Every major AI company — OpenAI, Anthropic, Google, Microsoft, Amazon — is racing to build more compute capacity. But the physical infrastructure needed to run that compute is running into a wall of problems that no amount of investment can easily solve. A new analysis by real estate firm JLL estimates that 30 to 50 percent of approximately 140 planned US data centers targeting 16 gigawatts of capacity may miss 2026 timelines or be canceled outright. The reasons are a combination of grid constraints, local political opposition, permitting backlogs, and supply chain bottlenecks converging simultaneously.
The Grid Problem Nobody Predicted at This Scale
IDC projects the US will need 40 gigawatts of new data center capacity by 2028 — roughly equivalent to adding 40 nuclear power plants' worth of electrical load in three years. According to the US Energy Information Administration, interconnection queues for new grid connections have grown from a combined 2,000 gigawatts of pending requests in 2023 to over 3,200 gigawatts in 2026, with average connection wait times stretching from 3.7 years to over 5 years in some regions. Hyperscale data centers routinely require 100 to 500 megawatts each. Getting that power connected to the grid in a reasonable timeframe is, in many US markets, simply not possible at the pace the AI build-out requires. Virginia's data center corridor has had an informal moratorium on new large connections in Northern Virginia since late 2024, forcing companies to look to Georgia, Texas, and Ohio — markets now facing their own constraints.
The Local Opposition Factor: Not Just NIMBY
Community opposition to data centers has moved beyond "not in my backyard" into organized political action. In Ohio, a ballot initiative aims to constitutionally ban hyperscale construction. In Texas, multiple counties have passed restrictive zoning ordinances. In Nevada, a proposal to limit data center water usage has advanced through committee. Opponents cite verifiable data on water consumption (hyperscale facilities can use 1 to 5 million gallons of water per day for cooling), grid strain impacts on residential electricity prices, and the mismatch between data center job counts and the scale of infrastructure they require. Before this wave of opposition, the standard data center development timeline was approximately 24 months from site selection to energization. After accounting for community opposition processes, that timeline has stretched to 36–48 months in many markets.
Supply Chain and Equipment Delays Adding Further Pressure
Even sites with secured land, permits, and grid connections face a third bottleneck: equipment. High-voltage transformers have lead times of 18 to 36 months. The global transformer manufacturing industry has finite capacity, and AI infrastructure demand has consumed most of the available 2025–2026 production. Liquid cooling systems — necessary for the densest GPU clusters — also have multi-quarter backlogs. Microsoft acknowledged in its Q2 2026 earnings call that transformer lead times were a material constraint on its Azure expansion plans. See our analysis of Nvidia GPU supply and AI data center demand in 2026 for the hardware side of this equation.
The Strategic Implication: Geographic Diversification
The combination of constraints is forcing AI infrastructure investment into less obvious locations — the American Midwest, the Southeast (Georgia, South Carolina, Tennessee), and international markets including Canada, Mexico, and Western Europe. Microsoft's $3.3 billion investment in Wisconsin and Amazon's data center expansion in Indiana both reflect this logic. The Middle East — specifically the UAE and Saudi Arabia — has become one of the fastest-growing markets for hyperscale capacity. As we covered in our report on global AI infrastructure investment trends, this geographic shift is one of the defining stories of the AI era.
What This Means for You
If you are an AI developer or enterprise buyer dependent on cloud GPU capacity: plan your infrastructure needs 12–18 months out. If you are an investor in data center REITs or infrastructure funds, the capacity constraint is bullish for existing capacity owners. If you work in local government or policy, the Ohio story is a template: communities now have documented evidence that the economic calculus of data center incentives needs serious renegotiation.
Frequently Asked Questions (FAQs)
Q: Why are so many US data center projects being delayed or canceled in 2026?
A: A combination of three simultaneous problems: electricity grid connection backlogs (average wait times now exceed 5 years in some US regions), organized community opposition slowing permitting, and 18-to-36-month lead times for critical equipment like high-voltage transformers and liquid cooling systems.
Q: How much new data center capacity does the US need for AI by 2028?
A: IDC projects the US will need 40 gigawatts of new data center capacity by 2028 to support AI workloads — equivalent in electrical demand to approximately 40 nuclear power plants. Current build rates, even before delays, are not on track to meet this target.
Q: Which US states are most affected by data center construction problems?
A: Virginia (grid moratorium in Northern Virginia), Ohio (tax break suspended, ballot initiative to ban construction), Texas (county zoning restrictions), and Nevada (water usage proposals) are all seeing significant data center development headwinds in 2026.
Q: Will the data center capacity crisis increase cloud computing prices?
A: In the short term, yes — undersupply of GPU compute capacity is already creating pricing pressure in spot markets for cloud GPU instances. However, the medium-term impact depends on how quickly international capacity and grid expansion can offset US delays.
The AI revolution runs on electricity, water, land, and transformers. The technology is moving faster than the physical infrastructure can keep up. That gap is America's biggest near-term AI risk — and it is entirely solvable, but only with the kind of policy urgency that has not yet materialized.
