Half of U.S. Data Centers Are Being Delayed or Canceled. AI's Biggest Bottleneck Isn't Software.

Alphabet, Amazon, Meta, and Microsoft plan to spend over $650 billion in 2026 expanding AI compute capacity. They've got the money. They've got the chips on order. They've got the land.

What they don't have is enough electricity.

Nearly half of the 140 planned U.S. data center projects slated for 2026 are now facing delays or outright cancellation. Not because of software problems or chip shortages. Because the physical power grid can't keep up. The AI industry is building the future, and it's tripping over the electrical wiring.

The Numbers Are Brutal Across those 140 projects, data centers representing at least 16 gigawatts of capacity were supposed to come online before the end of 2026. For context, 16 gigawatts could power roughly 12 million homes. But only about 5 gigawatts are actually under construction right now, and typical build times run 12 to 18 months. The math doesn't work.

Analysts expect 30-50% of 2026 projects to be delayed. The bottlenecks are unglamorous but very real: transformers, switchgear, batteries, and the grid infrastructure needed to deliver megawatts of power to a single building. These aren't components you can 3D print or overnight from Amazon. High-power transformers have lead times measured in years, not months.

And here's the part that makes U.S. policymakers nervous: imports of high-power transformers from China surged from fewer than 1,500 units in 2022 to more than 8,000 in 2025. The AI infrastructure buildout is heavily dependent on Chinese-made electrical equipment, right at a time when trade tensions and export controls are escalating. About 25% of planned projects haven't even disclosed how they plan to get enough power.

Why Software Companies Are Hitting a Hardware Wall There's something almost ironic about the situation. The companies building the most advanced software in history are being bottlenecked by century-old infrastructure -- power lines, transformers, substations. It's like building a spaceship and realizing nobody paved the road to the launch pad.

The AI industry's power appetite has caught basically everyone off guard. A single large AI data center can draw 100+ megawatts. A cluster of them in one region can strain the entire local grid. Utilities that were planning for steady, predictable demand growth are suddenly getting requests for the equivalent of powering a small city, and they need it built yesterday.

We covered Anthropic's deal with Broadcom and Google for 3.5 gigawatts of compute recently. That single deal represents more power demand than most mid-size cities consume. And Anthropic is just one company. Multiply that across every major AI lab, cloud provider, and enterprise building out AI infrastructure, and you start to understand why the grid is buckling.

The Bright Side: ASML Says Demand Isn't Slowing Despite the construction delays, the underlying demand for AI compute is only growing. ASML, the Dutch company that makes the machines that manufacture advanced AI chips, just raised its 2026 revenue outlook to 36-40 billion euros after beating earnings estimates. CEO Christophe Fouquet said chip demand is outpacing supply, and customers are fast-tracking expansion plans.

ASML can ship at least 60 of its extreme ultraviolet lithography systems in 2026 and 80 in 2027, but even that won't be enough. The company says supply won't meet demand "in the foreseeable future."

So the chips are being made. The demand is there. The software is ready. It's the physical infrastructure -- power, cooling, construction -- that can't keep pace. The AI industry has a last-mile problem, and it's measured in megawatts, not milliseconds.

What This Means If You're Not Building Data Centers You might be wondering why any of this matters to you if you're running a normal business, not a hyperscale cloud provider. A few reasons:

Cloud computing costs could rise. When supply is constrained and demand keeps growing, prices tend to go up. If the data center buildout falls behind schedule, expect compute costs to stay elevated for longer than anyone projected. That filters down to every business paying for cloud services, AI tools, and SaaS platforms.

The AI tools available today are built on infrastructure that already exists. This is actually the good news. You don't need to wait for new data centers to be built to start using AI for your business. The models, platforms, and tools available right now -- running on existing infrastructure -- are already capable enough to transform how you handle customer support, lead qualification, and sales.

While big tech fights over gigawatts and transformer supply chains, the AI chatbot tools you can deploy today work just fine on current infrastructure. The power shortage is a problem for companies training the next generation of models, not for businesses using the current ones.

Location will matter more. Data center construction is shifting toward regions with available power -- think nuclear plants, hydroelectric access, or states with friendly energy policies. That geographic reshuffling will affect which cloud regions offer the best pricing and latency. If your business depends on low-latency AI services, keep an eye on where your cloud provider is building.

The Uncomfortable Question The AI industry has been telling a story about exponential growth for three years now. Bigger models, more compute, faster everything. The data center bottleneck forces a question nobody wants to ask: what happens if the physical world can't keep up with the digital one?

Maybe the answer is more efficient models that need less compute. Maybe it's nuclear power and next-gen grid infrastructure. Maybe it's a slowdown that nobody in Silicon Valley wants to admit is possible.

For now, the smart move is the same one it's been all along: use the AI that's available today, don't wait for tomorrow's infrastructure to be built. The businesses getting value from AI right now aren't the ones waiting for 16 gigawatts of new capacity. They're the ones that started with a chatbot and worked from there.