AI Data Centers Face Delays Despite Massive US Infrastructure Investment

AI data centers are attracting hundreds of billions of dollars in investment, but the US buildout is moving much more slowly than the headlines suggest. Material shortages, limited chips, grid constraints, labor shortages, and permitting disputes are delaying new capacity. For teams building AI products, the practical takeaway is simple: planned compute is not the same as available compute, so infrastructure assumptions need more schedule and capacity buffer. The construction pipeline is already slipping.

The US pipeline is large, but mostly still planned

JPMorgan estimates that around $750 billion is being invested in AI infrastructure in 2026. Aterio has counted plans for 3,969 new US data centers, on top of 5,427 existing facilities reported at the end of last year. Yet only 802 of those planned facilities are currently under construction, according to the figures reported in the article.

That gap matters because developers often submit applications in several regions and later select only the sites that secure power, permits, financing, and equipment. The announced pipeline therefore overstates the amount of compute that will necessarily reach production.

Why AI compute capacity timelines are slipping

The main constraint is not a single missing component. It is the interaction between several bottlenecks. Data center construction typically takes 18 to 24 months, and those timelines are stretching as demand rises.

Chip supply is one limitation. Taiwan Semiconductor Manufacturing Company, or TSMC, fabricates leading AI processors including Nvidia Blackwell and AMD MI300X hardware. That concentration makes advanced accelerator supply dependent on a crucial part of the global manufacturing chain.

Power infrastructure is another major constraint. Data centers account for roughly 8% of US electricity use, a figure that the American Edge Project predicts could reach 12% by 2028. Wait times for generation step-up transformers have reportedly tripled, making it harder to connect new facilities even when a site is otherwise ready.

Construction labor is also scarce. Meeting the proposed buildout would require hundreds of thousands of additional electricians, welders, and plumbers, according to estimates cited in the report. In practice, a project can be delayed by the availability of specialized contractors as much as by the building itself.

What this changes for AI builders

Goldman Sachs estimates that only about half of the AI compute capacity scheduled to activate between now and 2028 will come online by its target date. Separately, Columbia Business School professor Stijn Van Nieuwerburgh expects only 180 gigawatts of the 565 gigawatts currently planned to be built over the next decade. These are estimates, not confirmed outcomes, but they point in the same direction: the AI compute capacity timeline is likely to be uneven.

For founders and product teams, that argues for flexible deployment plans. Avoid tying a launch to a single new region, accelerator class, or hyperscale capacity announcement. Build fallbacks around smaller models, inference optimization, reserved capacity, or multiple providers where the workload allows it. This will not remove supply constraints, but it can reduce the operational impact of infrastructure delays.

Permits may matter more than bans

Public resistance has become a visible part of the data center backlash. Roughly a dozen states have proposed moratoriums, and temporary bans have taken effect in New York and Texas. However, Goldman Sachs identifies permitting delays as a larger obstacle than outright bans in many cases. The regulation impact on data centers may therefore come through slower approvals, local conditions, and grid negotiations rather than a nationwide prohibition.

The spending continues despite the friction. US data center construction spending reached $68.3 billion in June, up 7% from the previous month and 46% from a year earlier, according to the report. That level of capital expenditure creates pressure to build, but it does not guarantee that every announced project is economically or technically viable.

The useful decision rule for AI builders is to treat new data center capacity as a probabilistic resource until power