For two years, the story of artificial intelligence was a story about chips. Whoever held the most GPUs held the advantage. That race is not over, yet a larger constraint has moved into view. The input that increasingly decides who can build AI, and where, is electricity. You can buy a chip in months. You cannot build a power plant or a grid connection in months. As a result, energy is overtaking hardware as the defining input of the AI era.
For Two Years, the Bottleneck Was Silicon
The early phase of the AI boom was defined by scarcity of one thing: advanced GPUs. Demand outran supply, lead times stretched, and access to chips became the clearest measure of who could compete. Companies signed multi-year supply deals and built strategies around the hardware they could secure. That framing made sense at the time, and it still shapes how many people think about the AI race.
Supply has since improved at the chip level, even as demand keeps climbing. More importantly, a different limit has come into focus. Every one of those chips needs power, and a lot of it. The question has quietly shifted from “how many GPUs can you buy” to “where will you find the electricity to run them.” That second question turns out to be far harder to answer.
Now the Bottleneck Is Power
The scale of demand is the first thing to understand. The International Energy Agency projects that electricity used by data centres will roughly double, from 485 terawatt-hours in 2025 to about 950 terawatt-hours in 2030. For comparison, that 2030 figure is close to the entire electricity consumption of Japan. Demand from AI-focused data centres grows faster still, tripling over the same period.
The growth is already visible in the numbers. In 2025, overall data centre electricity use grew 17%, while electricity use by AI-focused data centres surged 50%. The money confirms the trend. According to the IEA, capital expenditure by the largest technology companies exceeded 400 billion dollars in 2025. The capital spending of just five of them is now larger than global investment in oil and natural gas production. We covered the broader buildout in Why the $6.7 Trillion Data Center Buildout Still Will Not Be Enough.
Here is the catch. Spending money is fast, and building a chip factory is faster than building an electricity system. Capital can flow in a quarter. A new power plant, a high-voltage line, or a substation takes years. The result is a widening gap between how quickly AI demand grows and how slowly the systems that power it can respond.
Why Power Is Harder to Buy Than Chips
To connect new generation to the grid, a project must pass a series of interconnection studies that assess its impact and assign the cost of any upgrades. These studies have become a bottleneck of their own. According to Lawrence Berkeley National Laboratory, the median time from interconnection request to commercial operation has doubled. It was under two years for plants built in the early 2000s, and it now exceeds four years for those built recently.
The queue itself tells the story. By the end of 2025, more than 2,060 gigawatts of generation and storage capacity were waiting to connect in the United States, which is more than the country’s entire installed power capacity. Most of that will never be built. Only about 13% of projects that requested connection between 2000 and 2019 had reached operation by the end of 2024. For a data centre developer, this means the electricity needed for a new site may be theoretically available but practically years away.
Equipment shortages compound the delay. Orders for gas turbines surged 70% in 2025, straining a supply chain that cannot expand overnight. Transformers and power electronics face similar pressure. Because grid connections are slow, some United States developers are now building their own onsite gas generation. The IEA estimates that 15 to 27 gigawatts of onsite gas could power data centres by 2030. That is a workaround, not a solution, and it underlines the central point: power has become the rate-limiting step.
The Density Problem Nobody Planned For
AI does not just need more power overall. It needs an extraordinary amount of power in a very small space. The IEA notes that a single AI server rack, about the size of a household refrigerator, could by 2027 draw peak power equivalent to 65 households. Between 2020 and 2025, the power density of AI servers increased elevenfold, and a further fourfold rise is expected by 2027.
This density changes what a data centre has to be. Older facilities were not designed to deliver or cool that much power per square metre, so a large share of existing capacity simply cannot host modern AI hardware. The constraint is therefore not only the total electricity on the grid, but whether a specific site can deliver dense, reliable, high-quality power exactly where the racks sit. We explored these physical limits in The Power Problem: Data Center Energy Constraints.
What This Means for Anyone Building AI
The strategic lesson is straightforward. In a market where power is the binding constraint, the advantage belongs to whoever can reach electricity that already exists. A team that has to wait four years for a grid connection is four years behind a team that does not. Access to live, energised capacity is becoming as valuable as access to the chips themselves.
This is the logic behind the asset-light model. Rather than building new data centres and waiting in the grid queue, Axe Compute deploys on infrastructure across 200+ locations that is already installed and powered. Clients reach capacity in roughly 48 hours instead of years, because the hard part, the power, is already in place. We set out the economics of this approach in Neoclouds and Why the Asset-Light Model Wins.
None of this means hardware stops mattering. The right GPU for the workload still shapes cost and performance. The point is that the question has expanded. It is no longer enough to ask which chip to run. The decisive question now is where the power to run it already exists, and how quickly you can reach it. For a deeper look at planning across both, see Enterprise GPU Strategy in 2026.
Some companies treat energy as the first input, rather than an afterthought to the chip order. Those are the ones that will still be building when the grid queue stretches past the end of the decade. Energy is no longer the background condition of the AI economy. It is the defining one.
About Axe Compute
Axe Compute Inc. (NASDAQ: AGPU) is a neocloud AI infrastructure platform built on a fundamental premise: AI innovation should not be constrained by hardware choice or inventory limitations. Axe Compute gives enterprises and AI innovators choice across hardware, geography, and deployment speed through two delivery models: Axe Compute Access, providing the latest GPU compute options in as fast as 48 hours across numerous global locations, and Axe Compute Build, enabling enterprises to access large-scale dedicated AI factories, all backed by enterprise-grade SLAs and support. Axe Compute is headquartered in Pittsburgh, Pennsylvania. For more information, visit axecompute.com.
Reach power that already exists.
400,000+ GPUs · 200+ locations · 48-hour provisioning · Zero egress fees
Frequently Asked Questions
Is energy really becoming a bigger constraint than chips for AI?
Yes, for new capacity. A company can order GPUs and receive them within months, but securing the electricity to run them can take years. The International Energy Agency projects data centre electricity demand roughly doubling from 485 terawatt-hours in 2025 to about 950 terawatt-hours in 2030. Meanwhile, more than 2,060 gigawatts of power generation sits in United States interconnection queues, and the median wait to connect new supply has passed four years. Hardware is fast to buy. Power is slow to build. Axe Compute sidesteps this by deploying on infrastructure that is already energised.
How much electricity will AI data centres use?
The IEA projects total data centre electricity demand rising from 485 terawatt-hours in 2025 to roughly 950 terawatt-hours in 2030, around 3% of global electricity. Demand from AI-focused data centres grows faster still, tripling over the period. In 2025 alone, overall data centre electricity use grew 17%, while AI-focused demand surged 50%.
Why does it take so long to connect a data centre to the grid?
New power plants and large loads must pass interconnection studies that assess their impact on the grid and assign the cost of upgrades. According to Lawrence Berkeley National Laboratory, the median time from interconnection request to commercial operation has doubled from under two years to more than four years. Over 2,060 gigawatts of capacity is waiting in United States queues, and most of it will never be built. Transformer and gas-turbine shortages add further delay.
How does Axe Compute avoid the power bottleneck?
Axe Compute operates across 200+ locations in 93 countries on 400,000+ GPUs that are already installed and powered. Because the capacity is live rather than planned, clients do not wait years for a new grid connection. Axe provisions capacity in approximately 48 hours, with zero egress fees and 99% uptime, and pricing significantly below hyperscaler rates.
Sources
- IEA, “Key Questions on Energy and AI” (2026): data centre electricity 485 TWh (2025) to ~950 TWh (2030); 17% overall and 50% AI-focused growth in 2025; hyperscaler capex above $400B in 2025 and larger than global oil and gas investment; rack power density and onsite gas estimates
- IEA, “Energy and AI” news release: AI as the most significant driver of data centre demand growth
- Lawrence Berkeley National Laboratory, “Queued Up” (2026): over 2,060 GW in US interconnection queues end-2025; median interconnection-to-operation time doubled to over four years; ~13% of 2000-2019 requests reached operation
- Tom’s Hardware: Big Tech 2026 capex plans reaching ~$725 billion