The AI Build-Out Has Flipped the Breakers: Why Data Centres Are Stalling in 2026

Nearly half of all planned US data centres in 2026 have been delayed or cancelled. Not from lack of capital. Not from lack of demand. From transformers.

This is the defining infrastructure story of 2026, and most people are still not paying close enough attention to it.

The AI boom unleashed an unprecedented wave of data centre construction. Hyperscalers committed hundreds of billions of dollars. Governments declared it a national priority. Chip manufacturers scaled production. And yet, at the end of 2025, data centres requiring 241 gigawatts of electricity were sitting in the pipeline, representing a 159% increase from the start of that year alone. The bottleneck was never the chips. It was never the capital. It was the grid.

High-voltage transformers, the large steel and copper units that step down electricity from transmission lines into usable power for a facility, take between one and three years to manufacture. The global supply chain for these components is severely constrained, with lead times stretching to 80 weeks or longer in some markets. A significant portion of the components come from China, and tariff uncertainty has compounded the problem.

Beyond transformers, the issue runs deeper. The US grid interconnection queue, the formal process by which a new facility applies to connect to the power network, has become a multi-year backlog. In some regions, projects are waiting four to five years simply to receive permission to connect. Gartner has predicted that by 2027, 40% of AI data centres will face operational constraints specifically due to power availability. Goldman Sachs projects that data centres will consume 8% of all US electricity by 2030, growing at a 15% compound annual rate.

This is not a temporary supply chain hiccup. It is a structural misalignment between the pace of AI ambition and the pace at which physical energy infrastructure can be built.

The smartest operators in this space have already drawn the obvious conclusion: in this cycle, whoever controls the power controls the data centre. Google, together with Intersect Power and TPG, announced a $20 billion commitment to develop dedicated clean energy projects to power its future data centre builds, bypassing the public grid entirely. Microsoft has moved to restart nuclear capacity at Three Mile Island. Amazon is investing in small modular reactors. The pivot is unmistakable. The technology companies that move fastest are no longer competing for chips. They are competing for megawatts.

Geography is being redrawn as a result. The historically dominant data centre markets, Northern Virginia, Silicon Valley, and Dublin, are effectively closed to new large-scale development due to grid saturation. Attention is shifting to regions with surplus power capacity: parts of the US Midwest and South, the Middle East, and select markets in Southeast Asia where governments are proactively building energy infrastructure in anticipation of demand.

The Middle East deserves specific mention. The UAE and Saudi Arabia have both made explicit national commitments to AI infrastructure, and crucially, they are building the power generation capacity to support it. Abu Dhabi and Riyadh are not waiting for private grid operators to upgrade infrastructure. They are directing sovereign capital at the problem. For operators and investors who understand geography, this is a meaningful signal.

There are several layers to this story that informed observers should be tracking.

First, the energy infrastructure companies that service data centres are now more strategically important than many of the data centre operators themselves. The manufacturers of transformers, switchgear, and grid-scale power management equipment are experiencing demand that their production capacity cannot meet. This is a multi-year constraint, not a short cycle.

Second, nuclear is returning to credibility at pace. The combination of zero-carbon credentials, baseload reliability, and the sheer scale of power required has made nuclear the only technology that can realistically serve a hyperscale campus at the density AI workloads demand. Small modular reactors, while still largely pre-commercial, are receiving serious investment precisely because the alternative, relying on a strained grid, is becoming operationally untenable.

Third, the companies building AI applications are starting to face a hard commercial reality. If your data centre cannot be powered, your model cannot be trained or served at scale. This creates a meaningful competitive advantage for operators who secured power access early, and a meaningful disadvantage for those who assumed supply would follow demand.

Fourth, the geographic diversification of data centre investment has strategic implications beyond pure infrastructure. Countries that can credibly offer reliable, affordable, large-scale power to international operators are positioning themselves as genuine AI hubs. This is not just an energy story. It is a geopolitical story about which nations will host the physical infrastructure that runs the global digital economy.

The AI build-out is real. The demand is real. The capital is real. But none of it moves without power. And right now, power is the one thing that cannot be conjured quickly.

For anyone thinking seriously about where the structural opportunities sit in this cycle, the answer is increasingly not in the software layer. It is in the physical infrastructure that makes the software layer possible. Energy, grid technology, cooling systems, and the civil engineering of large-scale facility development are where the genuine constraints exist, and where the genuine leverage lies.