Top 7 Power Risks Facing America’s AI Data Center Boom

America’s AI buildout is running into a stubborn timing problem. Data centers can fill with a new generation of chips while the grid connection they need is still years away. That mismatch decides which billion-dollar campuses get power, where they get built, and how soon they can switch on.

A recent McKinsey analysis puts the emerging mismatch at 30 to 55 GW of additional US capacity needed by 2030, with data centers expected to account for roughly three-quarters of power demand growth over the coming decade. The national figure only gets interesting once it is pulled apart. Connection requests are running far ahead of credible demand, spare generation often sits far from the clusters that need it, aging plants are having their retirement dates reconsidered, and developers are bringing generation directly onto data center sites to secure earlier opening dates.

What looks from a distance like one enormous electricity shortage is a collection of very different constraints, some physical, some regulatory, and a few created by the data center industry itself. The AI buildout is encountering seven of its hardest challenges.

1. The megawatts are often in the wrong place

Electricity may be fungible on a spreadsheet, but grid capacity is stubbornly local. McKinsey expects much of the new load around Northern Virginia, Phoenix, Louisiana and parts of Texas, several areas where generation or transmission capacity is already tight.

PJM’s grid operator’s July 2026 capacity auction procured 138,318 MW and cleared at $325 per MW day, essentially at the imposed cap. Meanwhile, PJM’s latest forecast has summer peak demand rising by 65,733 MW over ten years to reach 222,106 MW in 2036.

Land, tax incentives and fiber still count, while a credible energization date can decide whether a campus opens in 2029 or remains a rendering on a developer’s website.

2. The queue contains the same demand several times

Here the numbers become almost comic. McKinsey calculates that large load interconnection requests amount to roughly nine times the new IT demand in its central 2030 scenario. Even contracted and high-confidence projects come to around twice projected demand.

Developers submit applications in several locations while hunting for the fastest connection, so a grid planner can see several requests representing the same future servers.

FERC ran into that problem in June, saying speculative requests were clogging studies and distorting forecasts before introducing tougher readiness requirements. McKinsey’s AEP Ohio example is even cleaner. After a new large load tariff, requested capacity dropped from 30 GW to 5.64 GW, an 81 percent fall.

3. Data centers are turning into power projects

When grid connection dates stretch, the building starts acquiring its own energy system. McKinsey’s 2025 industry survey found that 65 percent of respondents expected to deploy some form of onsite power. Nearly 60 percent expected permanent onsite generation by 2030 even after grid service becomes available, and 64 percent of those planning onsite generation expected to rely on natural gas.

Those dry percentages describe a major change. Onsite equipment once sat mainly in the resilience plan; it is moving into the main supply stack. FERC is also developing rules for colocated loads willing to limit grid withdrawals during constrained periods. Power design now enters development much earlier, alongside land, cooling and fiber.

4. Gas power trades a grid queue for a fuel queue

Gas engines and turbines can provide firm power on data center schedules, which explains their appeal. The catch arrives several miles away. McKinsey points to pipeline capacity, redundant fuel supply and rights of way as potential delays for facilities demanding very high uptime.

That gives Texas and Gulf Coast locations another advantage. Proximity to production and existing pipelines can shave infrastructure work from a project, while a remote site may spend years waiting for fuel connections.

5. Old power plants keep changing their leaving date

At the same time that AI load is arriving, a large block of conventional generation is approaching retirement. McKinsey estimates 50 to 75 GW of coal and aging thermal capacity could leave by 2030. The US Energy Information Administration says nearly 11 GW was scheduled for retirement during 2026 alone, including 6.4 GW of coal capacity. Several plants have already stayed open past earlier plans following federal emergency orders aimed at preserving reliability.

This is one of those details power planners tend to watch more closely than AI investors. Extending an old plant buys time while changing fuel demand, emissions trajectories and hyperscalers’ clean power plans.

6. The compute forecast gets fuzzier after 2030

McKinsey models data center IT demand reaching 121 GW in 2030 after growing at roughly 27 percent a year. The Department of Energy’s July transmission study cites estimates above 400 TWh of data center load growth by 2030 and earlier FERC projections of 13 to 55 GW in added peak demand. Virginia and Texas currently have the highest estimated demand, with Arizona and Oregon expected to add heavily.

Chip efficiency is improving quickly, enterprise adoption moves at a different pace from hyperscaler spending, and parts of the AI sector are being financed well ahead of proven revenue. Power infrastructure also has customers beyond the data center gate. Generation and transmission can serve factories, homes, and other commercial loads if compute demand lands toward the lower forecasts.

7. New nuclear arrives on a later timetable

The technology companies signing nuclear and geothermal deals are looking beyond the immediate squeeze. Through 2030, McKinsey expects additional US nuclear output mainly from upgrades and life extensions at existing plants. Small modular reactors and next-generation geothermal sit on a later clock, with material capacity expected around the middle of the 2030s.

For the next four years, the work is considerably more prosaic. Transmission approvals, substations, existing plants, gas supply, batteries and cleaner interconnection queues will determine how much announced capacity reaches operation. Developers are already adapting through onsite generation and more flexible arrangements with grid operators.

Bottom Line

The 30 to 55 GW gap is ultimately a timing problem playing out town by town, utility by utility. A 500 MW campus with a credible connection and fuel plan has a very different future from a 500 MW request sitting in three interconnection queues at once.

By 2030, America may have built plenty of power in aggregate while individual data center markets remain painfully tight.