Texas's pause on advancing data-center grid applications is a test of which proposed AI campuses are real enough to plan around. ERCOT's large-load queue grew from about 63 gigawatts at the end of 2024 to 474.7 gigawatts by June 2026, including 420.8 gigawatts of data-center requests. That is requested demand, not power already consumed, facilities under construction or an approved delivery schedule.
Ryan McEntush's a16z analysis frames the problem as queue quality rather than a rejection of AI investment. The distinction matters: a grid operator can overbuild around speculative demand, but can also delay genuine projects by discounting them too aggressively. Texas is trying to identify the credible portion before committing shared infrastructure and scarce engineering resources.
The first screen was far smaller than the headline queue
In July, ERCOT reported about 205 gigawatts initially eligible for Batch Zero, its first coordinated study under the revised large-load process. It also identified 315 projects without qualifying studies and 47 without required dynamic models. Those project counts should not be read as gigawatts or combined with eligible capacity as if they measured the same thing.
A dynamic model estimates how equipment responds to voltage changes, faults and other disturbances. Missing models prevent reliable assessment of how a campus affects the surrounding network. Even a financially credible tenant therefore needs technical readiness, not merely a reservation.
Governor Greg Abbott's August 3 directive required a comprehensive review before additional data centers could advance. It asks about ownership, public assistance, power sources, water, cooling and protections for neighboring properties, as well as electrical matters. Refusal can mean denial of grid access. ERCOT delayed Batch Zero while regulators worked out implementation.
This is a pause in advancement through a process, not a permanent statewide ban. It nevertheless changes the value of an early queue position: an application now needs evidence behind its promised load and timetable.
Multiple sites can represent one eventual customer
A developer may seek several locations because it does not yet know which can secure power. Different developers can also compete for the same prospective tenant. A cloud provider considering four sites might therefore appear behind four requests while intending to occupy only one.
Confidential negotiations obscure those overlaps. An early application provides valuable optionality and can enhance a parcel's appeal without proving that a customer, financing or final site exists. The engineering work still lands on ERCOT and utilities.
Verification has its own sequencing trap. Customers may want power certainty before signing, while a utility wants a committed customer before treating demand as firm. Site control, equipment orders and financial security reduce uncertainty but do not eliminate market risk. Well-funded companies can also maintain several options more easily than smaller credible developers.
The Associated Press reporting cited in the source describes similar concerns in other states. The Data Center Coalition supports stronger commercial-readiness assessment while urging Texas to distinguish responsible investors and move quickly. A broad pause can burden serious projects alongside weak ones if criteria and deadlines remain unclear.
SB 6 shifts some risk back to applicants
Texas Senate Bill 6 generally applies large-load standards to sites above 75 megawatts. Its requirements include contributions toward reasonable interconnection costs, disclosure of substantially similar Texas requests, evidence of site control, financial commitments, and disclosure of significant on-site backup generation.
These provisions make duplication more visible and put more applicant capital at risk. The wider framework uses study fees, capacity-linked security, milestones and possible reallocation. A developer withdrawing after transformer procurement should not automatically leave unrelated customers with all the cost.
Direct interconnection charges do not settle every shared expense. A line or substation triggered by one campus may later serve factories, homes or other loads. Charging the first applicant for everything can overstate responsibility; socializing everything can strand costs on residents. Electricity contracts alone do not reserve a physical path through an interconnected grid.
On-site generation is not an escape from those questions. Gas introduces fuel and emissions requirements; renewables need complementary supply or storage; batteries have duration limits; nuclear has a different development timetable. Even self-supplied facilities may use transmission for backup or exports.
Planned flexibility is not uncontrolled disconnection
SB 6 provides routes for projects bringing their own power or accepting binding curtailment, with some expedited arrangements requiring commitments of at least ten years. Some colocated arrangements must reduce load or switch supply during emergencies so existing generation remains available to the wider system.
A campus can defer suitable training, shift work elsewhere, use batteries for immediate response or generation for longer events. None is free. Real-time inference, healthcare and financial workloads may be difficult to move; training interruption wastes coordination; data movement takes time and energy, and privacy or latency can limit destinations.
ERCOT also needs voltage ride-through. Sensitive electronic equipment may rapidly disconnect during disturbances; simultaneous loss of several large loads can destabilize the grid. Ride-through requires staying connected under defined disturbances. Planned curtailment reduces demand when instructed; ride-through prevents a reduction at the wrong moment. They are complementary controls, not contradictory promises.
Credible flexibility therefore needs identified workloads, response times, duration, telemetry and maintained equipment. A contractual promise that arrives too late cannot relieve a constraint. Similarly, a generator needs fuel, operating reliability, synchronization and environmental compliance before its capacity is dependable.
The audit also addresses a local information deficit
The source cites a state survey in which only 28 of 377 contacted companies had responded by June. That does not prove wrongdoing by every nonrespondent. It shows why regulators lacked basic resource information.
Residents face water use, generator emissions, constant noise, construction traffic and lighting, not just a grid model. Hood County's consideration of the initially unidentified “Project Patriot” illustrates how project secrecy can undermine decisions on incentives and zoning. Construction spending and property taxes may benefit a community, but permanent jobs and long-term resource commitments are different measures.
Cooling choices require project-specific disclosure. Air cooling can reduce direct water demand while increasing electricity use under some conditions; closed loops still need initial supply and maintenance. Backup generators used more frequently for grid support can also change local impacts. Ownership, annual and peak consumption, water sources and operating plans therefore belong in readiness review.
Judge the repair by progression, not another request total
The next meaningful evidence is the revised Batch Zero population and annual capacity allocations for 2028–2032. A phased connection is different from immediate full load. The study allocates feasible capacity and upgrades; it does not itself create electricity.
Generation reaching construction, permits and firm operating dates would strengthen supply claims. Tested reductions at promised speed and duration would strengthen flexibility claims. Those assets and controls must remain maintained after approval, not merely pass once.
Texas's limited interconnection with neighboring grids increases the value of local supply and controlled demand, as well as the consequences of forecasting errors. The optimistic a16z case for technical solutions should be judged against those actual deployments. A smaller, credible pipeline would indicate queue repair; repeated unexplained delays would suggest an unresolved freeze. Neither outcome can be inferred from treating every requested gigawatt as inevitable AI capacity.