My father sits on the board of an electric cooperative in southwest Texas. Over the past year, our family conversations have shifted from cattle prices and drought conditions to something I didn't expect: data centers. Specifically, the wave of hyperscaler inquiries looking to build AI infrastructure in rural Texas.
It's one thing to read about AI's energy demands in industry reports. It's another to hear about it from someone fielding requests to bring net new power infrastructure to areas that, until recently, mostly worried about keeping the lights on during summer peaks.
That personal connection is why the numbers coming out of ERCOT this week hit differently for me than they might for most people reading tech coverage.
The Numbers Are Staggering, and That's the Problem
Texas's grid operator, the Electric Reliability Council of Texas (ERCOT), now has more than 230 gigawatts of large load interconnection requests, nearly four times the 63 gigawatts on the books at the end of 2024. More than 70% of those requests come from data centers.
To put that in perspective: 230 gigawatts is roughly three times ERCOT's total current capacity. Many individual requests exceed one gigawatt per site, which CNBC noted is equivalent to half the power produced by the Hoover Dam.
And yet, the number of projects actually approved or connected to the grid? Around 7.5 gigawatts.
That's a 30:1 ratio between what developers are asking for and what the grid can actually deliver. ERCOT is now considering revisiting some previously approved projects because the original planning assumptions have changed.
This isn't a story about AI succeeding too fast. It's a story about AI expectations outpacing physical reality.
The Batch Zero Reckoning
ERCOT's response to this chaos is a new process they're calling Batch Zero. Instead of processing interconnection requests one at a time, they'll evaluate batches every six months. Projects that have been waiting longest and don't require extensive restudies get first consideration.
Jeff Billo, ERCOT's VP of interconnection and grid analysis, put it bluntly: the current system, built for 40 to 50 projects, is now handling 225 new requests from last year alone. ERCOT is moving toward what Billo called "one study to rule them all" because they're on the verge of going back to developers who already cleared studies to say the original assumptions no longer hold.
The practical impact: if a developer requests 500 megawatts by 2028, but studies show only 100 megawatts are available until transmission upgrades complete in 2030, they get an "on-ramp" of 100 megawatts now and the rest later. After batch studies finish, developers have a set window to make financial commitments. No commitment, no guaranteed capacity.
This is the infrastructure equivalent of calling AI's bluff. Show us you're real, or get out of the queue.
The Hyperscaler Land Rush
The scale of investment is massive. Former Google CEO Eric Schmidt's new venture, Bolt Data & Energy, has partnered with Texas Pacific Land, a 150-year-old oil company controlling 882,000 acres of West Texas land. Their goal: 10 gigawatts of power generation capacity for AI data centers.
OpenAI and Oracle's Stargate complex in Abilene is expected to reach 1.2 gigawatts by 2026: enough to power more than a million homes.
These aren't speculative projects. They have funding, partnerships, and timelines. But they're competing for grid capacity that doesn't exist yet.
ERCOT projects that statewide power demand could double by 2030, with data center demand specifically rising to 78 gigawatts by 2031. Meanwhile, between 2024 and 2025, about 23 gigawatts of new generation came online, with another nine planned for early 2026. The math doesn't add up. Requests are growing faster than infrastructure can be built.
The Pattern No One Connects
Here's what the Texas grid situation reveals about the broader AI infrastructure conversation: the hype is creating physical strain before the value materializes.
I've written about this pattern at the enterprise level. In The AI Infrastructure Panic Is Self-Inflicted, I noted that 83% of companies predict their own infrastructure will fail under AI load, while 75% remain stuck in pilot mode. The infrastructure isn't straining under production AI workloads. It's straining under the weight of endless experimentation.
Texas is the grid-level manifestation of the same dynamic. The 230 gigawatts of interconnection requests aren't based on operational AI deployments generating revenue. They're based on anticipated demand from projects that mostly haven't broken ground, serving AI workloads that mostly haven't moved from pilot to production.
Some of these projects will succeed. Many won't. Texas Policy Research notes that projections that large load projects such as data centers are entering the queue faster than new generation or transmission lines can be built. That's not a sign of AI's success. It's a sign that speculation is outrunning execution.
Rural Texas as Canary in the Coal Mine
The reason my father's co-op is fielding inquiries isn't just land availability. It's grid topology.
A new data center model in Willacy County demonstrates the logic: position facilities near wind farms to capture energy that can't be efficiently transmitted to ERCOT's main grid. By consuming power locally, these projects avoid transmission constraints and use generation that would otherwise be curtailed.
This is creative problem-solving. It's also an acknowledgment that the main grid can't handle what's being requested.
Rural electric cooperatives, which weren't built for industrial-scale loads, are suddenly being asked to evaluate gigawatt-scale projects. The expertise, infrastructure, and capital requirements for that kind of transformation take years to develop. Meanwhile, developers are moving fast, trying to secure commitments before capacity tightens further.
That creates a tension my father sees firsthand: pressure to say yes to projects that represent generational economic opportunity, balanced against the practical reality that serving those loads requires infrastructure that doesn't exist.
What's Not Being Discussed
Coverage of ERCOT's data center challenge focuses on the regulatory and technical dimensions: queue management, transmission planning, generation capacity. What's missing is the underlying question of whether the demand projections driving all this activity are realistic.
AI data centers consume 10 to 30 times more energy than traditional facilities. But how many of the 230 gigawatts in ERCOT's queue will actually be built? How many of those that get built will run at projected capacity? And how many will generate the economic value needed to justify the infrastructure investment?
We've seen this movie before with cryptocurrency mining, which drove previous waves of Texas grid requests. Some of those facilities operate. Many never materialized. The ones that did have faced periods of curtailment during grid stress, exactly when they'd otherwise want to run.
AI workloads are different: they're more valuable, more persistent, and backed by better-capitalized companies. But they're also riding a wave of investment that, as Oxford Economics noted in analyzing AI-related hiring trends, may be based on anticipated returns that haven't been demonstrated at scale.
The Discipline Texas Needs
ERCOT's Batch Zero process is a step toward rationality. By requiring financial commitments before transmission investments, it filters speculation from genuine projects. By processing requests in batches, it forces developers to compete on readiness rather than just timing.
But the broader discipline required goes beyond queue management.
Match infrastructure planning to demonstrated demand. The 230 gigawatts of requests are projections. Planning should weight actual commitments more heavily than speculative asks. If developers aren't willing to put capital at risk, their capacity requests shouldn't drive grid expansion.
Build flexibility into rural infrastructure. Co-ops and rural utilities being asked to serve data centers need modular approaches that can scale if projects succeed but don't strand assets if they don't. Senate Bill 6, passed in Texas in 2025, gives ERCOT authority to mandate backup generation or usage curtailment during emergencies. That's a start.
Question the growth projections. ERCOT's forecast of 78 gigawatts of data center demand by 2031 is based on developer submissions. Those submissions are based on AI demand forecasts. Those forecasts are based on... what, exactly? If 95% of enterprise AI pilots fail to produce ROI, as MIT research found, the demand underpinning these projections may not materialize.
The Real Story
Texas isn't just a venue for AI's energy demands. It's a real-time experiment in what happens when AI hype collides with physical infrastructure limits.
The 30:1 ratio between requested and approved capacity isn't a bug in the process. It's the process working as intended, creating friction that forces developers to demonstrate commitment before infrastructure gets built.
For my father's co-op and others like it across rural Texas, that friction creates breathing room to make thoughtful decisions about which projects to support and how to serve them. The alternative, saying yes to everything based on projected demand, is how you end up with stranded infrastructure when the projections don't pan out.
The AI boom is real. The investment is real. But the physical infrastructure to support it exists on a different timeline than the hype cycle. Texas is learning that lesson in real time. The rest of the country, and the enterprises making AI infrastructure decisions, should be watching.
