Somewhere between the press release and the groundbreaking, hundreds of billions of dollars in announced data center investment has run into something that does not respond to capital: time. Not the abstract passage of time, but the very specific, bureaucratic, physical kind that governs interconnection queues, local zoning hearings, environmental reviews, and the order backlog at every major gas turbine manufacturer on the planet.

An analysis by JPMorgan found that more than 60 percent of data center capacity planned for completion in 2027 has not yet entered construction, with an additional 7 percent already marked as delayed. That figure deserves more attention than it has received. The companies running the artificial intelligence race, including Alphabet, Microsoft, Amazon Web Services, and Meta, have announced capital commitments at a scale that implies operational capacity. The infrastructure reality is considerably more compressed.

Capital commitments have become a form of competitive signaling

Most coverage treats the construction gap as friction, the ordinary lag that accompanies any large capital program, correctable by throwing more money at permit lawyers and construction crews. The more useful frame is sequential dependency. Clear the zoning and you still wait years for a transformer. Secure the grid agreement and you still cannot get a turbine delivered before 2027.

Utilities in high-demand markets are telling developers to expect waits of three to seven years for additional grid capacity. The Trump administration has urged hyperscalers to bring their own power. Some are doing exactly that, strapping gas generators to semi-trailers. That is not an engineering solution. It is a sign that the conventional buildout path has stopped working.

Google did not buy a power company. It bought time.

Alphabet's $4.75 billion acquisition of Intersect Power, a clean energy and data center infrastructure developer, looks different once you understand the interconnection problem. The standard hyperscaler playbook involves power purchase agreements with utilities or independent generators. That model assumed two things that no longer hold: that the grid has sufficient headroom to deliver electrons wherever they are needed, and that interconnection queue timelines are manageable. Neither is true in 2026.

Intersect had pioneered a behind-the-meter model in which solar generation and battery energy storage systems are co-located with compute infrastructure, bypassing grid injection entirely. One facility, using 500 megawatts of solar and 2 gigawatt-hours of storage, completed construction in 18 months while a conventional grid connection for the same site would have required a five-year queue. Alphabet is acquiring a 10.8 gigawatt pipeline of similar assets targeted for delivery by 2028. The deal is not about clean energy credentials. It is about procurement speed.

The cost of building a new combined-cycle gas turbine, or CCGT, plant has risen from under $1,500 per kilowatt of generating capacity in 2023 to $2,157 per kilowatt last year, and it now takes 23 percent longer to complete one, according to reporting by TechCrunch. For developers who cannot wait for a CCGT to be delivered, the alternative is a simple-cycle peaker unit, which can be built in 18 to 24 months but burns 50 to 60 percent more gas per megawatt. The efficiency penalty does not appear in capital cost projections. It appears in operating costs and in scope-1 emissions that cloud providers have committed to reducing.

Nuclear solves 2035. The power problem is 2027.

Microsoft revived Three Mile Island and secured 837 megawatts of nuclear output in a 20-year agreement. Amazon Web Services acquired a campus adjacent to the Susquehanna Steam Electric Station in Pennsylvania for $650 million. Google signed a deal with Kairos Power for small modular reactor supply. Coverage treats these moves as evidence that hyperscalers are closing the gap. Nuclear plants have historically taken over a decade to build. Small modular reactors have not been deployed at commercial scale anywhere. The 2027 supply question remains open.

JPMorgan's broader infrastructure analysis projects $5 trillion in global data center and AI spending through 2030, requiring participation from public capital markets, private credit, and government involvement. Large-frame gas turbines ordered today carry lead times of five years or more, with Wood Mackenzie putting the average delivery window at nearly five years as of last year. Adding 150 gigawatts of generation in a relevant time window is, in JPMorgan's own language, a remarkable challenge.

The state bans are not the outlier

Maine's legislature passed the country's first statewide moratorium on large data centers earlier this year. The governor vetoed it in April, citing a carve-out dispute over a single project, but conceded in her veto message that a pause would be appropriate given the environmental and rate impacts seen elsewhere. Moratorium bills have been introduced in at least a dozen states. Residents near proposed sites are raising objections about electricity rates, water consumption, and noise. These are not fringe positions.

The Trump administration's directive to hyperscalers to supply their own power creates a different problem for the communities hosting them. Utilities generally pass the cost of new generation to all ratepayers, not just data center customers. On-site generation avoids that cost-sharing mechanism but introduces its own political friction when it involves natural gas combustion in areas that did not anticipate an industrial generating facility next to a server farm. Behind-the-meter permitting in residential or mixed-use zones is no faster than grid-connected projects.

Data centers have become a proxy for a broader argument about who pays for the power infrastructure that artificial intelligence requires. That argument is not settled, and it will shape the permitting environment for the next several years regardless of how much capital a hyperscaler has committed on a quarterly earnings call.

The number your vendor quotes is not the number that will serve your workload

Goldman Sachs projects U.S. data center power demand rising from 31 gigawatts in 2025 to 66 gigawatts in 2027. That growth assumes a construction and interconnection pace that the current pipeline does not support. The same research estimates that roughly half of capacity scheduled over the next two years will materialize on time when supply chain and labor constraints are accounted for.

The gap between announced and operational capacity matters for enterprise technology buyers in a specific way. When a cloud provider or AI platform vendor quotes capacity, availability, or performance, the underlying infrastructure promises rest on a construction schedule that is not publicly disclosed. The 60 percent of 2027-targeted capacity that has not yet broken ground is not uniformly distributed across providers. Some vendors are ahead of the average. Others are not.

Before committing to a multi-year infrastructure agreement, a technology leader should ask not whether the vendor has announced capital — that part has proven easy — but whether the vendor can show operational facilities, confirmed interconnection agreements, and power supply already on the grid or behind the meter.