The promise of unlimited AI compute is running headfirst into a very physical constraint: the electrical grid.

According to a new analysis reported by Tom’s Hardware, close to half of all planned US data center builds for 2026 are projected to be delayed or outright cancelled. Not because capital dried up. Not because demand evaporated. Because there is not enough power, and the components needed to deliver it have lead times stretching up to five years.

This is a story that goes beyond hyperscaler infrastructure headaches. It has direct implications for every business that has built or is building AI strategies on the assumption that compute will keep getting cheaper and more available.

The Numbers Are Stark

Of the roughly 140 large-scale data center projects, representing approximately 12 gigawatts of power capacity, planned to come online in the US in 2026:

• Only a third are currently under construction
• The remaining two-thirds are stuck in pre-production and considered unlikely to open on schedule
• For 2027, just 6.3 gigawatts worth of projects are under construction against 21.5 gigawatts already announced

That gap between announced and actually-building represents hundreds of billions of dollars of AI capacity that exists only on paper.

What Is Actually Causing the Delays

The instinct is to blame permitting, NIMBY opposition, or financing. The reality is more mundane and more difficult to solve: transformers, switchgear, and grid connections.

These are the physical components that get power from the utility grid into a data center building. They are not exotic technology. They are not new inventions. But demand for them has exploded alongside the AI infrastructure buildout, and the manufacturing capacity to produce them simply has not kept pace.

Lead times for large power transformers now stretch to four or five years in some cases. In a sector where deployment cycles run under 18 months, that mismatch is catastrophic for project timelines.

The problem compounds when you add trade restrictions into the mix. Chinese electrical equipment, which has historically been a significant part of the global supply for this kind of hardware, is now constrained by tariffs and trade friction. Companies that planned around Chinese sourcing are scrambling to find alternatives, often at higher cost and longer timelines.

Why Big Tech Cannot Simply Spend Its Way Out

Alphabet, Amazon, Meta, and Microsoft collectively plan to spend more than $650 billion in 2026 expanding AI infrastructure. That level of investment is unprecedented. But money alone cannot compress five-year transformer lead times into 18 months.

Projects involving some of the largest names in AI are expected to miss projected deadlines by more than three months, according to reporting citing satellite imagery and project tracking data. Companies are not publicly acknowledging the delays, but the evidence on the ground tells a different story.

The scale of the ambition creates its own problem. When dozens of hyperscaler projects all hit the market for the same components at the same time, suppliers get overwhelmed, prices increase, and timelines stretch further. The competitive race to build has created a collective action problem no single player can solve unilaterally.

What This Means for Business

If you are a business leader building AI strategy today, this story is worth understanding clearly.

AI services will remain more expensive than the hype suggests. The assumption that compute costs would continue falling rapidly relies on massive new supply coming online. If half of that planned supply is delayed or cancelled, pricing for AI inference and training stays elevated for longer. Factor that into your AI investment plans.

The hyperscalers will prioritise their own needs first. When capacity is constrained, cloud providers allocate it strategically. Enterprise customers on standard contracts may find capacity limitations showing up in ways they did not anticipate, particularly for GPU-intensive workloads.

This creates real advantage for businesses that act now. If you secure AI capabilities, build internal workflows around them, and train your team on AI tools while supply is constrained, you create a structural lead over competitors who are waiting for costs to drop further. The companies already generating returns from AI today will have more resources, more data, and more institutional knowledge by the time the infrastructure situation resolves.

On-premises and hybrid AI strategies look more attractive. For some workloads, the volatility of cloud AI pricing and availability makes a case for deploying smaller models in-house. This is not right for every business, but it is worth having in your thinking.

The Bigger Picture

The AI infrastructure crisis is a sign of genuine, enormous demand rather than a sign that AI is slowing down. The world is consuming AI capacity faster than the physical world can build it. That is, in an odd way, a validation of the technology.

But it is also a reminder that transformative technology runs on very unromantic foundations. Copper wire. Steel cabinets. Grid interconnection agreements. Power purchase contracts. The exciting AI demos run on hardware that has to be physically built, in buildings that have to be connected to grids that have to be expanded.

For businesses, the practical lesson is this: the AI advantage window is real, but it is not unlimited. Infrastructure constraints will resolve eventually. Companies that have already built AI fluency, established working AI workflows, and accumulated proprietary data and experience will have a compounding advantage over those who waited.

The grid will catch up. The question is whether your competitors will have lapped you by the time it does.

---

If you’re deciding where to start with agents, start here. The free Working With Claude field guide walks through the ecosystem, Claude Code, and a real rollout plan. Get your copy.