A gigawatt campus of AI compute now costs roughly $38 billion to build, according to Epoch AI's total-cost-of-ownership modeling. Five facilities at that scale are scheduled to come online globally in 2026, each run by a different hyperscaler. [1] None of these are data centers in the sense the term meant five years ago. These facilities happen to contain servers, and the industry building them has started calling them what they are: AI factories.

The label extends beyond marketing. Nvidia CEO Jensen Huang has spent the better part of two years arguing that a data center running continuous AI inference behaves less like IT infrastructure and more like a production line. He explains this as a facility that consumes power and raw silicon on one end, and outputs tokens on the other. [2] Whether or not you accept the framing, the capital structure behind it agrees. Amazon, Microsoft, Google, and Meta have guided toward a combined 2026 capital expenditure of more than $700 billion, with most of it directed at AI infrastructure. [4] McKinsey puts cumulative AI-related data center investment at $5.2 trillion by 2030 in its base case and $7.9 trillion in its high-growth scenario. [5]

The problem is that nobody running these projects has done one before. The playbook for a hyperscale data center, one the industry has iterated for two decades, assumes three things: a 40–500 MW facility, a conventional utility interconnection, and a construction schedule measured against comparable recent builds. The AI factory breaks every one of those assumptions simultaneously, and it is doing so at a fast pace. It has already outrun the standard risk controls capital allocators use to underwrite megaprojects.

What Scale Looks Like Now

Meta's Hyperion campus in Louisiana is planned to scale to 5 GW (a footprint Mark Zuckerberg has said would cover a significant part of Manhattan). Since no utility in the region could absorb that load through conventional interconnection, the site requires dedicated natural gas power. Three plants are under construction, and seven more are planned. [6]

Purpose-built gas-fired power is one way hyperscalers are working around grid limits. Photo: coffeekai/Envato Elements

Meanwhile, xAI's Colossus 2 in Memphis is targeting gigawatt scale within 12 months of groundbreaking, which would make it the fastest of the five such facilities due online in 2026. [1] When the larger picture clicks, the gap is obvious. Simply to clear the U.S. interconnection queue, a conventional generation project faces a backlog of over 2,060 GW, even after a wave of withdrawals. The typical project still takes more than five years to reach commercial operation. [7]

Bypassing the conventional interconnection queue has become its own strategy. Photo: Visual__Production/Envato Elements

An AI factory is larger than a conventional megaproject. It proceeds on the assumption that the constraint defining every other large power project in the country simply doesn't apply to it.

The 128-Week Wait

As AI-driven demand collides with a global manufacturing base that spent a decade underinvesting in transformer capacity, large power transformer lead times have stretched to roughly 128 weeks on average, and as long as four to five years at the extreme. [8][9]

Demand for generator step-up transformers, the large units that connect a power plant to the grid, has grown 274% since 2019. [10] Roughly half of the 12 GW of U.S. data center capacity originally planned for delivery in 2026 is now at risk of delay or cancellation. Only 4–5 GW actually remain under active construction against that figure. [8]

The Skills Gap: Why Hyperscalers Are Training Their Own Workforce

Data center-related job postings have more than doubled over the past two years [11], and the labor shortage is now acute enough that Meta has committed $115 million in the first year of a tuition-free trades academy.

Skilled trades are now a bottleneck for AI infrastructure. Photo: ckstockphoto/Envato Elements

It is explicitly designed to guarantee graduates a hard-hat job with one of its contractors. [12] Meta’s behavior reflects a company that’s not confident in its supply chain of skilled electricians and pipefitters. It has concluded the labor market cannot produce them fast enough on its own.

The Financing Shift

To fund Hyperion alone, Meta entered a roughly $27 billion joint venture with Blue Owl Capital. Blue Owl covered about 80% of the cost. [13] That structure is becoming the default, not the exception. Even a hyperscaler with an investment-grade balance sheet now moves gigawatt-scale construction risk off its own books and into a project-finance vehicle. That is because the risk profile of an AI factory doesn't resemble the risk profile of the balance-sheet capex that came before. Then, a conventional data center lease was underwritten against occupancy and tenant credit. Now, an AI factory is underwritten against equipment delivery dates, transformer allocation, and a labor pipeline that, in several U.S. markets, doesn't exist yet at the scale required. Lenders and infrastructure funds are pricing a category with almost no real-world performance data.

Europe's Full, the Middle East Is on Pause

Geography compounds the problem rather than diversifying it. In the established European colocation markets (Frankfurt, London, Amsterdam, Paris, Dublin) vacancy stood at 6.4% in the second quarter of 2026, close to its record low, and just 3.1% in Frankfurt. [3] That looks like healthy demand until you notice what it actually means: there is essentially no slack left in the system to absorb a schedule slip. In the Middle East, the problem is different. Its pipeline, 2.6 GW in development against 1.6 GW of live capacity, has stalled as regional conflict delays major completions, though JLL says projects are being deferred rather than dropped. [3]

None of this means the AI factory is a bad bet. Nvidia CEO Jensen Huang said in March that he sees at least $1 trillion in orders for its Blackwell and Vera Rubin platforms through 2027, double the $500 billion through 2026 that he cited in late 2025. [14] This by itself implies a multi-gigawatt construction wave across the U.S., Asia, and Europe regardless of how any single project's risk profile shakes out. The capital is committed and the demand signal is real. What remains missing is the operational maturity to match it. This is the accumulated, boring, hard-won institutional knowledge that lets a capital allocator look at a megaproject and know, six months before the delay becomes visible, that the equipment order is slipping or the labor plan won't hold.

That knowledge doesn't exist yet for this category, because the category itself is so new. The first gigawatt-scale AI data centers are only now coming online. [1] On a conventional megaproject, a slipping transformer order or a thinning labor plan shows up as a schedule variance months before it becomes a headline. These signals lie buried in procurement data and subcontractor staffing levels long before they reach a steering committee. The AI factory generates the same signals, but the industry doesn't yet have anyone whose job is to read them at the speed the category is being built.

First Movers Win: Closing the Knowledge Gap

That gap will close. Every capital-intensive industrial category eventually develops its own institutional memory. It’s the instinct for which procurement delay is routine and which one is the first data point in a schedule collapse. The AI factory will get there too, probably within a construction cycle or two. The only question that matters right now, for developers signing 12-month schedules, lenders pricing a risk category with almost no track record, and hyperscalers writing checks against production dates their suppliers can't yet guarantee, is whether to build that institutional memory before the current wave of projects needs it, or after.