What happens when the hunger for artificial intelligence collides with the limits of an already stretched power grid? In Texas, the answer arrived faster than almost anyone expected. By mid-2025 the state’s grid operator was staring at a queue of nearly 474 gigawatts of proposed large new customers. Roughly ninety percent of that mountain of demand came from data centers. That figure sat more than five times higher than the grid’s record peak load. Then, almost overnight, the state hit pause. New permits froze. Interconnection studies stalled. Developers who had already poured money into land and equipment found themselves waiting.
I’ve watched energy markets long enough to know that sudden freezes rarely come from a single cause. Politics play a role, especially with elections on the horizon. Yet the deeper story is about how the grid works, who pays for upgrades, and whether America can still build the infrastructure its most ambitious industries need. This is not just a Texas story. It is a preview of the friction every region will face as AI scales.
The Real Reasons Behind the Pause
Two forces collided at once. The first is pure volume. Grid planners designed their review process for a few dozen large loads a year. Suddenly they faced hundreds of applications, many of them speculative. Developers routinely filed projects at multiple sites, hoping one would stick. Some had never connected a single GPU to the grid. Sorting the serious from the speculative became nearly impossible under the old rules.
The second force is community tension. Residents worry about noise, water use, emissions, and rising power bills. When companies approach local officials under code names and request tax breaks before revealing their identity, trust erodes quickly. Only a fraction of companies answered state surveys about their resource needs. That lack of transparency made elected officials cautious, and caution turned into a freeze.
Understanding the Interconnection Bottleneck
Connecting a large data center to the grid is no longer a simple paperwork exercise. A decade ago, a company like a major cloud provider could find a utility with spare capacity, sign a large-load agreement, and match its consumption with financial power contracts. The wires themselves had room. That era is over. Almost every new large load or generator now requires physical upgrades to substations, transformers, or transmission lines.
Power does not travel in a straight line from plant to customer. It flows across every available path. Buying electricity from a specific generator does not reserve a private highway to your building. Grid studies must therefore examine what happens under stress. Planners simulate transformer failures, line outages, and sudden generator trips. Maintaining reliability standards often means building extra infrastructure that no single project wants to fund alone.
A newer risk has emerged that makes planners even more careful. Many data centers switch instantly to backup generators at the first sign of a voltage dip. They do this to protect sensitive hardware. The result can be an abrupt drop of hundreds of megawatts from the grid in seconds. In an isolated system like Texas, losing more than about 3.2 gigawatts at once can create serious stability problems. New voltage ride-through rules now require data centers to stay connected through ordinary faults. Some operators even keep dummy computing jobs running after a training run fails just to avoid sudden load loss.
How Cost Allocation Creates Friction
Who pays for the upgrades is often the slowest part of the process. Texas traditionally allocates transmission costs based on a customer’s demand during summer peak hours. A large load that deliberately ramps down on the hottest afternoons can avoid much of that charge. Regulators have proposed counting all twelve monthly peaks and treating large loads as if they operate at full size. In short, buyers would pay according to the capacity of the “pipe” they require, not how much power actually flows through it most of the time.
That sounds fair on paper. In practice, upgrades built for one campus often benefit existing customers or create headroom for future growth that is hard to value up front. If a project runs over budget or never materializes, other ratepayers absorb the difference. New rules require a flat study fee plus a deposit of fifty thousand dollars per megawatt. A one-gigawatt project posts at least fifty million dollars. Yet if the project loses its place in the queue after two years of delay, it forfeits only twenty percent of that deposit. The deposit also typically covers only project-specific upgrades, not the larger regional lines whose cost is shared across the system.
Fully connecting a large data center can take five to ten years. Phased connections have become more common. They give utilities clearer intermediate targets and allow developers to prove they can actually use the capacity they requested. Right now even those phased projects sit in limbo while an audit determines which applications in the current batch are legitimate.
Why Developers Turn to On-Site Power
Idle GPUs cost far more than the electricity that runs them. Time to power has become the single most important metric for many developers. That is why so many projects now plan to bring their own generation “behind the meter.” Almost every site I have examined still intends to connect to the larger grid as soon as possible, because on-site power almost always costs more. The private plant simply bridges the gap until the wires arrive.
Some campuses plan large on-site gas plants designed to run independently for years. Others pair wind or solar with batteries and a partial grid connection from day one. A few describe themselves as pure microgrids that will never draw from the public system, yet their site plans still include switchyards next to high-voltage lines. The pattern is clear: independence is temporary. The grid remains the long-term goal.
Running an islanded system is harder than it looks. The larger grid provides inertia, fault current, voltage stability, and black-start capability that a private plant must supply itself. AI workloads add another layer of difficulty. Rapid swings of ten to twenty megawatts several times per second can stress turbine shafts. Operators have responded by installing large battery systems to smooth those swings. Fuel supply creates its own risks. Natural gas plants keep little fuel on site. A severe winter storm can interrupt pipelines just when demand peaks. Solar replaces fuel risk with weather risk, and covering a cloudy winter week at gigawatt scale requires enormous battery capacity.
Full reliability on an island is extremely expensive. Nuclear plants, among the most reliable sources, still average only about ninety-two percent uptime because of refueling outages. True redundancy means building essentially a second power plant. That is why many behind-the-meter designs favor modular gas engines that can be added or repaired in smaller increments. Even then, maintaining hundreds of engines is no small task.
Some hyperscalers have shown they will accept lower reliability if it means faster deployment. Designs targeting ninety-nine percent uptime, roughly eighty-eight hours of downtime per year, are appearing. With GPU hours this valuable, that trade-off can still beat waiting years for a firm grid connection.
The Limits of Private Power Networks
Texas, like most states, tightly restricts who can sell power to whom. You may supply your own campus, employees, or tenants. You may run a private-use network that sells surplus into the grid and buys when short. Selling to a neighboring factory usually requires a retail electric provider certificate. Building your own transmission line often collides with the local utility’s service territory.
Proposals exist to loosen those rules and allow private utilities to serve multiple large customers across their own network. A few other states already permit large loads to contract for fully off-grid systems. If demand remains this intense, the states with the most flexible frameworks may capture a larger share of the buildout. The risk, of course, is a utility death spiral in which fixed grid costs fall on a shrinking number of remaining customers.
How Flexibility Changes the Equation
A campus that can keep itself running when the grid is tight becomes easier for operators to accommodate. Flexibility can take several forms: shifting compute to other hours or locations, switching to on-site generators, or discharging batteries. Studies suggest that modest flexibility could unlock significant capacity without new generation. One analysis found that ERCOT could theoretically add about ten gigawatts of new load if that load reduced its grid consumption by only half a percent over the year. Average curtailment events last roughly two hours, a duration batteries handle well.
In practice, flexibility must be credible and always available when operators need it. Generators already operate under connect-and-manage rules that allow early interconnection in exchange for potential curtailment during congestion. Data centers that have promised customers high uptime find that bargain harder to accept. Some Bitcoin mining operations have made the trade successfully. New batch processes offer large loads two optional paths: bring your own power and count it toward your size if you can cut load within one minute of a plant failure, or accept automatic cuts above a guaranteed share when lines are congested.
Distributed resources such as aggregated home batteries are beginning to prove themselves in wholesale markets. When trusted by planners, they can add capacity far faster than a new peaker plant or transmission line. Advanced hardware like solid-state transformers can measure and steer power flows more precisely than traditional equipment. Dynamic line ratings and higher-capacity conductors can also squeeze more capacity from existing wires. The challenge is that planners still size upgrades for a flexible load’s full planned demand. Flexibility speeds connection but does not yet shrink the required infrastructure.
Scaling to Hundreds of Gigawatts
On-site power and flexibility will determine how the next few campuses come online. The larger question is how the industry reaches the scale its leaders discuss after 2030. One hundred continuous gigawatts of load would consume roughly a fifth of current national electricity use. Planned additions of solar, wind, and gas across the entire country for the coming year fall well short of that figure when realistic capacity factors are applied.
Gas remains popular because it can run when needed, yet scaling it aggressively raises emissions questions. Equipment for reciprocating engines and fuel cells is sometimes easier to obtain than large combustion turbines. Solar has the manufacturing scale the industry needs, but most of that capacity sits outside the United States. Geothermal and advanced nuclear offer firm low-carbon power, yet both still face long development timelines. Transformers, switchgear, and skilled crews remain bottlenecks regardless of fuel source.
Transmission is its own story. Major new high-voltage lines take years to permit and build. Federal permitting reforms under discussion could help, but nothing moves as fast as demand is growing. My own view is that a handful of repeatable templates will dominate: on-site gas and batteries as a bridge, solar layered where geography allows, and geothermal or reactors arriving once they prove commercial. Who owns the plants is still unclear. Some technology companies are buying generation developers. Oil and gas firms are exploring broader grid roles. New “neo-utilities” may emerge if regulation allows private networks serving multiple campuses.
What Happens After the Freeze Ends
Checking that projects in the current batch are real is reasonable. Credible developers who have funded early phases and posted substantial deposits should be able to move forward. A blanket permit freeze that treats a funded first phase the same as a pure speculation is harder to justify. Narrowing the freeze to exclude emergency-only generators and projects that have already committed serious capital would restore some momentum without abandoning oversight.
Looking ahead, clearer rules would help everyone. A developer that pays its deposits, funds reserved capacity, and hits phased milestones should receive a firm connection date. Agreeing to curtail when the grid is stressed should accelerate that date further. Utilities building the upgrades should face accountability for delays comparable to the accountability placed on developers. Competitive bidding with cost caps and performance penalties is one idea worth exploring.
Power will remain tight for years. The bottleneck will simply move from one piece of equipment or one regulatory step to the next. Data centers are not unique in the hardware they need. They simply hit the limits first. In that sense they are a rehearsal for every other industry that wants to expand or reindustrialize.
There is also an opportunity. AI companies will pay almost any price for reliable power and can locate almost anywhere. The revenue they bring can fund upgrades the grid needs anyway. In the best case, those upgrades lower average costs for everyone else by spreading fixed expenses across more sales. Early examples in other states show data center revenue reducing household bills or funding community benefits. The key is transparent deals in which developers truly cover the costs they impose rather than shifting them onto existing customers.
Trust matters. A campus that levels with its neighbors and follows through makes the next project easier to welcome. Get that relationship wrong and political resistance hardens. None of the current friction justifies a permanent pause. It does justify getting the rules and incentives right while the industry scales. Most large buyers would rather connect to the public grid than build permanent private islands. Letting them do so under clear, fair terms serves the broader interest.
Texas built the energy fast lane once before. It now has the chance to design a better one. How it handles this moment will tell us a great deal about whether the country can still construct the physical systems its most ambitious technologies require. Power, more than almost anything else, is where America discovers whether it can still build at scale.
