For most of the last two years, the story of artificial intelligence was a story about chips. Who could secure GPU allocation, who was at the front of Nvidia’s line, who had the compute to train the next frontier model. That scarcity defined strategy, valuations, and the pecking order of the entire industry.

That story has quietly inverted. The constraint that now decides which AI data centers actually get built is not silicon. It is the unglamorous electrical hardware that delivers power to the rack: transformers, switchgear, and the grid connections behind them. Capital is fully committed. Compute is available. What is missing is the physical layer that turns billions of dollars of investment into energized megawatts.

At Acuvity Consulting, California’s boutique Big 4 alternative in business and technology strategy, we think this shift matters far beyond the data center industry. It carries a lesson that sits at the center of how we advise clients across healthcare, manufacturing, technology, and the public sector.

The numbers tell a stark story

The scale of the disconnect is hard to overstate. BloombergNEF estimates that the capital expenditure of the 14 largest publicly owned data center operators will approach 750 billion dollars in 2026, up from a little under 450 billion the year before. The money is there. The buildout is not keeping pace.

The gap between announcement and construction has become the defining metric of the cycle:

• Roughly half the pipeline is stalling. Of the announced U.S. data center capacity for 2026, only about one third is under active construction, with 30 to 50 percent of the pipeline expected to be delayed or canceled.
• The deferred value runs into the tens of billions. Analysts have framed the shortfall as roughly 30 to 70 large AI training campuses that were promised but cannot be delivered on schedule, each representing 1 to 4 billion dollars in capex.
• Demand that capital cannot unlock. Microsoft has described an 80 billion dollar Azure backlog driven largely by power availability rather than weak demand. Customers are ready to buy capacity that cannot be switched on because the electricity to run it is not yet connected.

The reason is not money, and it is not GPUs. It is lead time on power equipment.

Transformers are the new GPUs

The single clearest signal of where the constraint has moved is the humble power transformer. The numbers are striking:

• Lead times have roughly doubled. Wood Mackenzie analysis puts power transformer lead times around 128 weeks, with generator step-up units at 144 weeks. In severe cases, delivery can stretch to as long as five years, against AI data center deployment cycles that run under 18 months.
• Prices have surged. Transformer prices are up 70 to 150 percent since 2020, with demand for substation transformers up 116 percent since 2019 and generator step-up units up 274 percent.
• A small cost with outsized power. Transformers and switchgear represent a small share of total project cost but sit on the critical path. A delay in any single element of the power chain can halt an entire campus.

The pressure is structural rather than temporary. The same categories of long-lead equipment are being chased simultaneously by data center developers, utilities running their own grid upgrades, renewable interconnection projects, and industrial electrification. Everyone is competing for the same constrained pool. And the demand curve is steepening: NEMA’s Grid Reliability Study projects that data center electricity consumption will rise roughly 300 percent over the next decade and account for 38 percent of net U.S. electricity demand growth through 2037.

This is no longer a procurement footnote. It is the first go or no-go question in any serious AI infrastructure development plan.

The market is already adapting in revealing ways

When a constraint becomes this binding, behavior changes fast, and the workarounds tell you where the real value is moving. Three responses stand out.

1. Bypassing the grid entirely. Rather than wait years for a utility connection, developers are building their own power on site. One comprehensive analysis identified 46 data centers with a combined 56 gigawatts of capacity planning to generate power “behind the meter,” roughly 30 percent of all planned U.S. capacity, with 90 percent of those projects announced in 2025 alone. A strategy that was a curiosity a year ago is now mainstream. Meta’s Hyperion campus in Louisiana, set to come online by 2028, will be powered by three on-site natural gas plants costing a combined 3 billion dollars.

2. Paying almost any price for speed. The economics explain the urgency. An AI data center can earn as much as 10 to 12 billion dollars per gigawatt, so getting online a few years early can produce a windfall. Efficiency is out; speed to power is the only metric that matters. That logic is pushing companies into genuinely unusual decisions, including one developer placing a 1.25 billion dollar power generation order with a supersonic aircraft startup that had never sold a power product, and others strapping gas turbines to semi trucks to get capacity online.

3. Buying the power, not just the building. The winning asset in a constrained market is increasingly land plus permits plus a credible path to electricity. Alphabet’s 4.75 billion dollar acquisition of clean energy developer Intersect was explicitly about bringing data center load and power generation into closer lockstep. Logistics platforms with large land positions and established utility relationships now command a premium, because location decisions that used to turn on connectivity and cost now turn on speed to power.

The Acuvity view: the advantage has moved to sequencing and ecosystem position

Here is what we take from all of this, and why it shapes the way we work with clients.

When capital is abundant and compute is available, neither one is a differentiator anymore. The companies winning this cycle are not the ones announcing the largest dollar figures. They are the ones who understood the physical constraint early, locked in long-lead equipment ahead of the queue, and positioned themselves around the parts of the value chain that everyone else overlooked.

That is an execution advantage, not a capital advantage. It is also exactly where business strategy and technology strategy meet, the intersection Acuvity Consulting is built to bridge. And it rewards two things in particular.

Sequencing discipline. Lead time has effectively become a currency. Developers who secure equipment earliest are buying schedule certainty, and schedule certainty is buying market share. The teams that treat procurement timing as a core strategic decision, rather than back-office paperwork, are the ones still building while competitors stall.

Ecosystem positioning. The binding constraint has shifted off the headline component, the GPU, and onto a long tail of less visible suppliers in the AI data center hardware ecosystem:

• Transformer and switchgear manufacturers, where domestic players like Eaton, Vertiv, and Schneider Electric remain capacity-constrained and years away from absorbing the gap.
• High-speed gas generation and on-site power integrators, now central rather than peripheral to a project’s timeline.
• Cooling systems, batteries, and the firms that can compress the time between capital and energized capacity.

The independent hardware players who can move faster than the incumbents are suddenly indispensable. Understanding where a company sits in that ecosystem, and where the genuine bottlenecks and opportunities are, is now a strategic question rather than a technical one. It is the kind of go-to-market and market-entry question we help manufacturing and technology clients answer.

This is the work we find ourselves doing repeatedly: mapping an ecosystem to find where the real leverage sits, helping a client see that the binding constraint is not the obvious one, and separating the headline narrative from the part of the value chain that will actually determine who wins. The AI buildout is simply the largest and clearest example of a pattern we see across industries. Strategy is only as good as the execution layer beneath it, and the execution layer is usually constrained by something nobody put on the slide.

What this means going forward

There is a healthy debate about whether the buildout is overbuilt. One camp points to record demand and historically low vacancy, arguing the constraint is delivery rather than appetite, with CBRE framing 2026 as defined by execution risk rather than demand uncertainty. Others warn that a meaningful share of planned capacity may be delayed or never built and that the industry may be running ahead of demand. Both can be true at once. The capital is committed regardless, and the constraint is real regardless.

What is not in doubt is the lesson for anyone making large, irreversible bets in a capacity-constrained market:

• Money does not move the constraint. Foresight does. The organizations that lead the next phase will be the ones who looked past the obvious scarcity, found the real one early, and built their strategy around it.
• The bottleneck always moves. It was GPUs. Now it is transformers and power. Wherever it lands next, the winners will be the teams watching the execution layer, not just the headline.

That is the difference between a plan and a result. It is also, not coincidentally, the difference we try to make for the companies we work with.

Acuvity Consulting is California’s boutique Big 4 alternative, bridging business strategy and technology strategy for healthcare, manufacturing, technology, and public sector clients. From AI infrastructure strategy and go-to-market planning to ecosystem mapping and M&A target screening, we help organizations find the real constraint before it finds them. Let’s talk.