs or grid stress may trigger regulatory or community backlash, complicating further expansion.
- Climate trade-offs may undermine corporate and national climate credibility—opening reputational and regulatory risk.
What needs to be done
- Faster grid investment & policy reform — accelerate permitting, transmission build-out, upgrade of local distribution networks.
- Power-demand management — data centres and utilities must collaborate on variable workloads, demand-response, and scheduling compute to smooth peaks. Research shows AI-HPC data centres can provide grid flexibility at relatively lower cost.
- Diversified energy mix — pairing renewable generation (solar/wind) + storage + onsite generation (fuel cells, SMRs) to deliver reliable power. E.g., fuel-cells and hydrogen turbines are being considered for data-centre loads.
- Transparency & community fairness — ensure that rate-payers, utilities, and tech firms coordinate so growth of data centres doesn’t unfairly burden local consumers.
- Location strategy — distribute data-centres more geographically, locating them where power, transmission and cooling resources are abundant, to avoid “all eggs in one grid basket”.
The AI revolution isn’t just about smarter algorithms or fatter data-centres. It’s increasingly about electricity infrastructure. The U.S. tech giants may hold the chips, capital and plans — but without sufficient, reliable power, the ambition could stall. If this power bottleneck isn’t addressed, the U.S. risks losing ground in the global AI race, ceding leverage to countries where infrastructure can keep pace.
In short: The next frontier of AI dominance may depend not just on compute, but on watts.