The city pause and county permit covered different ground. I mapped the six parcels RGV Business Journal linked to the project against the Census Bureau’s 2026 Harlingen boundary.

Caption: Harlingen’s 2026 Census boundary, as of Jan. 1, 2026, and six deed-reported parcels linked to the Fluidstack project. The nearest parcel edge is about 1,070 feet from the boundary. Parcel group and distance are approximate; no surveyed campus boundary is available.

All six parcels fall outside Harlingen’s 2026 Census boundary, beyond the territory covered by the city’s application pause. The appraisal records still list previous owners, so the parcel group is a proxy for the project footprint.

What the worksite permit covers

Cameron County’s Sept. 10 notice says its worksite permit allows site changes such as clearing and grading land, installing utilities and building surface infrastructure. The county expected site improvements to begin after issuance. It also says Fluidstack will need future permits before building construction.

The permit authorizes site work. Building permits, campus energization and a grid-service date remain unconfirmed in the public record. Fluidstack has said it expects to purchase electricity from AEP Texas, its transmission provider, and a private third-party company, and pay for its own substation.

Harlingen still entered the project’s local infrastructure discussion through stormwater. Fluidstack asked to route rainfall runoff into a city-owned drainage ditch. The city tabled the request in September pending a public town hall. On Oct. 21, commissioners denied a proposed drainage agreement with Cameron County and directed the city manager to negotiate directly with Fluidstack. The drainage request concerns rainfall runoff; the public record here gives no detail on cooling-water supply or campus power service.

Moratorium exposure and project delays

In September, SemiAnalysis mapped more than 300 U.S. data-center moratoriums. Its analysis estimated about 20 GW of proposed capacity inside active restrictions and classified 1,525 MW as delayed by them. One figure measures capacity inside a restricted area; the other measures capacity SemiAnalysis attributes to project delays.

ERCOT’s 12.2 GW figure is countywide

ERCOT’s September Batch Zero presentation lists 12,235 MW of large-load capacity in Cameron County’s 2032 base-plus-studied view. It combines projects in the base case with projects still under study. It covers large loads across the county and does not identify data centers or Fluidstack’s project status.

ERCOT does not publish Cameron County’s exact base-versus-studied split. Cameron County is absent from the five largest base-load county totals; the fifth-place figure is about 2,330 MW. Using that as an upper bound for Cameron County’s base component gives 12,235 minus 2,330, or at least 9,905 MW in the studied category. The 9,905 MW floor is our calculation from the public table. The table does not give an individual project’s status or expected service date.

What a month of earlier power is worth in one model

Occam Edge and Royal Powers Energy estimate about $410,000 per MW for one month of earlier service in their adjusted base case. Their model assumes $11 billion in annual revenue per GW of operating capacity, a 20-year operating life and a 7% discount rate. It adjusts the value to account for ramp-up, commissioning and tenant acceptance, producing a 60.8% adjusted value.

Occam models the present value of bringing a 20-year revenue stream forward by one month.

Applying the reported rate to the 9,905 MW lower bound gives about $4.06 billion; applying it to a rounded 10 GW gives about $4.1 billion.

The scenario applies Occam’s rate to every MW in the studied lower bound, assuming each receives a month of earlier service. It leaves actual advanceable capacity, local deal availability and Fluidstack’s share of any value unresolved. The scenario is separate from the reported $4 billion first-phase cost of the Fluidstack project. Occam’s piece describes the trade as a proposal and says there are no publicly disclosed implementations.

The next record

The next useful records are the project’s ERCOT study status and its expected service date, neither of which appears in ERCOT’s public Batch Zero materials.

Method note

The parcel group comes from six parcels RGV Business Journal linked to the project; appraisal records still show previous owners. The approximate 1,070-foot distance is measured from the nearest parcel edge to the 2026 Census boundary (EPSG:2279; closest parcel PROP_ID 424402, approximately 1,069 feet). ERCOT’s 12,235 MW figure covers countywide large loads. The 9,905 MW studied figure is a lower bound derived from the fifth-place base-load county total. The $4.06 billion figure applies Occam’s reported rate to that lower bound.

Sources and research

• Harlingen, moratorium record; KRGV, May 23 report on the 120-day pause and Aug. 25 report on a proposed extension.

• Cameron County, Fluidstack worksite permit notice, Sept. 10, 2026.

• RGV Business Journal, Fluidstack project and parcel reporting, Sept. 10, 2026.

• KRGV, Harlingen tables Fluidstack’s stormwater request, Sept. 3, 2026; City of Harlingen, Oct. 21 City Commission meeting recording.

• U.S. Census Bureau, 2026 TIGER/Line Shapefiles, legal boundaries as of Jan. 1, 2026.

• SemiAnalysis, “Everyone Says Datacenter Moratoriums Are Killing the US Buildout. We Mapped All 300 of Them”, Sept. 15, 2026.

• ERCOT, September 2026 Batch Zero Update, Item 14.

• Occam Edge and Royal Powers Energy, “Structuring Gigawatts”, Oct. 1, 2026.

A note on process. I use AI to help explore research, draft and revise writing, and create visuals. I verify the claims against their sources, and the final choices and any errors are mine.

I mapped 853 ERCOT price points to see where that difference sits. Rabbit Hill ranks fourth among 495 catchments. North Fork, also in Williamson County, averages $16.90 above its zone. Every Bell County catchment is below $2.50. The map puts the price boundary through the middle of Williamson County.

Rabbit Hill sits near the Salado–Hutto transmission corridor, between Bell and Williamson counties. The corridor’s ERCOT constraint bound in 9.8% of dispatch runs in the year ending October 2026, about 850 equivalent hours. The September 2026 ERCOT GIS screen found no battery projects queued at the modeled Rabbit Hill-area nodes. Williamson County has about 210 MW of operating batteries. Harris County listed 26 queued projects totaling 6.2 GW in the same snapshot; my Houston model shows much of its local price premium disappearing at that storage level. The Rabbit Hill screen leaves the gap unexplained, and private connection points may hide projects aimed nearby. One possible explanation is structural. ERCOT’s standard market products pay batteries for energy and ancillary services. Occam’s study assigns the value of earlier service to the data-center load, outside those standard battery revenue streams. Brattle’s report for Google and the Burleson study examine ways storage could help large loads connect sooner. A local agreement would have to direct some of that value to the battery owner.

Day-ahead spreads by price point, year ending Oct 2026. Rabbit Hill averaged $28.20/MWh above its load zone; River Valley, in the same county, averaged $7.60.

The original ten-project screen puts about 125 MW of reported capacity around Round Rock and Georgetown near the grid nodes most exposed to the line. A further roughly 1 GW appears in selected Hutto and Taylor records, with weaker modeled exposure to this constraint. Those are selected project records from the initial screen. June reporting counted 30 operating or proposed facilities across nine Williamson County campuses, with 1.6–1.9 GW possible at full buildout. The 125 MW estimate has not been rerun against the newer filings. The full project list, with sources, is in the table at the end.

Since that screen was assembled, Skybox filed Hutto Building 4, an 846,300-square-foot core-and-shell building scheduled for construction from January 2027 to October 2028. The building is part of the existing 600 MW PowerCampus plan, which keeps the campus figure at 600 MW. Taylor City Council approved a development agreement on Oct 8 for a separate 664-acre Taylor Technology Campus. Data Center Dynamics reports a campus designed for up to 1 GW, with a 500 MW first phase, a proposed 750 MW gas plant, and first power targeted for H1 2028. The city describes on-site generation and closed-loop cooling using reclaimed water. Taylor belongs in a separate power-path category from the grid-connected sites near Rabbit Hill.

ERCOT’s March 2026 regional planning update lists project 80546E, an upgrade to the Hutto and Round Rock-Salado 138 kV line, with a May 2027 estimated in-service date. Taylor’s first power is reported for H1 2028, so the schedules nearly touch. The planning list gives the upgrade’s estimated date. The PUC filing reports post-contingency thermal loading on the Salado Switch–Hutto line before and after the project. Battery dispatch and any effect on a data center’s service date fall outside those analyses (Oncor filing in PUC Docket 55999).

State policy is moving too. The governor’s Sept 21 directive asked TCEQ to halt permits sought by data centers pending ERCOT’s audit. The governor requested a compliance update by Oct 19. TCEQ told FracTracker on Oct 14 that it had paused relevant permits and reviews. A review published Oct 19 by FracTracker and Public Citizen found records of permit activity continuing after the directive, including public notices, stormwater approvals, and applications moving through review.

Could a battery near Rabbit Hill earn from price swings while helping nearby grid-connected loads get service sooner, before the planned line upgrade and new campuses arrive?

The second job

I backtested a 200 MW, four-hour battery at Rabbit Hill using each day’s four lowest and four highest day-ahead prices. The model estimates about $16.7 million in annual arbitrage revenue, compared with $9.8 million at the zone average. It assumes one cycle a day and 85% round-trip efficiency. At that rate, an 80 MW battery models to about $6.7 million a year.

A battery could also relieve a constraint that limits service to a large new load, potentially helping it connect sooner. Occam Edge and Royal Powers Energy, in Structuring Gigawatts, estimate $41 million of adjusted value for one month of earlier service to a 100 MW tranche, based on their revenue, ramp-up, commissioning, and tenant-acceptance assumptions.

At those modeled rates, $41 million equals a little over six years of an 80 MW battery’s trading revenue. The estimates describe different parties: earlier-service value to the data-center load and trading revenue to the battery owner.

The timing matters. Over the past year, 81% of the hours when Salado–Hutto bound fell between 9 a.m. and 6 p.m., peaking around 3 p.m. Texas batteries often charge in those hours, when solar makes power cheap. At Rabbit Hill, congestion raises the local price at the same time, which already nudges a battery to discharge instead. A four-hour battery would cover about three-quarters of the observed binding time. The longest episode ran 25.5 hours, which no four-hour battery covers. I compared it with the observed timing; I haven’t simulated battery dispatch. A utility counting on relief would need to know when storage can charge, discharge, and recharge against the constraint.

ERCOT’s binding record shows when the constraint affected dispatch. A local engineering study would need to establish whether that limits service to nearby data centers, how much dependable relief storage could provide, and whether it would move a service date.

The Burleson comparison

In Flexibility without Curtailment, Piq Energy and Base Power modeled a hypothetical 100 MW data center at Burleson Switch. Their study found that roughly 80 MW of targeted home batteries could alleviate constraint violations under the contingencies they evaluated. Their 80 MW result describes the Burleson substations, load, and contingencies. Salado–Hutto needs its own engineering analysis.

Brattle’s report for Google describes a possible flexible-service arrangement with a defined load, dispatch triggers, linked studies of the load and supporting resources, and utility controls.

The map changes the battery question

Texas had about 4 GW of operating batteries at the end of 2023 and 18.4 GW by August 2026, according to EIA.

For local officials, countywide project totals can hide different power paths for nearby parcels.

For a battery developer, the next number to price is how much dependable relief a project could provide and how many months it could bring a nearby data center’s service forward.

The projects behind the line

The initial ten-project screen, sorted by reported size. MW figures are reported maximums; operating load is unknown. Exposure uses the nearest priced node’s response to Salado–Hutto congestion. Taylor Technology Campus is tracked separately because its approved plan uses on-site generation.

ProjectCityMW (reported)StatusExposure to Salado–HuttoSourcePowerCampus Austin (Skybox + Prologis)Huttoup to 600 (campus plan)Building 1 operating; Building 4 construction 2027–2028WeakDCD, ConnectCREKDC Project ComalTaylor360 (trade press)Annexation and zoning approved Mar 26, 2026WeakCity of Taylor, Spectrum NewsSkybox Round RockRound Rockup to 75, grid-connectedRezoning approved Feb 12, 2026DirectCommunity ImpactBlueprint TaylorTaylor60 (30 MW Oncor agreement reported)Planned, phases 2025–2027WeakConverge DigestBlueprint GeorgetownGeorgetown25PlannedDirectConverge DigestSwitch AUS05Round Rock25 (one listing)Under construction, 2028DirectBaxtelColovore Project RaptorHuttonot publicApproved; completion Dec 2028 reportedDirectConnectCREIron Mountain Taylor MeadowsHutto areanot publicProposed; grid buildings around 2030WeakBaxtelProject Sequel (Skybox + Prologis)Huttonot known until 2028RezonedUnknownCRE DailyZydeco Project HuttoHutto70Withdrawn Apr 17, 2026Direct (withdrawn)KUT, City of Hutto

Directly exposed, excluding the withdrawn project: about 125 MW plus Colovore. Weakly exposed: about 1,020 MW plus Iron Mountain.

Texas operating batteries reached 18.4 GW by August 2026. Williamson County has about 210 MW operating; the September screen found none queued at the Rabbit Hill-area nodes.

What is measured, reported and estimated

• Measured: congestion hours, months and timing come from ERCOT’s published binding-constraint records. Prices come from ERCOT day-ahead settlement points. Battery capacity comes from EIA’s generator inventory and ERCOT’s interconnection report.

• Reported: every project MW figure comes from developers, city records or trade press. Several rest on one source (Project Comal, Skybox Round Rock, Switch). Only Blueprint Taylor reports a utility power agreement (30 MW with Oncor).

• Estimated: exposure uses each site’s nearest priced grid node and price-based estimates of how that node reacts to the line. Actual connection points are private, and large campuses with their own substations may connect elsewhere. The battery backtest is a model on past prices.

• Ceilings: proposed MW is the most a project could draw. Approvals, power allocation and ERCOT’s large-load review will decide what connects, and when.

Sources and research

A note on process

I use AI to help explore research, draft and revise writing, and create visuals. I verify the claims against their sources, and the final choices and any errors are mine.

• Occam Edge and Royal Powers Energy, “Structuring Gigawatts”,

Oct. 1, 2026. Source for the cost-of-earliness scenario and flexible-load economics.

• Piq Energy and Base Power, “Flexibility without Curtailment”, Sept. 2026. Hypothetical Burleson Switch load case and aggregate battery dispatch study. Base Power’s summary.

• The Brattle Group, “Solutions for Timely Interconnection of Large Loads: Enabling Non-Firm Transmission Service”, prepared for Google, Sept. 2026.

• ERCOT, DAM Settlement Point Prices (NP4-190-CD), source for node and load-zone day-ahead prices.

• ERCOT, SCED Shadow Prices and Binding Transmission Constraints (NP6-86-CD), source for constraint-binding records.

• ERCOT, Generation Interconnection Status (GIS) Report (PG7-200-ER), source for queued battery projects.

• ERCOT, March 2026 Oncor and LCRA TSC regional planning update, project 80546E, estimated May 2027 in-service date.

• City of Taylor, Taylor Technology Campus agreement and project conditions, including on-site generation and reclaimed-water cooling.

• Data Center Dynamics, PowerHouse and Big Watt secure approval for a 1 GW behind-the-meter Taylor campus, source for reported campus capacity and phase, gas-plant capacity, and first-power target.

• Data Center Dynamics, Skybox and Prologis file Hutto Building 4, source for building area and construction schedule.

• U.S. EIA, Preliminary Monthly Electric Generator Inventory (EIA-860M), source for operating battery capacity.

• Office of the Texas Governor, directive to TCEQ on data-center permits, Sept 21, 2026.

• Austin Free Press, Williamson County data-center inventory and buildout estimate, source for the 30 facilities, nine campuses and 1.6–1.9 GW.

• Project records: links in the project table above (City of Taylor, City of Hutto, Community Impact, Spectrum News, KUT, CRE Daily, DCD, ConnectCRE, Converge Digest, Baxtel).

Abbott moved first. On September 14 he directed the Texas Water Development Board to start enforcing water reporting against data centers that had been ignoring it, saying data centers must share the duty to protect Texas water. Ten days later Paxton opened an investigation into “hundreds of data center developments”.

Paxton is also running for Senate against James Talarico, and the eighteen counties cover every major media market in the state. Read into that what you like. The underlying problem predates the campaign by three years.

I spent the weekend trying to get from those eighteen counties down to specific sites. Three filters:

• where the water is stressed

• where new development is planned

• which of those sites are already energized

I expected drought to do most of the cutting. It did almost none. North and Central Texas has dried out fast since early summer, and moderate drought now covers nearly the entire 46-county forecast area around Dallas and Fort Worth. Dallas went from July 15 to September 15 without measurable rain, against a normal of more than four and a half inches. Wichita County hit exceptional drought on October 1. Screening Texas counties for water stress in October 2026 is close to screening for counties in Texas.

Which left energization.

What the state has

The Texas Water Development Board has been mailing a mandatory water use survey to data centers since 2023. The response rate has been flat and bad the whole time. In 2023 the board surveyed 22 facilities and 32 percent answered. In 2024 it surveyed 67 and got 28 percent. In 2025 it surveyed 329 and got just under 30 percent. Among the 267 facilities new to the board’s tracking that year, 22.5 percent answered. These are statewide figures, not county ones.

At a House Natural Resources hearing on June 23, lawmakers were told the number was 17 percent. Temple McKinnon, who runs water supply planning at the board, noted that a response is required for state financial assistance and for certain environmental permits, which is to say the incentive to file already exists and most facilities are declining it anyway.

The denominator keeps growing. Texas Tribune research counts at least 596 operating and planned data centers across the state as of September, against the 329 the board surveyed.

Authority over this sits with the state. Hood County tried in February, voting 3 to 2 against a pause on data center development, which would have been among the first such measures in Texas. Before the vote, Senator Paul Bettencourt, who chairs the Senate Committee on Local Government, sent Paxton a letter warning that counties hold no constitutional or statutory authority to impose development moratoriums.

So the pressure has nowhere to go but the state, and the state’s instrument is a mail survey. The Public Utility Commission ran its own data center survey alongside the board’s, except theirs is voluntary, and agency officials said there is no incentive for operators to submit anything. One agency can compel a response and gets 30 percent. The other asks politely and gets less.

Abbott’s directive and Paxton’s investigation are the first real pressure applied to any of it. The state’s picture still resolves to the county line and stops there.

Looking for energization dates

Drought is a county-level fact. Planned development is close to one, since an announcement comes with a county and sometimes a town. Energization is a site fact.

It also sorts by stage, which a county list flattens. A facility that has been announced may hold land, may hold a filed interconnection request, may hold nothing beyond a press release and an option. A facility that is energized has power on site and has been running long enough to produce the monthly figures the survey asks for. The first kind may still be years from its first reportable month. The second kind has twelve of them on the books already.

A facility that is energized has power on site and has been running long enough to produce the monthly figures the survey asks for. The first kind may still be years from its first reportable month. The second kind has twelve of them on the books already.

ERCOT sorts large loads into five tiers: Observed Energized, Approved to Energize but Not Operational, Planning Studies Approved, Under ERCOT Review, and No Studies Submitted.

Those tiers are published in aggregate. NPRR1267, the protocol revision that created the report, requires ERCOT to publish monthly and to aggregate requests “across multiple dimensions with as much specificity as possible while maintaining the confidentiality of Customer data.” The public series runs from September 2022 through June 2026, 831 cells, each one megawatts by load zone, transmission provider and size bucket. Any category with fewer than five customers gets rolled up with its neighbors.

The April 2026 interconnection update shows what per-project disclosure would look like. Eighteen rows, LLI-9001 through LLI-9018, transmission provider, megawatts, and a Batch Zero eligibility column. The owner names are Big Data, ABC, XYZ, Project 1. The slide notes it as a precursor to the Large Load Portal, still in development.

Four other routes, same outcome. ERCOT’s monthly generation workbook covers solar, wind and storage. Settlement data resolves to load zones. The PUC dockets I found set interconnection standards and batch study procedure rather than site agreements. TCEQ air permits carry an address and a received date, though a backup generator gets permitted well before a building energizes.

Across nine named projects inside the eighteen counties, I have zero dated energization statuses. The closest I got: Cipher has stated a 300 MW interconnection at Barber Lake, with no approval-to-energize date attached. Abernathy has a reported 345 kV tie and no megawatt figure. A September ERCOT deck lists Haskell County among the top base-load counties at 2,689 MW, which is a county total under a provisional classification.

Water use runs through a mandatory survey that roughly 30 percent of facilities return. Power runs through a mandatory interconnection process that every large load completes. The water figure goes missing when operators decline to file. The power figure is collected, sorted into five tiers, then aggregated past the point where it can name a building, under a protocol that lists customer confidentiality as a design requirement.

Paxton is asking operators to self-report the resource the state cannot independently measure, while the resource it does measure sits behind a rollup. If a facility in Ector County reports nothing, there is no second record to check it against.

ERCOT’s Batch Zero process under PGRR145 has paused energization of new large-load data centers pending a December audit, with roughly 8,927 MW already holding Approval to Energize as of the June Large Load Working Group update. That audit will go to site level. Whether any of it reaches the public is a separate question.

The cooling trade

The standing argument from the people building these campuses (hyperscalers and neoclouds mostly) is that the water problem is handled. Closed internal loop, water gets reused, done.

The loop is very real but the electricity it takes to run one shows up on a different bill.

Cooling is a trade between two resources. Evaporative systems shed heat by spending water and use less electricity doing it. Dry and closed-loop systems spend electricity instead and use very little water. Same heat leaving the building, different resource paying for it. Google has committed to closed-loop air cooling on its new builds, which is the trade being made deliberately and announced.

The size of that trade is smaller than I assumed and lands somewhere stranger. Evaporative cooling improves PUE by 15 to 30 percent over mechanical cooling in suitable climates, and evaporative assistance cuts peak hourly PUE by roughly 10 to 30 percent against purely dry cooling. Call it 1.2 PUE evaporative against 1.35 to 1.5 dry in a Texas summer, a range consistent with modeling work that treats PUE 1.2 as the efficient water-cooled baseline. For a 100 MW facility that works out to 15 to 25 percent more total energy, somewhere around $8M to $10M a year.

That same 100 MW facility running evaporative cooling consumes roughly 1.9 billion liters a year, close to 500 million gallons, at the 1.8 L/kWh industry average. Dry cooling eliminates nearly all of it. Which prices the swap at roughly $16 to $20 per thousand gallons avoided, against industrial water rates that run a small fraction of that in most Texas municipalities. The arithmetic has pointed one direction for years and operators have followed it. “Solved” describes a decision to pay several times the going rate for water, and nobody makes that decision unless something forces it.

The interconnection is where this gets expensive. A utility service limit caps total facility power, and cooling and compute draw from the same number. At a 100 MW service limit, PUE 1.2 leaves 83 MW for IT equipment. PUE 1.5 leaves 67 MW. Everything else about the building is identical: same land, same shell, same substation, same transformers, same three-year interconnection wait. You paid for 100 MW of infrastructure and can sell 67.

That 16 MW lands on the revenue line. It recurs every month for the life of the asset and scales with the lease rate instead of with equipment prices. At $50 per kW per month it runs $9.6M a year. At $120 it runs $23M. The research frames the same point from the other side, noting that evaporative assistance releases 10 to 30 percent more power capacity for compute under the same site power budget.

And it concentrates. Dry coolers reject heat to ambient air, so their penalty tracks outdoor temperature. In January the two systems nearly converge. In August the gap is widest, which is also when ERCOT scarcity pricing hits and GPU demand peaks. The capacity that disappears is the capacity that was worth the most.

There is one escape, which is to decide early. Size the interconnection at 125 MW instead of 100 and dry-cool the full load. That path is mostly closed right now given paused interconnection requests and multi-year queues. The megawatts requested in 2024 are the megawatts available in 2027.

Closed loops still drink. Meta’s Indiana campus runs closed-loop liquid cooling and still requires around 8 million gallons a day at full buildout, because the heat still has to leave the site once it leaves the chips.

Which system a given site chose stays inside the design documents. A facility in Ector County and a facility in Harris County can carry identical load and wildly different water draw, with the permit trail reading the same for both. It does leave traces. Load profile against installed capacity narrows the range, since the two designs put different shares of total draw into cooling. Whether there is any water to withdraw nearby narrows it further. A high-draw site with no withdrawal permit and no utility connection within reach is running dry by necessity. A moderate-draw site sitting on supply is the one worth a phone call.

Each piece here is circumstantial. But together they provide an estimate, assembled from records that were already public, which is more than the state currently has (or willing to share with the public). Paxton is enforcing a filing requirement and the filings will eventually arrive.

Which sites were asked first can be narrowed now, from load data and water infrastructure, without waiting for anyone to fill in a form.

What I am building next

DOE keeps a library of model buildings it uses to test energy codes before they become law. Sixteen commercial types, from small offices to hospitals to warehouses, each built in EnergyPlus and run across nineteen climate locations. 5,168 models in the current set. One of those nineteen representative cities is El Paso, which is also one of Paxton’s eighteen counties.

The sixteen types cover offices, schools, hospitals, warehouses. Data centers arrived separately, in 2020, when a team at Lawrence Berkeley added two prototype models: a small one representing a computer room served by CRACs, and a large one representing a stand-alone facility served by CRAHs with a central chiller plant. Good work, built with input from the ASHRAE 90.1 committee and implemented in OpenStudio-Standards. It is also six years old, and the large model describes a building that would now be a rounding error on a Texas campus.

So I am building the scaled version. Same method, same climate data, sized to what is actually under construction here, with evaporative and dry variants of each so the trade shows up as hourly output.

That gives a figure I have yet to find published: what a hyperscale campus should be drawing in water and power, by climate zone, measured against what these sites report.

Until then, the honest answer to how much water a Texas data center uses is that the state is asking the operators, and the operators are the only ones who know.

Four counties carry the headline

The current Cleanview snapshot assigns roughly 35 GW of planned capacity to four counties with zero operating MW: Carson at 17.0 GW, Hill at 7.2 GW, Pecos at 6.9 GW, and Caldwell at 4.0 GW. More than one-third of the statewide planned total sits in those four counties. Caldwell and Pecos sit at opposite ends of the transmission map. Caldwell has a 345 kV line within a mile of its mapped sites. Pecos has a comparable line crossing the county: LCRA’s North McCamey to Sand Lake line, energized in July 2026. Both are at zero operating MW, but the zero means something different in each place.

A line gets you to the site

Caldwell has the line, a signed agreement with Bluebonnet Electric Cooperative, county incentives, and public support from state and local officials. It is still at zero because energization is scheduled for Q1 2028 and no customer has been confirmed. The missing pieces are the customer and the calendar. Pecos takes a different route: Amazon and Pacifico’s GW Ranch, along with the nearby Prometheus project, are building off-grid on purpose. The projects have named customers, air permits, construction schedules, and first power targeted for 2027. The move is a direct response to the uncertainty in ERCOT’s interconnection queue.

Three counties, three clocks

Hill County is the policy case. The county imposed a moratorium in May, faced a $100 million lawsuit, and rescinded the moratorium in June. In August, the governor ordered a statewide audit affecting up to 300 projects across Texas that were still in the ERCOT interconnection queue. Any Hill project waiting on a conventional grid connection is running on that clock now.

Carson is a different kind of number. Its 17 GW total is heavily shaped by Fermi Matador records, including gas and nuclear entries. I am keeping the figure in the dataset comparison because it is in the tracker. Treating it as 17 GW of conventional data-center load requires checking each entry more closely.

The zero column is doing too much work. Pecos is building around the grid, Caldwell is waiting for a customer and a date, and Hill is caught in a state policy freeze.

Several clocks are running at once, and the statewide total collapses them into one number.

Method note

The statewide starting figures come from Cleanview’s Texas tracker, updated September 2026. The tracker lists 348 planned projects and 137 operating projects, with 97,151 MW planned and 7,072 MW operating. The map contains 465 geocoded records, while 20 additional records do not have map coordinates.

County totals are grouped from the map’s records by county and stated MW. They should be read as a snapshot of the source dataset, not as an independently verified state total. Planned capacity is not the same as approved, financed, interconnected, under construction, or likely to operate.

It is building three paths, and on August 28, 2026, they stopped moving together.

Two Oncor segments advanced:

Dinosaur–Longshore.
Longshore–Drill Hole.

The Public Utility Commission of Texas deferred a decision on Howard–Solstice.

The Bell County East–Big Hill–Sand Lake path remains unresolved in separate proceedings.

All three came out of the Permian Basin Reliability Plan required by the 2023 legislative mandate. They are related, but they are not interchangeable.

This is the same map I published a year ago, updated with what happened next.

My old version compressed the system into one corridor. That hid the important part: three infrastructure bets now have three different timelines.

The approved path runs through Abilene, where Crusoe’s Lancium Clean Campus is now running Microsoft workloads, with roughly 1.2 GW planned across the site.

AWS’s GW Ranch and Microsoft’s other West Texas campuses sit closest to Howard–Solstice, the path whose commission decision was deferred.

Detail: the 640 MW and 600 MW labels are mapped load signals within the broader roughly 1.2 GW planned campus, not an additional total.

Same basin. Same reliability plan. Different power path.

That is the discovery.

A regulatory status change is becoming a site-selection variable. One cluster can point to a path that has advanced through the state process. Another is still underwriting a route whose decision has moved out. A third is waiting on a separate proceeding before the path can be treated as settled.

The power problem predates AI.

A 2022 study commissioned by Chevron, ConocoPhillips, and ExxonMobil projected that Permian oil and gas could need 17.2 GW by 2032. The grid had 4.2 GW.

Data centers did not create that deficit. They arrived with much larger loads and shorter delivery clocks.

That is why transmission status belongs in site selection.

“Near a transmission line” is not enough. The useful questions are:

• Which path serves the site?

• What has the commission actually approved?

• What remains a filing, recommendation, or open proceeding?

• Who owns and builds the line?

• What does the interconnection agreement really secure?

• If the line is delayed, is onsite generation a bridge or part of the permanent design?

A credible power path is an asset.

A proposed path is an option.

A deferred path carries a different timeline, cost, and underwriting problem.

The map does not say Howard–Solstice will fail. It does not make every site on the approved path deliverable. It shows that these projects can no longer be read as if the infrastructure behind them had one timetable.

The question is no longer only who announced the largest campus in West Texas.

It is who has a power path with a decision behind it.

I am still looking for a better public read on the Howard–Solstice easement maps, especially in Val Verde County. If you were involved in that proceeding, or have public material I have missed, I would like to compare notes.

Sources and method

The three-path structure comes from ERCOT’s Permian Basin reliability work and the associated transmission filings. The relevant PUCT matters include Project 55718 and CCN Dockets 59315, 59029, 59336, 59475, and 59182. The August 28 action advanced the two Oncor segments, while the Howard–Solstice decision was deferred and the Bell East–Big Hill–Sand Lake segments remained in separate proceedings.

The data-center locations and company relationships are based on public announcements, utility filings, and the map’s project records. “Roughly 1.2 GW planned” refers to the broader Lancium Clean Campus site, not a claim that all of that load is currently operating.

ERCOT independent review of the three-path project.

The biggest power consumers in the state are going up nowhere near the demand they used to sit near. These are huge loads, gigawatt-scale, and they’re going up far from the cities.

I’ve counted at least 360 large compute sites so far. 126 of them sit out west, next to the wind farms, in places like Armstrong County: population under 1,800, spread across a county the size of Rhode Island. These sites draw a city’s worth of power, with no city in sight. The load Texas spent decades routing toward Dallas and Houston is landing on ranchland instead. There’s cattle, a couple of county roads, and a fenced-off campus pulling half a gigawatt off the grid.

It makes sense on paper. Texas spent the late 2000s and early 2010s building CREZ, a $7 billion, 3,600-mile transmission backbone the state built specifically to haul West Texas wind power east to Dallas, Austin, and Houston, where the people are. If you’re building a data center now, and power is your biggest cost, the smart move is to skip the wire entirely. Go stand next to the wind farm instead of paying to ship the electrons east.

This isn’t the first time Texas, or power companies more broadly, went after big consumers to make a grid pay for itself. A century ago, Samuel Insull built Chicago’s electric utility by signing up streetcars and factories specifically because their demand hit at different hours than home lighting, filling in the valleys and letting him build bigger plants more cheaply. Data centers looked like a modern version of that: steady, 24/7 load, the kind that should smooth a system out. Except these loads don’t stay steady when the system needs them to. They’re first to be cut.

Works fine most of the year, too. Then February comes.

Not much solar output before 8am in the winter. Turbines ice over in a hard freeze. The blades can’t spin, and de-icing them isn’t fast. The gas fleet that could cover the gap is sitting four hundred miles away, already committed to the cities it was built to serve. So the capacity that shows up on paper, the capacity everyone modeled around, isn’t actually there when it’s needed.

This isn’t a hypothetical February. NOAA’s Climate Prediction Center put the odds of a very strong El Niño this winter above 90% in its August update, which would make it one of the strongest events since 1950. That doesn’t mean a repeat of 2021’s Arctic outbreak. El Niño winters usually run close to normal in Texas, sometimes even a touch warmer. What it does mean is a wetter season and a more active storm track, and forecasters are already flagging the exact combination that catches wind and solar off guard: a cold rain that turns to ice with only a small shift in temperature. You don’t need a deep freeze for turbines to ice over. A storm with the wrong few degrees attached is enough.

When supply runs short in a freeze like that, somebody has to be dropped, and ERCOT has decided who. Big industrial loads, including data centers, go first, before residential load gets touched. That’s the law now, and honestly, it’s the right call. Texas’s official count from February 2021 is 246 deaths, mostly hypothermia. Researchers who compare how many more people died that month to a normal February put the real number well above that, since the official count only includes deaths a coroner directly tied to the storm. Before this rule existed, the only lever ERCOT had left was cutting power to homes. Nobody wants to run that experiment twice.

So if you’re building a data center in Texas right now, the obvious move is to protect yourself from that shed order. Build your own generation. Sit behind the meter, next to a wind farm you built or contracted, and when ERCOT needs to cut load, you’re covered. You’re not drawing from the grid, you’re drawing from your own turbines.

Except that’s not how it’s playing out.

Last month, PUCT ruled on the first case under a law called SB6, and I think it’s going to define how this works going forward. There’s a 525 MW data center campus, Crusoe’s, run for a customer called Ensign, sitting right behind its own 265 MW wind farm in Armstrong County. PUCT told them to drop the entire load, all 525 MW, within thirty minutes of an ERCOT emergency signal. Not the portion the grid was covering. Not the gap between what the wind farm produces and what the campus needs. All of it. Breaker open, physical disconnection if necessary, and no pay for the privilege. The order also bars the site from paid demand-response programs.

I read through coverage of the order a few times because I wanted to make sure I wasn’t missing something. The logic is pretty simple. This campus kept its interconnection to the grid, kept the wire, kept the meter, because a 265 MW wind farm can’t run a 525 MW campus through an ice storm. Wind alone doesn’t cover the load on the exact day the load needs covering most. So the campus stayed a grid customer. And customers get shed. Owning generation next door doesn’t change that, not under SB6, not according to PUCT.

Which means the workaround doesn’t actually work. You can build your own wind farm, spend the money, do everything right, and PUCT can still open your breaker and drop 525 megawatts in half an hour, same as if you’d never built anything out there at all.

The data centers losing power for thirty minutes isn’t the real story. The real story is who pays for the wires.

Since the first power plants went up in the early 1900s, Texas built wires to go where the demand was. Houston needed power, so power plants and transmission lines got built to serve Houston. Now that’s flipped. ERCOT’s own queue shows the shift in scale: hundreds of gigawatts of large-load requests, nearly 90% of it data centers, most of it chasing cheap power out west instead of the other way around. The demand is moving out to where the generation already exists, rather than the generation getting built to where the demand is. That should mean less new transmission. Skip the four-hundred-mile line. Build next to the generation instead.

Except SB6 and the Armstrong County ruling mean these sites can’t actually leave the grid, even when they build their own power. They have to keep the interconnection for the days their own generation doesn’t cover the load, which, in Texas, in February, is exactly the day it matters most. So the grid still has to be built and maintained out to wherever these sites land, plus enough headroom to actually serve them during a freeze on the days their private wind farm is iced over.

What this does to ERCOT’s long-term planning is still an open question, and not just for me. It could make the state more conservative about new interconnections in far-west counties, since PUCT has now signaled these loads can’t be trusted to disappear during the exact hours the grid is stressed. It could also change how these campuses get financed. Lenders are going to want to know what “you might get cut off with thirty minutes notice, unpaid” does to a facility’s uptime guarantees.

One thing is sure, regardless of how the open questions shake out: the wires get built either way. West, to serve the compute campuses that moved out to the wind, instead of east, to serve cities. And ratepayers, the households ERCOT is protecting by shedding these big loads first, are the ones who cover the cost of the wires in both directions. The data center didn’t get to skip the grid. It just moved where the grid has to reach.

I get it. Cleanrooms want to be big boxes. A fab floor has to be contiguous and vibration-damped. A wafer can’t cross a seam. Under it, by my math, they need something like 8,700 miles of process piping.

That geometry isn’t a style choice. It’s everything around it that we don’t yet see.

Drawn in pieces.

It isn’t true that nobody is planning anything. In January 2022, a local investor group bought 4,860 of the site’s 6,166 acres. This includes the reservoir, the dam and the spillway with the stated intent of future residential development. That was four years ago. There has never been a public plan.

And there will be more. Much more. Developers will plan tracts. Utilities will plan easements. The school districts are already at the table negotiating tax abatements. Every one of those parties will do competent work inside their own box.

So what does this place look like when all is said and done?

The county

Grimes County is located about an hour northwest of Houston. Its 2025 population estimate is around 34,000. Navasota is the largest town with 10,000 residents. Anderson, the county seat, has fewer people than most suburban high schools have students.

Insert a construction program running for years, then at least 3,000 permanent jobs, followed by a second-order economy of suppliers, contractors, and the people who feed and house and teach all of them.

As one developer put it during the Intel buildout in Ohio: every job you create comes with a commute and a housing unit. That’s math, not opinion.

Why no one can convene it

Texas counties do not have zoning authority. A county can review subdivision plats, issue septic permits, regulate floodplain and abate nuisances. It cannot say this parcel is a neighborhood and that one is a truck yard. Regulatory power in unincorporated Texas is narrow by statute.

So the coordination problem isn’t a failure of will. There is just nobody with the authority to do anything. The county can’t. Navasota’s extraterritorial jurisdiction reaches only so far. The groundwater district covers water and nothing else. The school districts control just the schools.

Someone raised this with me on LinkedIn recently, in the context of digital twins, the idea being that you could model the whole thing at once and actually see how the pieces fit. The technology isn’t the obstacle. The obstacle is that every actor controls one piece and only one piece. The fab company builds the fab. The developer plats the tracts. The county issues the permits. None of them has any reason to model the whole, because none of them owns the whole. That’s not a Grimes County failure. It’s the default condition of unincorporated land: everyone optimizes their parcel, and nobody is responsible for the place.

The clock is ticking

This land is a decommissioned coal plant. Gibbons Creek Steam Electric Station ran from 1982 until 2018. In early 2021 Charah Solutions acquired the roughly 6,166-acre site, which includes the 3,500-acre reservoir, its dam and spillway, and took on the remediation obligation. By October 2021 they had imploded the plant and begun cleaning up the coal ash, working with the TCEQ on the environmental remediation the property requires. TMPA’s adjacent lignite mine, a separate asset, is undergoing its own reclamation.

The contamination is real and documented. A 2019 analysis of federally required monitoring data found unsafe groundwater contaminants under sixteen Texas coal plants, and Gibbons Creek was named among the sites exceeding federal health benchmarks.

Under the state’s Voluntary Cleanup Program, a site is cleaned to Texas Risk Reduction Program standards according to the applicant’s desired closure strategy — and a certificate of completion then releases future landowners and lenders from state liability. So the cleanup standard depends on what the owner intends to build.

Of course, residential closure is a stricter and more expensive bar than industrial. Which means the landuse decision is not downstream of remediation. It is part of the remediation process. This decision is being made right now, parcel by parcel, by private parties, in a county with no zoning authority to check it. Whatever gets certified to industrial standard is functionally spoken for. Housing can’t go there later without someone paying to reopen it.

There is a window on where a town can legally exist, and it is closing.

The anchor might not hold

A coal plant sat on that land for thirty-six years.

That history is worth remembering, because SpaceX’s own IPO paperwork described Terafab as a general framework, with no binding commitments and no obligation for either side to continue with the project. Ohio ran this experiment. Intel announced its New Albany project in 2022, then slowed construction and threatened to halt it. Developers who had bought land for workforce housing began reconsidering their plans, while school districts pulled back expansion plans. The counties had already spent money on infrastructure. Brookings has documented that the jurisdictions absorbing most of that growth had no role in the siting decision.

There are more than a few rural communities in Texas where hyperscalers are buying up the land around them. Those communities aren’t necessarily opposed to the development. They just want to know that, in five years, there will be something left besides a fence line and a tax bill. That’s the right expectation, and it has a design answer.

A town with its own reason to exist can survive the anchor changing its mind. Speculative subdivisions can’t. Planning for a place isn’t the romantic option here. It’s the only way this works over the long term.

They used to do this on purpose

When the United States needed a secret industrial complex in eastern Tennessee during World War II, it did not build only the plants. The Army hired Skidmore, Owings & Merrill to help plan Oak Ridge as a town alongside them: housing, schools, shops, and a town center.

The first plan assumed roughly 13,000 residents. By 1945, the population had reached 75,000—just three years after the site had been rural farmland. Oak Ridge had real problems. It was segregated by design, paternalistic, and company-run. I’m not proposing that we repeat its governance. But the planning competence was real, and it was recent.

The assumption that a giant industrial facility should arrive with a town attached was, within living memory, ordinary. We didn’t lose the ability. We lost the assumption.

The dozen

Last Sunday morning, while getting the toddler ready, I asked an AI to draw the alternative: a town on this site. The same land. The real reservoir, parcels, roads, and power nodes. Not one answer. A dozen. Some denser, some looser—wrapped around the water, strung along the roads.

The toddler took longer choosing a bathing suit than it took to generate twelve master plans. These plans incorporated some of the constraint layers that actually govern the site: the ash landfill, the cap boundaries, the parcels likely headed for industrial closure, and the probable locations of monitoring wells.

Make no mistake: These are not real site plans. This is a quick experiment designed to show scale, speed, and spatial relationships. A town. A place where people might live. But drawing was a gate for a long time.

A credible alternative master plan used to require a design firm, a fee, and two months—which meant it was only made after someone with money had already decided the question was worth asking.

That sequencing is what broke. The alternative can now exist before the decision instead of after it. Cheap enough to be speculative, specific enough to argue with. That’s a much smaller claim than “AI will design our cities.” It’s also the more useful one. A rough map that people argue over beats a good model nobody commissioned.

Who draws the rest

There’s an answer, and it’s boring, which is why it never makes the press release. The commissioners court. The school districts. The groundwater conservation district. Navasota’s ETJ. The remediation applicant choosing the closure standards. The investor group that bought the reservoir four years ago. Whichever developer files the first plat. Most of these decisions will be made over the next twenty-four months, before the fab is finished. After that, the land is spoken for and the pattern is set for the next fifty years. So far, the box is the only part anyone is drawing as one thing. Who draws the rest?

That’s probably how most people are reading the Texas data-center pipeline right now. Impressive if you count the blocks. Less impressive if you ask how many get built. I think this is right about the blocks and wrong about what happens next.

What actually happened

On August 3, Abbott directed the PUCT and ERCOT to run a comprehensive verification and audit of every data center moving through the interconnection process, and paused new interconnections until it’s done. The letter says ERCOT is looking at roughly 474 GW of connection requests, about 90% of it data centers, which is more than five times the grid’s record peak demand.

Texas already had a framework for this. SB6 passed in 2025. The PUCT adopted 16 TAC §25.370 in February and §25.205 in March. Batch Zero, built through NPRR1325 and PGRR145, got PUCT approval on June 18 and replaced the old project-by-project queue with a structured batch study.

None of that is new. What’s new is the consequence. This is the first time the state is trying to make the disclosures determine whether a project gets power at all. An announcement is still free. What it no longer buys is a place in line.

Announcements were always free

For two years, announcing a Texas data center cost basically nothing. A press release, a rendering, a gigawatt number, a capital figure at a scale nobody can really picture. The incentives all pointed one way. It moves the stock, it stakes a claim, it signals seriousness to investors and to counties competing to host these things.

None of it required the project to be real. You could announce a gigawatt campus with a site you didn’t control, power you hadn’t secured, and water math nobody had checked. Plenty of people did.

So the queue got enormous. The funnel underneath it did not. ERCOT’s March TAC report shows about 9,042 MW with approval to energize and observed peak large-load consumption around 3,883 MW. Set that against a queue in the hundreds of gigawatts and you get a request-to-operating conversion of roughly 1.6%.

To be fair, queues always look like this. ERCOT’s generation queue has carried similar attrition for years and nobody calls it fiction. That’s fair. The difference is that generation queues cull themselves quietly over three or four years and nobody certifies the survivors.

Second, most of this was going to wash out anyway. Also fair, and it’s why “the audit kills projects” is not much of a prediction. It’s half trivial and half unfalsifiable. The interesting question is what the audit creates.

The $50K we MW deposit was never a filter

People point to the ERCOT security deposit, roughly $50,000 per megawatt, as the thing that separates real from fake. A gigawatt project posts $50 million. That sounds like a lot until you look at what money is actually doing in this sector.

In November 2025, Anthropic announced $50 billion in American compute infrastructure with Fluidstack, with the first sites in Texas and New York. In June 2026, Apollo and Blackstone closed a $35 billion private credit package to buy Google TPUs and lease them to Anthropic. Reports say a second facility of similar size is already in discussion.

Against that, $50 million is just a small line item. Expect the size isn’t the interesting part. The financial structure is.

Lenders in that deal did not underwrite Anthropic’s ability to build anything. Broadcom covers the residual-value shortfall on the senior tranches if the chips have to be sold for less than the notes are owed, which effectively swaps in Broadcom’s investment-grade credit for an unrated startup’s.

Google guarantees the rent on the data center side if Fluidstack stops paying. Pension funds and insurers can hold the paper because somebody else’s balance sheet answered the credit question.

Capital markets stopped testing whether the developer is real at almost exactly the moment Texas started.

I know what you think. Big claim, but one buildout is not a market, and Anthropic is not the median Texas announcer. This is example is what the capital stack looks like for a top-tier borrower with unusually strong counterparties. If those vendor backstops sour, the picture changes fast. But that asymmetry is the point. The well-backed can convert money into grid access. The speculative sponsor cannot, no matter how liquid he is. Run the stack forward. Capital is available and structurally indifferent to whether the borrower can build. Chips are financeable. Buildings are financeable. Every input has a lender behind it.

Except one. Nobody can vendor-finance a queue position. Broadcom cannot guarantee a transmission path. Google cannot backstop a water permit. There is no supplier whose credit makes an interconnection agreement appear.

The bet

I’m not betting that most of the pipeline disappears. I’m betting that most of it gets sold.

A queue position was never a plan to build. It was an option, and options have buyers. The developer who announced a gigawatt with no site control and no water math was never going to pour concrete. But he is going to be standing there holding a study, a deposit, and a date when someone who can build shows up needing exactly that.

The exit economics push the same way. Under the proposed PUCT large-load rules, a developer who withdraws from Batch Zero pays all incurred utility costs and gets back only 20% of the remaining per-megawatt financial security. The other 80% stays with ratepayers. If quitting is punitive and selling is not, people sell.

This new audit isn’t a filter. It’s a repricing. It takes the one asset that used to be free, a credible position in line, and makes it scarce and verified and therefore worth paying for. Every project the state culls raises the value of the ones that clear.

That inverts who’s in trouble. The loser isn’t the small speculative developer. He now has an exit he didn’t have before. The loser is the operator with real money, a real customer, and no grid position, who has to buy in at a price the audit just created. Fluidstack is a decent picture of that buyer: a customer, a backstop, lenders, and a need for sites.

The winner is whoever already holds the boring stuff. A brownfield substation. A dead crypto site. Water rights. Behind-the-meter generation, which may fall outside the directive entirely if it never touches ERCOT. An interconnection agreement signed in 2023 by someone who had no idea what he had.

Texas set out to audit the pipeline. What it’s actually doing is minting a title system for grid access. The first thing a title system does is create a market in the thing it certifies.

What would prove me wrong

Not the gigawatt count. I am watching transactions. Capital flows. Asset sales, JVs, and option-to-purchase filings on sites with existing interconnection. Price per MW of position against price per acre of raw land.

If the queue collapses and nobody buys the parts, I’m wrong and it really was just vapor. If the announced total stays roughly flat while the names on the projects change, I’m right.

I’ve been mapping the physical side of this, the transmission, parcels, permits, and visible construction, and the permit-versus-announcement count is top of mind. If you’re closer to the queue than I am, tell me: is the audit a real filter, or does most of this clear once developers catch up on their filings? Reply and tell me where I’ve got it wrong.

So I gave it a shoot. Gave a “frontier” model a real Texas parcel, the kind of thing you’d actually be looking at: rural county, dirt roads, real infrastructure in the area. I wanted to see what the workflow would do end to end, not in a demo, on something messy and real.

It came back fast and sure of itself. Exact coordinates. And then the map lit up around them: transmission lines, gas, fiber, substations, road access. A clean little diligence packet. Looked like a hell of a site.

Here’s the problem. None of it was wrong, and all of it was.

The layers were real. The join wasn’t.

Here is exactly how it breaks, because it’s not where you probably think it is.

The infrastructure layers were real. The transmission line is really there. The gas is really there. The math the system did, distance to substation, distance to fiber, proximity to existing activity, was real math, correctly placed on the map. If you audited the calculation, it would pass.

The failure happened one step upstream of all of that. The issue was that the location join was weak. All because The address was messy (in the way real addresses often are) and so the geocoder did what geocoders do: it didn’t say “I don’t know.” It returned the nearest thing that looked plausible. In this case it resolved to the wrong road, a few miles off from where the parcel actually is.

And once that wrong point is sitting on the map, everything downstream inherits the error and none of it looks wrong from your seat. The distances are computed correctly but all from the wrong place. The proximity claims are real, about the wrong location. The packet is internally consistent and externally false. It’s all based on bad coordinates wearing a confident diligence uniform.

That’s the part that bothers me. Not that the AI made an obvious mistake you could catch. That it made an invisible one, wrapped in real infrastructure and correct code, that reads as grounded right up until someone acts on it.

Why “I don’t know” is the missing feature

A geocoder is built to return an answer. That’s its whole job. Give Google, or Mapbox, or Esri, or the Census geocoder a clean, complete address and it does beautifully. Give it something incomplete or messy or slightly off, which is most of what real leads look like, and it usually won’t refuse. It returns the closest plausible match and moves on.

That match can be a lot of things. The wrong road. A ZIP centroid, which is the geographic middle of a postal code and often a field in the middle of nowhere. A county centroid, when it really has no idea. A mailing address instead of the actual site. A “corrected” version of the address that quietly changed what you typed in. Every one of those comes back with a latitude and longitude attached, and the coordinates make it look exact.

But exact and accurate are different things, and that gap is the whole problem. Precision is the AI giving you six decimal places. Accuracy is those decimal places pointing at the right dirt. You can have all the precision in the world and be pointing at the wrong county.

This isn’t a quirk of one bad address. After I ran this, I came across a piece by Will Cadell, who runs the geospatial firm Sparkgeo, writing about the same failure mode from the other direction: geocoders hand back a confident result with no reliability signal, and the errors are worst in exactly the rural, sparsely-addressed places where the messy inputs live. Which means the problem is well understood by the people who work in it. It just disappears the moment the output leaves their world and becomes a coordinate in someone’s workflow.

The reason this matters more in geography is that a wrong location doesn’t announce itself. A hallucinated quote you can often catch. A made-up statistic you can check. A confident pin on the wrong road, surrounded by real infrastructure, calculated correctly, is designed to survive review. It looks exactly like a right answer.

I keep coming back to how this works in navigation. A navigation system that’s right 90% of the time isn’t 90% good. It’s 100% bad, because you can’t trust any single instruction, which means you can’t trust the system. Ninety percent is worse than useless, because it’s good enough to make you stop checking. Spatial diligence is the same. A workflow that places the point correctly most of the time trains you to stop verifying, and then quietly puts the point in the wrong place on the one that matters.

The fix isn’t a smarter model

The instinct, when AI gets something wrong, is to reach for a better model. More parameters, a newer version, a bigger context window. That doesn’t fix this, because this isn’t a reasoning failure. The reasoning was fine. It’s a discipline failure, and it happens before the model ever reasons.

So I built the discipline in. Twenty-nine claim-gate skills for AI maps, small checks that sit between the raw data and any spatial claim the system is allowed to make.

The idea underneath all of them is simple, and it’s borrowed from how careful people already work. Separate what’s known from what’s inferred from what’s blocked. A coordinate checker that flags a weak or fallback location instead of trusting the first lat/long. A proximity gate that won’t let “near transmission” become a fact until the location it’s measuring from is actually defensible. Status checks that don’t let a planned project get mapped as an operating one. A claim ledger that, for every assertion, records whether it’s grounded, guessed, or unverified, and what evidence would move it from one to the next.

One rule ties the whole set together: no verified join, no spatial claim. If the location hasn’t earned its place on the map, the system is not allowed to say it’s near anything. Keep the lead in the ledger. Just don’t let it lie yet.

The point

None of this is an argument against using AI for spatial work. I use it constantly, and the version with the gates in front of it is genuinely powerful, because you can finally trust what comes out the other end. The gates don’t slow the useful part down. They stop the useless part from looking useful.

The point is narrower and, I also think, more important than “AI good” or “AI bad.” It’s that the dangerous failure in real estate AI isn’t the model being dumb. It’s the model being confident and wrong upstream of the answer, in a way that shows you real infrastructure the whole way down and reads as diligence.

A weak assumption became a map. The map became a memo. The memo became a decision. And nowhere in that chain did anything look wrong.

That’s the thing to fix. Not the model. The join.

here’s the link to the skills -

https://drive.google.com/file/d/15KPGr1hFZKAnFfaU-KzNJxrQmTCTqIHh/view?usp=drive_link

Last week I made a map about two Texases. It showed how the first one was built around people and the second is now being built around compute. I still think that’s basically right. But it was missing something, and the thing it was missing is that compute itself is splitting.

If you count sites, existing or planned, the map still looks urban. Dallas. Houston. Austin. San Antonio. Cities pull compute toward them, so the footprint clusters right where you’d expect: near people, fiber, customers, labor, logistics. That map is not wrong.

But if you stop counting buildings and start counting megawatts, and you see something very different.

Near the metros: 30,402 MW:

Remote counties: 81,383 MW:

Almost three times the load, in places with very few people around.

So now there are two real maps pointing in opposite directions. The count map says compute is urban. The power map says compute already moved, west and north, out past the cities into counties most Texans would have to look up.

This is worth paying attention to, not because it proves some grand theory, but because it breaks a shortcut people lean on all the time: more sites means more load. That used to be roughly true. It isn’t anymore. The place with the most buildings is no longer the place with the most power, and the gap isn’t small.

What the permits won’t tell you

Permits don’t say what any of these sites will actually run. Training, inference, cloud, colocation, enterprise. None of that is in the filing. That’s just not how permits work.

You could use common sense and say the small sites near cities are for inference and the giant remote loads are for training. Maybe. Probably, even, in a lot of cases. But not always, and a permit will never prove it. So I’m not going to pretend the data says something it doesn’t. What I can say is narrower and, I think, more useful: site count and power capacity are no longer telling the same story about Texas. They’re telling opposite ones.

Why it matters

If you’re buying land, building power, underwriting load, or just trying to read where this market is heading, which map you use changes the answer.

Count buildings and you see a metro play, where being near a city is the thing that matters. Land near the triangle, grid interconnection, competing with warehouses and logistics for the same parcels. Count megawatts and that story falls apart. The load is going where the power is, which is out where almost nobody lives, and the calculus there looks nothing like a metro build. Gas, water, land, transmission, and a relationship with the grid that a downtown developer never has to think about.

Same industry. Same state. Completely different map depending on what you decide to count

Where it’s going

There’s a second split hiding in the numbers, between what’s operating now and what’s planned or under construction. The early read is that the current footprint still leans metro, the sites already running are closer to the cities. But the pipeline, the stuff being planned and built right now, widens the power split rather than closing it.

If that holds, the remote power map isn’t just a snapshot of where things are. It’s the direction. The metro map is the present. The power map is the future, and the future is being poured in counties nobody is looking at.

That’s the next piece, once I’m sure the pipeline read holds up. For now the point is simpler than any theory about training or inference or what the hyperscalers are thinking. Texas hands you two honest maps of the same thing, and they don’t agree. Most people are looking at the one with the buildings. The one with the power is the one that moved.

Reply or comment “method” if you want access to the live Texas map to see where the megawatt figures come from. Part 3, on the pipeline, when the read holds.

The Texas work continues; this is the machine behind it.

New technologies show up in graduate theses years before they show up in your life. Long before the product, the pilot program, the utility interconnection queue, or the public fight at a zoning hearing, somebody sat in a library and wrote two hundred careful pages about an idea most people hadn’t heard of yet.

In 2010, before data centers became the public face of the AI buildout, MIT theses were already treating them as physical infrastructure. Not as an abstract computing topic, but as power loads, cooling systems, airflow models, operating cost, and energy efficiency. One thesis piloted low-cost efficiency improvements at a Raytheon data center in Garland, Texas, reporting a 23% reduction in electricity use. Another modeled above-plenum airflow with CFD to understand cooling losses above a raised-floor data center.

The market would later call this a real estate, energy, or permitting problem. The Archive shows it arriving first as a building systems problem.

That gap, between when an idea enters the archive and when it enters the world, turns out to be measurable.

Back in 2020, we built a tool at MIT’s Real Estate Innovation Lab to trace it. It took a lab, a team, and an industry partnership with JLL. This month I rebuilt it: $4 in compute, under an hour.

You can explore Attention Archive here, or just talk to the LLM Version.

This essay is how it works and what it found.

The $4 Rebuild

The $4 isn’t a gimmick. It’s a finding before the findings.

The hard part of the 2020 project wasn’t computers. It was reading time. What changed is that reading became a thing you can buy by the token. The whole archive, every thesis individually, for less than a typical Starbucks order. Any question that was ever too expensive to ask of an archive is now worth asking.

So here’s what asking looks like.

Attention Archive reads 26,562 MIT theses across Civil and Environmental Engineering, Architecture, EECS, Urban Studies and Planning, and Management. Not just keyword search. An AI reader goes through each thesis and records what it is actually about: the technologies discussed, the methods used, the geography, the themes, and the one thing that turned out to matter most: how the thesis talks about each technology.

That “how” gets scored on a five-stage scale, from the language of wondering to the language of taking for granted:

speculative → pilot → evaluative → critical → infrastructural

A speculative thesis asks whether something will matter. A pilot thesis builds with it and reports back. An evaluative thesis measures whether it worked. A critical thesis asks what it breaks. An infrastructural technology barely gets discussed at all. No thesis argues for concrete. Concrete is just there.

Score decades of writing this way and every technology traces a path. You can watch attention move: sustainability overtaking everything after 2000, computing going from footnote to method, AI going vertical, and data centers moving from equipment rooms into the built-environment conversation.

Keyword vs. Intention

Here is the uncomfortable part.

Most of what anyone believes about research trends, including what I believed in 2020, comes from dashboards that count keywords. Trend reports, horizon scans, “state of the field” charts: pattern matching, mostly, all the way down.

Pattern matching can be wrong in a very specific way, and now we can measure it.

Keyword mode works the way most dashboards work. Here is the actual rule for the “Data, GIS & Computing” theme, straight from the config, because methods should be checkable:

/\bgis\b|geographic information|\bdata\b|comput|simulation|\bmodel|algorithm|machine learning|sensor|digital|software/i

AI-read mode asks a different question: not does this thesis mention data, but is this thesis actually about data, GIS, or computing?

Same archive, same year, very different answers. In 2024, keyword matching tags 64% of built-environment theses as computing research. The AI reader says 15%. Skim the disagreements and the pattern is obvious: a pharma logistics thesis “uses data,” a housing policy thesis “builds a model,” a planning thesis cites simulation. Keyword counts all of it as computing research. On actual reading, only about one in five of those keyword hits turn out to be about computing at all.

That’s the difference between keyword and intention. A word appearing in a document tells you the author touched a subject. Only reading tells you whether they meant it.

And the gap itself is the story. After 2000, the keyword line climbs toward 70%, not because everyone switched to computing research, but because data, models, and simulation became how all research gets done. Computing is both a subject and a tool, and those are not the same thing. The distinction matters because it is exactly how technologies become infrastructure: first they are strange enough to be studied directly, then they become useful, then they become assumed.

The old instrument could not tell the difference between a field’s topic and a field’s toolkit. Reading can.

If your dashboards run on keyword matching (and most of the world’s do), that 64-vs-15 gap is worth sitting with.

The Data Center Question

This is why the data-center example matters.

If you ask when data centers first appear, the archive does not just return a trend line. It returns receipts. The early built-environment examples are about energy efficiency, facilities management, airflow, cooling, and operations. By 2025, data-center attention climbs to its highest level yet in the tracked series. The archive makes visible a shift that now defines the market: computation becomes real estate when it starts needing land, power, cooling, capital, and permission.

That crossing, from one discipline’s equipment room to another discipline’s land fight, took about fifteen years. And the archive caught a second crossing that is happening right now.

Data centers made that crossing between 2010 and 2025. Generative AI is making it now, and this time the instrument is watching in real time.

That is also why this connects back to Texas air permits. The permit archive is not a separate project. It is downstream of the same question: when does a technology stop being software and start becoming infrastructure?

Ahead of the Patents

One more early result, flagged as early.

I ran the same AI reader on a pilot set of built-environment patents, using the same schema, and lined up the attention curves. For the first overlap set of technologies, thesis attention peaks roughly sixteen years before patent mention volume. Median, four technologies. A pilot, not a law.

But the direction is interesting. Economists Comin and Hobijn estimated that technologies can take decades to fully diffuse into use. The MIT corpus sits at the front of that lag. What this means is that the academic archive is not just an old record of what already happened. It is a leading indicator of what is coming, sitting in public, machine-readable, and mostly unread.

Kick the Tires

Everything above is checkable, and that is deliberate. The keyword baseline is published. Technologies enter the lifecycle chart by rule rather than taste. Small samples are marked. Answers include thesis links.

And as of this week, you don’t have to take my word for it. Ask the archive directly.

Ask when data centers became an energy problem. Ask when GIS first appeared, and in whose thesis. Ask where keyword search and AI-read disagree. The system answers with numbers, sample sizes, charts, and links to the theses themselves. It also tells you when the archive can’t answer.

If you want the extraction schema and the keyword-vs-AI-read comparison notebook, reply to this email or comment “method” and I’ll send it.

What should this Reader read next? Patents at scale are the obvious candidate. Funding rounds and trade press would test whether the lead-lag holds downstream. A city’s planning archive is already in the works.

If you have access to a corpus and want to see it read, my DMs and inbox are open.

Attention Graph: archive.infrastructure-research.com

LLM version: ai.archive.infrastructure-research.com

It’s the map every Texan carries in their head. The triangle. The I-35 spine. A century of growth following one rule so reliably we stopped calling it a rule: demographics is destiny. People arrive, jobs follow, power lines follow the jobs, and the map fills in around the cities.

That map is now wrong. Or rather — it’s only half the map.

What the queue shows

Pull up the ERCOT large-load interconnection queue and sort by requested megawatts. The biggest new loads in the state are not landing in the triangle. They’re landing in counties most Texans couldn’t find on a map: Abilene and the Big Country. The Permian. The Panhandle.

These are not places with people. Some of these counties have lost population for decades. What they have is power — or more precisely, the raw ingredients of power: gas at the wellhead, wind and solar already built for a grid that couldn’t always take it, cheap land, and nobody nearby to object.

The new rule is simple: compute follows power, not people.

What the queue doesn’t show

Here’s the part that took me a while to understand, because it looks like a contradiction: the queue is evidence of the shift, but it also understates it.

The most aggressive builds don’t appear in the queue at all. They skip it. Onsite gas turbines, behind the meter, generating their own power on their own land — never asking the utility for permission, never waiting the four to five years an interconnection request now takes. New Mexico’s Project Jupiter ran this play one state over: a campus that went effectively off-grid to build far beyond the scale it first announced.

So the visible data shows the geography flipping — and the invisible data, the builds you can only find through gas nominations, air permits, and satellite imagery, says the flip is bigger than anyone’s official numbers admit.

The sixth Texas

There’s a beloved scene in Linklater’s Bernie where a Carthage local carves Texas into five states, right there at the diner counter: the flat ranches of West Texas, the Dallas snobs, the carcinogenic coast, the People’s Republic of Austin, and East Texas behind the Pine Curtain — where, he notes with pride, the South begins.

Every Texan carries some version of that map. Drawing it is practically a state pastime — I did it for my toddler this morning. And the monologue holds up fifteen years later, except for one thing: every one of his five states is defined by its people. Who lives there, how they talk, what they think of Austin.

Draw the map by megawatts instead, and none of his lines survive.

There’s a sixth Texas now, and nobody lives in it.

Layer the two distributions — where the people are, where the new load is going — and you get something strange: two Texases occupying the same borders.

The first Texas is the one we all know. Bernie’s five states, really one state: built for people, dense along I-35, anchored by the triangle, powered by a grid designed to serve homes and offices.

The sixth is being built for compute. It’s out west and up north, in the empty quarters, pulling city-scale power into counties with no city. A single campus in a near-empty county will soon draw the load of a mid-size metro. No people. Just machines. The Carthage local couldn’t have described its residents, because it doesn’t have any.

On the map, the two barely touch. That’s what makes the visual so striking — and so misleading.

They’re not separate. They’re competing.

Geographically, the two Texases pass like ships. Economically, they’re in a bidding war.

Every gas turbine that goes behind the meter at a data center is turbine capacity — and gas supply, and pipeline space — not serving the first Texas. Every “off-grid” campus still wants backup interconnection, transmission optionality, water rights, and skilled labor drawn from towns that don’t have surplus of any of it. And ERCOT has to plan a grid around enormous loads it can only partially see.

The competition clears through prices. West Texas compute demand tightens the gas market, the capacity market, the labor market — and the first Texas feels it, eventually, in its power bill. The people-map pays a tax to the compute-map, and almost nobody on the people-map knows the second map exists.

The map I’m sending

This is the map in the video: population density as the base layer, then the new load — queued and estimated behind-the-meter — layered on top. The connective tissue between them is the part I’m most interested in, and it’s where the next few essays are headed: which corridors bind the two Texases together, where they’ll collide first over gas and water, and which counties are quietly becoming the most important places in the state despite having almost no one in them.

The line between the two Texases isn’t one I drew. It falls out of the data. That, to me, is the story: the state’s own infrastructure network, left to speak for itself, describes two different places wearing the same shape.

My toddler’s map still has the cities on it - Bernie’s five states, waiting for her to learn their names. By the time she’s old enough to drive I-20 west and wonder why the horizon glows in a county with no town, she’ll need the other map. The sixth state won’t be on anything they teach her. But it will be paying for more of Texas than any of the five that are.

Demographics built the first one. Megawatts are building the second.

If you want the data and the notebook behind the map, reply to this email or comment “method” — I’ll send it over. And if this was forwarded to you, you can subscribe below for the next installment: where the two Texases collide first.

If you track permits, ownership, and the coordinates underneath the announcements, it reads as the opposite — a company that never had a single bet, building quietly while the headline number absorbed all the attention.

Here’s the whole thing, in the order it actually happened.

The timeline

October 2025. Poolside announces Project Horizon — a 2 GW AI campus in Pecos County, on Longfellow Ranch, anchored by a lease with CoreWeave. This is the version that made the maps and the trade press. Big number, big names.

Eight days later. A second Poolside entity, “Poolside LF Phase 1 DC Ops LLC,” comes online at Puckett Road, on the northern end of the same ranch, about 20 miles from the Horizon site. No announcement. No press. Just a registration and a permit.

March 2026. Poolside files a TCEQ permit for Phase 2 at the same Puckett Road location.

April 2026. The CoreWeave anchor lease on the 2 GW Horizon flagship collapses. The headlines write themselves: Poolside’s Horizon deal is dead.

Read in sequence, the collapse isn’t the story. The story is that Phase 1 came online eight days after Horizon was announced — before the ink was dry on the flagship — and Phase 2 was filed weeks before the collapse became public. The quiet track wasn’t a reaction to the flagship dying. It was running in parallel from the first week.

The collapse didn’t end Poolside’s presence in Pecos. It just removed the loud thing standing in front of the quiet one.

Same ranch, two parcels

This is where the coordinates matter — and where my map was wrong before I fixed it.

Both Puckett Road permits describe the same directions: from Fort Stockton, south on US-285 for 26.7 miles, then onto Puckett Road. That puts the active site about 27 miles south of town. The Horizon flagship sits further southeast. The real distance between them is about 33 kilometers — roughly 20 miles. Not two unrelated sites. Two parcels on the same property.

Longfellow Ranch is a single 350,000-acre holding owned by the Mitchell family, operated through Longfellow Ranch Partners. Poolside’s own announcement named “568 acres of land on the Longfellow Ranch owned by the Mitchell Family.” The “LF” in “Poolside LF Phase 1/2 DC Ops” is the ranch. The same landowner relationship covers both the dead flagship in one part of the property and the live buildout in another.

A note on method, because it matters as much as the finding: the Puckett Road site was geocoded wrong in my own atlas — pinned to Fort Stockton city center instead of the site 27 miles south, because the geocoder found “Fort Stockton” in the directions and stopped there. The permit text is what corrected it. If you underwrite off coordinates, the coordinates have to be right, and the only way to know is to read the filing, not trust the pin.

What’s documented, what’s inferred

I’d rather be precise than dramatic, so: the entity names, the dates, the shared address, the landowner, and the parallel timeline are all documented. What the permits don’t specify is whether “DC Ops” means AI compute or something else — a standard TCEQ air registration covers the power equipment, not the load. The AI read comes from who Poolside is, an AI model-training company — not from the permit text. “Data center buildout by an AI company” is documented; “AI compute” is a reasonable inference. The difference is the whole point of working from filings.

Why this is the pattern

Pecos is clean, but it isn’t unique. The shape repeats: a large, named, headline-grade project absorbs the coverage while the real, permitted capacity goes up quietly nearby — often under a different entity, sometimes on the same land, usually without a press release.

There’s a reason quiet works. Across the country, data center projects are being killed by local opposition over power and water. The announcement draws the fight. The quiet build avoids it. Whether that’s deliberate or just how phased construction looks from the outside, the effect is the same: the loud project is gone, and the quiet one is still going.

That difference in the message is the whole game. The announcement is a number a company chose to publish. The filing is what they’re actually building. The headline gives you the gigawatts. The permit gives you the coordinates.

This is part of an ongoing series mapping where the AI buildout actually lands in Texas — permits, power, ownership, and the coordinates underneath. If something here is wrong or incomplete, tell me. The data gets better when practitioners push back. If you have the TCEQ application text for the Phase 2 permit — the portal was erroring when I pulled this — send it; it’s the one piece I couldn’t source directly.

So I priced one. Meta broke ground in February on Domino, a $10 billion, 1-gigawatt campus at Indiana’s LEAP District in Lebanon. 1,500 acres, 13 buildings, two main phases. It’s under construction now. The state and city have posted parallel fact sheets on scale and timeline. That’s the story on the wire. What the wire doesn’t resolve is whether the equity in a deal like this is mostly real estate and execution, or mostly pricing power tied to AI demand that can outrun new supply.

Why this campus

A gigawatt is a lot of power — roughly the draw of 750,000 homes. It’s also not the frontier anymore. Meta’s own Hyperion campus in Louisiana is heading toward 5 GW. Amazon, Microsoft, and Google are all building large sites in Indiana too. Domino is about 8% of Meta’s 2026 capex.

So why price the small one. Because it’s the most readable one. The cleanest. Big enough to matter, small enough to see clearly, recent enough to model at today’s assumptions. And because Meta has no cloud-rental business, Domino is a clean read on a single question. A merchant operator’s campus reflects its lease deals. Domino reflects only Meta’s own conviction about compute demand. There’s no third-party tenant in the numbers to muddy what they’re actually wagering on.

What I ran

I built a spatial model I run myself for this, very much inspired by the great Daniel Fink and his open source Rangekeeper framework.

Click a building footprint on the map, the polygon converts to square footage, square footage converts to megawatts, and a 15-year phased cashflow model runs underneath: construction S-curve, two phases (249 MW around 2028, 751 MW around 2030), a utilization ramp, MISO power costs, Lebanon’s tiered water rates, the LEAP abatement, 60% leverage. About 855 lines of custom Python.

Again, the distribution approach is borrowed from the Rangekeeper, the hyperscale economics are bolted on top.

One disclosure before any numbers, because it defines the experiment.

Meta doesn’t pay rent to itself. The model prices capacity at dollars per kilowatt per month, a market proxy, the same way you’d value owner-occupied real estate at market rent per square foot. So I’m not computing Meta’s P&L. I’m asking a cleaner question: if this capacity were priced at the market, what is it worth. That framing also sharpens the captive point.

A colo operator whose pricing collapses can re-tenant the building. Meta can’t. If its own compute demand softens, there is no backup tenant, which makes Meta’s downside structurally worse than a merchant’s, not better.

How pricing works in the model

Capacity rent isn’t a flat input. Demand compounds at a default 40% a year. Supply follows with a two-year lag. Price moves with the demand-to-supply ratio, clamped between a $120 floor and a $230 ceiling, starting from a $160 base.

Under the defaults, demand stays ahead of supply and price climbs into the ceiling, so even the base case reflects a tight market. To be clear about that $230: it’s a model cap, not an observed quote. And the 40% growth rate mostly determines how fast you reach the ceiling, not the level. That detail matters more than it looks, and the simulation is where it shows up.

How I stress-tested it

Three hand-picked scenarios will tell you a story, but you pick them, so they tell you the story you already believe. I’d rather not choose. So I ran the whole thing 8,000 times.

I pulled the actual Lebanon filings first. The campus is developed through a shell entity, Orla LLC. The water agreement has dated delivery tiers (Tier II in January 2028, Tier III in January 2031) that cap how much capacity can come online and when. I fed those dates, plus substation energization and Phase 2 completion, into the model as distributions rather than fixed inputs: on-time, or slipping 6, 12, or 18 months, weighted 40/25/20/15. Pricing ran as a triangular distribution across the $120 to $230 band, mode at $230. That’s a joint Monte Carlo, the same distribution primitives Rangekeeper uses, applied to a phased campus with the Orla water cap wired in.

Here is what came back.

Base case: 16.3% IRR, about $4.84B in value created, on current slider inputs with no slips. Then I let the constraints randomize. Across the 8,000 draws, the median outcome is 12.5% IRR and $1.64B. The tenth percentile is 7.0% and a $1.64B loss. The ninetieth is 15.1% and a $3.72B gain.

One thing a careful reader will catch: the base case sits above the P90, which looks wrong until you see why.

The base case holds pricing at the $160 input. The simulation lets pricing range up to the $230 ceiling and lets delivery slip, so the band isn’t a confidence interval around the base case. It’s the distribution of outcomes once you stop assuming the constraints land on time and at one price.

A positive base case with a negative tenth percentile is the correct read, not a glitch: it says the deal works on today’s inputs and still has a very real downside once you price the uncertainty.

Then the part that decided the post. I decomposed the variance to see which input drives the spread. Capacity pricing explains about 98% of it (98.2%, to use the run’s number). Power timing, Phase 2 timing, and the water tiers together explain the remaining 2%.

The largest single deterministic move in the model is pricing: hold everything else at base and compress the rent ceiling from $230 to $160, and you lose about $5.8B of value.

A 12-month slip on Phase 2, by comparison, costs about $480M. A 12-month water-tier slip costs about $250M. One caveat on method: the variance shares are an OLS attribution on standardized inputs, not a full Sobol decomposition, so read them as approximate. The ranking is the point, and the ranking isn’t close.

The water fight is the cleanest test of that. Meta has requested up to 3 million gallons a day, phased through 2031, and residents near Eagle Creek have pushed back.

It’s a real permitting and political risk. But the campus uses direct-to-chip, liquid-to-liquid cooling, a closed loop that consumes far less water than the evaporative towers most people picture, so the consumption is below what the 3-million-gallon headline implies. More to the point, the model says water timing moves when value lands, not whether it’s there. It’s friction on the schedule, not a threat to the thesis.

That’s the asymmetry I keep thinking about. Demand growth, water, power timing, construction slippage: the model absorbs all of them. The one input it can’t absorb is the rent compressing, and the rent compresses for one reason, which is supply catching demand.

Where the value actually lives

The question I asked next is where in time the value sits.

At year 15 the model adds a continuation value, stabilized NOI capitalized at 6.5%, worth roughly $25B. Strip it out and the operating years alone don’t carry the deal; NPV goes negative without it.

For a merchant operator that terminal value is a sale. Meta never sells, so here it stands for something else: the present value of this capacity still being scarce and useful after 2041. Which means the bet shows up twice.

The rent curve asks whether demand stays ahead of supply over the next decade. The continuation value asks whether compute is still scarce in year 16 and beyond.

If the demand curve bends, both answers turn at once. Nobody pays a 6.5% cap on an asset whose pricing just collapsed. One bet, priced in two places, and they fail together.

So the honest one-line summary of a $10B hyperscale campus in 2026: it’s a bet that compute capacity is still scarce and valuable in 2041.

The structural bet

Meta’s implicit bet isn’t that Indiana is cheap. It’s that AI compute demand stays steep enough that the demand curve never bends in a way that forces capacity rent down across the portfolio. And the scarcity it’s betting on isn’t only about demand. It’s about demand staying ahead of supply, and every hyperscaler is building supply as fast as it can.

Demand can keep growing and the bet can still fail - if the industry’s own buildout brings enough capacity online to close that gap.

Domino is one reading of a bet Meta is placing across the whole fleet: Hyperion at 5 GW, the rest of the buildout, $125–145 billion of capex in 2026 alone, up from $72 billion the year before. A pricing hit doesn’t add up linearly across one-off projects. It correlates across the portfolio, because market rent per kilowatt is the moving part in all of them, and there’s no third-party tenant anywhere in the fleet to catch a fall. The correlation isn’t a side risk. It’s the position.

What to do with this

Separate “can they build it” from “will rent hold.” Different diligence stream, and the first one is mostly settled at this point. They’re in the ground.

Stress the pricing path before you stress four separate cost levers. The variance decomposition and the scenarios say the same thing: one input dwarfs the rest.

If you hold multiple hyperscale exposures, model the portfolio as one rent curve, not N independent IRRs. The correlation is the risk.

And ask where the value sits in time. If most of it lives past year 15, and here it does, you’re not underwriting a building. You’re underwriting a view on whether compute is still scarce in 2041.

The model is open

Full breakdown, the interactive map, the scenario logic, every assumption, what the app actually loads. The numbers get better when people check the work. If you think base rent starts somewhere other than $160, or the $230 ceiling is wrong, or 6.5% is the wrong cap for this asset in 2041, tell me where. Those three inputs are most of the argument.

Sources: Meta — Lebanon campus announcement; City of Lebanon — announcement; Project Domino — Plan Commission staff report (Docket 25-35); Project Domino — Water & Wastewater Agreement; Rangekeeper framework (conceptual reference); Meta 2026 capex guidance (Q4 earnings). Model and methodology are my own; numbers are from the current 8,000-draw run and will move if inputs change.

This is really a data problem. Memory fades fast, and the version of a place the photo shows you isn’t the version you actually liked. You may just have a lot of photos, which would be a selection bias. So I did the thing I do with data and fab sites. Put it all on a map. Then stream the layers on a timeline.

Here’s the finding first, then how it got built — including the parts that didn’t.

The gap between “popular” and “worth it” is wider than I expected. Google and IG’s algorithms optimize for popular — volume, foot traffic, the spots everyone already knows. What I wanted was worth it.

Put the local signal next to the algorithm’s signal and two different cities appear in the same place. Packed with tourists, thin on local ratings: trap. Quietly rated by locals, nearly invisible in search: the one you walk twenty minutes for. The map’s whole job is making that gap visible.

Now the build.

There are five layers, and each one answers a different question.

The route — where we actually were. Every photo carries metadata, and 4,543 of the 4,717 had a usable timestamp and GPS coordinate — about 96%. I pulled those straight from the file (lat, long, creation date). GPS-derived, not guessed. The other 174 were missing data and got dropped. Order everything by when it was taken and the route draws itself: four weeks of movement, in sequence, as lines and markers.

What we ate — going from “I was here” to “what was here.

A GPS point says we stood on a corner in Osaka at 9pm. It doesn’t name the yakitori place. So I clustered the photos into stops — anything within 90 meters and 90 minutes collapses into one place we lingered. Then search Google Places within 200 meters of each cluster, keeping matches inside 160.

The first pass resolved 196 visited or likely-visited food and drink stops:

74 in Korea, 122 in Japan.

In retrospect this is suboptimal of course. And limited by Google Places API quota at 200 per day, which I had to batch. So the map is partial because I ran out of API calls. and time.

What locals think — the layer that’s the point.

Local-platform ratings, Tabelog in Japan and Naver in Korea, matched against the places we’d been: 50 local records, 35 that matched cleanly to a stop. This is the layer that turns a travel diary into the popular-vs-worth-it comparison — and it’s where the gap between the two shows up.

The roads not taken. Around the dense areas — Tokyo, Kyoto, Osaka — I ran a separate Japan-only search for places near where we’d spent time but never actually went, excluding everywhere we’d been. That added 184 candidates: 76 in Tokyo, 68 in Kyoto, 40 in Osaka. The midnight spots we’d clocked and missed, geocoded for next time.

The conversation layer — and this is the one I didn’t expect. Before and during the trip, a lot of the planning happened in long conversations with Claude — where to eat, what to skip, what was worth the detour.

I pulled the places out of those conversations by hand — 16 of them, 13 in Japan, 3 in Korea — each with the excerpt in JSON format and the date, and dropped them onto the map as their own layer. Matched to the photos from the same day.

So the planning shows up right where the decision was live.

The place we talked about standing next to the place we went. It’s small and hand-built, not an automated “platform”. But seeing the conversation plotted against the route is the part of this map I keep going back to.

I did not build a magic “best ramen” algorithm. That is a future post.

There’s no “single score” ranking every place on the map. I built a layered system — my path, what Google could see near it, what locals actually rate, the roads not taken, and the planning conversation underneath.

The value is in comparing the layers, not pretending one number replaces judgment. The map makes the gap visible. The call (and memory) is still mine.

It’s the same method underneath the data-center work. Geocode messy sources, fuse the layers that usually live in different sources. Surface the signal sitting under the obvious metric and with limited bias — with AI in the loop doing boring extraction work.

Ramen shops or HBM fabs, planning conversations or interconnection queues.

Same move.

This is just the version where the dataset is delicious.

Build your own. If you’ve got a trip’s worth of geotagged photos, you can run this: pull GPS + timestamp from the metadata, cluster into stops, match against Google Places, layer a local-rating source the algorithm under-weights (Tabelog, Naver, your region’s equivalent), and — if you planned the trip with an AI — pull those places into a layer of their own. The recipe matters more than my specific spots.

Going to Japan and want a similar map? let me know.

— Yair

Specifically HBM (high-bandwidth memory), which is now about 63% of what it costs to build a frontier AI accelerator — up from 52% just two years ago (Epoch AI). It’s the most expensive layer in the chip, and the supply is brutally tight.

If you want a sense of how tight: at Computex last week, Jensen Huang made a point of walking up to the SK Hynix booth, picking up a marker, and signing a wafer with a note — “Please make more.”

SK Hynix’s entire 2026 production is sold out — customers have offered to prefund new lines just to get ahead in the queue. A few days later, NVIDIA signed a multiyear deal to lock in SK Hynix supply for its next-generation systems.

Here’s why that matters. Everything else in the stack, you can route around — build your own power, bring your own cooling. HBM is the one input you can’t. Three suppliers, and no fourth door.

So what even is HBM?

Picture regular memory as Houston. Everything’s spread out — you drive twenty minutes to get groceries, twenty minutes back. Now picture HBM as New York. Same groceries, but the store’s on the ground floor of your building. You’re back upstairs before a Houstonian finds his keys.

That’s the whole trick. HBM is DRAM stacked vertically and bonded right beside the processor, so the data barely needs to travel. The short distance is the point — it’s what feeds a GPU fast enough to keep up.

Only three companies in the world make it: SK Hynix, Samsung, Micron. And the four largest AI chip designers ate roughly 90% of the supply last year. It’s the most concentrated, most expensive layer in the entire stack.

Which makes it a useful lens for the whole supply chain. Because an AI chip isn’t one thing built in one place. It has different layers, and each one gets built where its own constraints are cheapest to solve. Those constraints are different enough that the layers land in completely different states.

Three layers. Three binding inputs. Three different maps.

Fabs follow power and water. A leading-edge fab needs a credible path to 200–600 MW before anything else, drinks around 10 million gallons of water a day, and needs flat, vibration-free land. That’s most of the siting decision right there. It’s why Micron’s New York megafab sits in the Niagara hydropower zone — power at a 75–80% discount to market.

Packaging follows talent. Stacking and bonding chips isn’t a megawatt problem, it’s a workforce problem — how fast can you staff and train. So it clusters near research pipelines. SK Hynix’s $3.87B advanced packaging plant went to West Lafayette, Indiana — greenfield, next to Purdue. Chosen for people, not power.

Data centers follow power too — but stranded power. They used to chase fiber. In 2026 they’re power-first, landing in places like West Texas and North Dakota where energy is abundant but isolated. Musk’s Colossus in Memphis is the extreme case: sited for speed-to-power, gas turbines trucked in, everything else secondary.

Three layers, three binding inputs, three maps that don’t overlap.

With one developing exception.

Texas is where the layers start to meet.

And the numbers are concrete. Samsung’s Taylor fab is a $17B project — the largest foreign direct investment in Texas history — backed by about $4.75B in finalized CHIPS funding. It makes leading-edge logic: 2nm-class chips, trial production by the end of 2026, mass production in 2027.

It was sited for textbook fab reasons: 25km from Samsung’s existing Austin campus so the two share infrastructure, plus state incentives and an established Central Texas ecosystem.

What makes Taylor different from Micron’s New York fab or SK Hynix’s Indiana plant is the customer. Taylor’s anchor client is Tesla — an in-state buyer — in a chip deal worth at least $16.5B running to 2033, with Musk saying he’ll personally walk the process. So the fab layer and the demand layer are both Texas entities, on the same grid, under the same state policy. The chip gets fabricated in Taylor; a major buyer builds compute a few hours away; xAI is expanding in the same region.

Fab plus demand plus the data-center buildout, all under one grid and one incentive regime. That convergence doesn’t exist anywhere else in the US at this scale.

And here’s where HBM becomes the whole story.

It’s the one layer Texas doesn’t have.

The memory that is two-thirds of an AI chip’s cost is being onshored almost entirely into Micron — three of the five US sites, around $35B, in New York, Idaho, and Virginia — while SK Hynix and Samsung keep their HBM lines in Korea. Taylor makes logic, not memory.

So even Texas — the one place the chain is genuinely converging — still imports its most expensive layer. The loop doesn’t close. The state can fab the logic, buy the compute in-state, power some of it with ERCOT and local gas, and still the HBM ships in from Icheon.

That’s the thing worth watching.

I track data centers obsessively, because they have such a big physical footprint. I started tracking fabs because of CHIPS. But the layers underneath sit on different maps, on different clocks — most of these memory sites don’t even produce until 2027–2029 — governed by different rules.

Texas is the first place those maps start to overlap. And the fact that the most valuable layer is the one still missing — at least until someone decides to fix it — is exactly why it’s the place to watch what gets built next.

Full video:

All of it needs the same thing eventually: a legal path to discharge wastewater. Through a publicly owned treatment works (POTW), through a permitted industrial outfall, through a regional water authority. Not “any river that looks close on a map.” Not the storm drain. A National Pollutant Discharge Elimination System (NPDES) permit, reviewed against the receiving plant’s actual capacity to handle the load.

That capacity is finite. And it's not showing up as a primary screen in most site selection — at least not yet for the AI buildout.

Why now

Three things are happening:

One. Intel’s Ohio One campus is in active construction. The first fab’s process wastewater profile will hit Ohio EPA’s NPDES process in the next 24 months. Industrial process water at fab scale is not municipal sewage — it negotiates limits, pretreatment requirements, and monitoring against the receiving plant’s headroom. The Licking Regional Water District just approved a new $200M treatment facility at Raccoon Creek specifically to handle this expansion. That’s one fab, one regional response, after a multi-year permitting fight.

Two. Ohio EPA proposed its first-ever general NPDES permit for data centers in late 2025 (OHD000001). The public comment period closed January 16, 2026. The final permit is pending. Whether it’s approved as drafted determines whether every new data center in Ohio gets fast-track coverage under one framework or has to file individually — which means longer timelines, site-specific review, and case-by-case headroom negotiations.

Three. The queue is forming. Every project that files first locks in allocation against finite plant capacity. Every project that files late pays for upgrades, hauling, redesign, or relocates. This isn’t a future problem. The order is being set right now, by which permit applications hit Ohio EPA’s queue first.

That’s the urgency. Not in 2030 when the fabs come online. Now, when the permits are being written.

What I measured

I started with Ohio’s NPDES universe. About 1,850 geocoded facilities, drawn from Ohio EPA’s individual permits database, crosswalked to facility coordinates. Underneath those facilities sit 6,400+ outfall-level rows — the actual discharge points, each with design flow (what the permit allows) and recent flow (what’s actually being used).

For each outfall where both numbers are reported, headroom is the difference:

headroom = design flow − recent actual flow, floored at zero.

I rolled those outfalls to county polygons and compared county-level headroom to a geocoded data center site list in the same counties. One caveat to surface up front: only about one in five outfalls has both design and recent flow data in the extract.

The map shows every geocoded facility. The headroom calculation is conservative — only about one in five outfalls report both design and recent flow, so the others get counted as zero

Counties with low reported headroom may have more slack than the data shows; counties with high reported headroom may have constraints that don’t surface until a specific project applies. The methodology is conservative — it counts only what’s observable.

That conservatism is the point. Even under the strictest “observed only” rule, the concentration is stark.

The concentration

The top 10 Ohio counties hold roughly 70% of statewide observed discharge headroom. Eighty-eight counties total. Ten of them control most of the usable slack.

That’s not how power distributes. Ohio’s grid is constrained but the constraint shows up almost everywhere — interconnection queues are long across most of the state. Wastewater headroom is the opposite: highly concentrated, geographically uneven, and not correlated with where projects are clustering.

Spearman correlation between county headroom rank and data center count rank: 0.40. Moderate, not strong. Clustering is not following headroom. It’s following power and fiber.

That’s the gap.

What this map doesn’t show

Headroom is hydraulic — raw gallons of permitted discharge capacity. It’s not a measure of whether a specific data center can lawfully send its cooling tower blowdown there. That’s a separate gate: chloride, total dissolved solids, conductivity, residual chlorine, treatment chemistry. Whether a POTW can accept industrial process water depends on the receiving plant’s existing permit limits and the chemistry of what arrives.

I could only pull chemistry data on about 3% of Ohio outfalls. So the map shows where the gallons are. It doesn’t show where the data centers can go. That’s a separate ruler, and almost no public dataset holds it well

Who loses when the constraint binds

Three groups, in order of how soon it shows up in their cost basis.

Late movers in tight corridors. Projects that announce after major permits and pretreatment agreements harden don’t just pay more for land. They pay for plant upgrades, hauling, or redesign. The first three or four big projects in a county set the operating envelope. Everyone after that operates inside it or pays to expand it.

Communities that assumed river access equals disposal access. NPDES is the legal gate. High-salinity industrial waste, semiconductor process water, and battery manufacturing byproducts often can’t route through municipal sewage without explicit permission and plant capacity that doesn’t yet exist.

A general permit that treats all data centers the same — regardless of cooling technology, water chemistry, or location — folds what should be three separate technical conversations into one. Whether that's the right tradeoff is a real policy question, and it's a question Ohio EPA hasn't yet answered.

Anyone pricing land on interconnection alone. Megawatts queued and gallons withdrawn are the wrong first screen if discharge is the long-run constraint. The site that has power and water on paper but no clear discharge path is a site that hasn’t actually cleared diligence.

What the map shows

County-level discharge headroom, choropleth from aggregated outfalls. NPDES facility locations and outfall points layered underneath. Data center sites in the same frame, so you can see where slack exists vs where projects are clustering.

The comparison rules match the tables. No fantasy layers, no projected capacity, no forecast headroom. Only observed headroom from reported design and recent flow.

The map shows Franklin lit up with both headroom and clustering. Belmont and Lorain showing significant headroom and nothing else. Licking in the middle, Intel’s anchor visible. Most of southern and eastern Ohio with headroom that no project is queueing against.

That’s the geography of the next five years if nothing changes.

What’s next

Switchyard currently covers Texas. The Ohio dataset behind this analysis — facility-level NPDES, county-level headroom, the data center geography — is the foundation for the Ohio expansion. If you work in Ohio and want to search a specific address, the wait is real but the data is being readied.

The bottom line

Land in Ohio is still being priced like power is the bottleneck. The real picture is more complicated. Hydraulic capacity is finite and concentrated. The chemistry that determines whether a specific site can actually use it is even more uneven. Both rulers point in the same direction: the site that has power and water on paper but no clear discharge path hasn’t actually cleared diligence.”

Intel’s 2030 timeline is when the fab’s wastewater story stops being slides and starts being permits and meters. The data centers stacking up behind it will be deciding their discharge paths in the same window.”

The constraint isn’t the fab. It’s the pipe out of it.

Sources and methodology

Primary sources

• Ohio EPA — NPDES permitting program

• Ohio EPA — Draft General NPDES Permit OHD000001 (data centers)

• U.S. EPA — National Pollutant Discharge Elimination System (NPDES)

• U.S. EPA — National Pretreatment Program (POTW / industrial)

• Licking Regional Water District — Raccoon Creek permit (Columbus Dispatch, May 2026)

• Intel Ohio One — production timeline (ENR)

• Alliance for the Great Lakes — OHD000001 public comment

Methodology

Headroom is calculated as design flow minus recent actual flow, floored at zero, only for outfalls where both flows are reported in the Ohio EPA extract. About one in five outfalls meet this criterion in the current dataset. County rollups sum outfall-level headroom within county boundaries. Data center sites are geocoded from a separate site list and spatially joined to the same county polygons.

The 70% concentration figure compares the top 10 counties’ summed headroom to the statewide total under the “observed only” rule. Concentration remains high under conservative bounds that zero out missing recent flow values.

The queue for large loads is at 230 GW.

That gap is the shadow grid forming. Most of those queue requests won’t clear — Andrea Himmel projects up to 93% wash out under SB-6’s August attestation gate. But the campuses with capital aren’t waiting. They’re going behind the meter — building their own gas turbines, on-site fuel, battery storage. Filed under names that don’t connect to their operators in any public registry.

The Stargate example

The most public version of this campus is also the cleanest example of how the naming works.

Stargate Abilene — the $500B AI build-out announced by Trump from the Oval Office in January 2025, with SoftBank, OpenAI, Oracle, and Crusoe at the table — sits on 825 acres at 5502 Spinks Road in Taylor County. 1.2 GW of interconnect approved. 10 buildings being built, expanding to 20. $3.4B in financing from Blue Owl Capital. It’s the most discussed AI infrastructure project in the world.

It is filed at TCEQ as Abilene DC 1, LLC.

The campus name on the state’s power station tracker is Longhorn Data Center.

The operating permit lists 62 emergency backup generators, 170 MW of total backup capacity, 28 of them powered by Baudouin engines that have a sparse distribution network in the U.S. consisting “largely of marine engine shops” (per Hunterbrook Media’s reporting). When generator supplier IGSA Power was asked about its orders for the site, the representative referenced “Abilene DC3, DC4, DC8” — implying a serialized naming convention covering at least eight separate filing entities.

Stargate does not appear anywhere in the permit.

Neither does OpenAI. Neither does SoftBank. Neither does Oracle.

The largest AI infrastructure project in the country is filed under a name that doesn’t appear in a single press release.

This is not unique

Stargate is the loudest version. It is not the most extreme.

A partial inventory of campuses filed under serialized or shell entities across Texas, from public TCEQ records:

Galaxy Helios I LLC. Dickens County. 1.63 GW approved, 120 emergency diesel generators, CoreWeave as the public tenant. $1.4B in financing. The permit was filed January 2025. The announcement still hasn’t come.

Crusoe Energy Systems LLC. Multiple Texas campuses, including the Lancium build-to-suit at Abilene. Public-facing operator with shell entities filing specific permits.

Lancium Clean Campus filings. Multiple Lancium-affiliated entities filing for separate BTM capacity at the same campus, with different sub-entities holding different portions of the interconnect approval.

ElectriGen. 1.8 GW of behind-the-meter generation, filed in the last several weeks. Tenant entity not disclosed in public filings.

The pattern repeats. The public name announces. The shell entity files. The connection between them lives in trade press reporting and forensic database work, not in the public record.

What this means for the grid you can see

A 300 MW data center with on-site gas turbines and battery storage is physically a power plant. It has the equipment, the fuel supply, the generation capacity, and the grid-interactive controls to inject power back into the system, shed load when ERCOT calls a conservation alert, or operate as an island during a grid event.

In February 2021, when the Texas grid collapsed during Winter Storm Uri, 246 people died. The grid had a 30 GW shortfall at peak. The BTM campuses being built and announced now — the ones surfacing in TCEQ filings, the ones being financed without ERCOT queue positions — could individually be sized to cover that shortfall many times over. They won’t be.

It will not be used to prevent the next Uri.

Not because it can’t be. Because nothing in the financing, the permit structure, the tenant contracts, the SLA architecture, or the ERCOT relationship requires it. Uptime is the product the campuses are selling. Sharing the asset back means risk to that product. The entities financing these campuses didn’t model bidirectional flow because nobody asked them to.

Why this isn’t being tracked as one thing

Three structural reasons.

One. The data sources don’t talk to each other. ERCOT’s interconnection queue tracks load that wants to connect to the grid. TCEQ tracks air permits for combustion equipment, regardless of whether the equipment is grid-connected. The shadow grid lives in TCEQ data because it’s burning gas. It doesn’t live in ERCOT data because it isn’t connecting. Nobody operating ERCOT’s load forecast is reading TCEQ’s permit filings as a leading indicator.

Two. The naming conventions break standard linkage. Public registries link entities to operators through name matching. When a campus is filed under “Abilene DC 1, LLC” and operated under “Stargate Abilene,” standard database joins return nothing. The link exists in the world. It doesn’t exist in any database that aggregates either set of filings.

Three. The reporting cadence is wrong. TCEQ updates its pending application list every Friday in a pipe-delimited text file. Most journalists, analysts, and even some industry trackers don’t ingest weekly regulatory filings. By the time a campus appears in trade press, it’s been in the regulatory record for months.

The shadow grid is hidden in plain sight. The data is public. The tools to read it as one system don’t exist outside of a small number of forensic operations.

In December, the Georgia PSC approved nearly 10,000 MW of new generation. $16 billion in capex. Roughly four-fifths of the load attributed to data centers. Five new gas plants. Eight months from request to approval.

The water hasn’t cleared at all.

One site tells the math.

Project Sail in Coweta County surfaced in DRI-era filings at roughly 9 million gallons per day. The county uses 7.7 MGD total. One campus, proposed at more than a hundred percent of countywide consumption.

The number was later revised down — the developer now markets a closed-loop air-cooled system at peak 300,000 GPD. Maybe that holds at full buildout. Maybe it doesn’t. Plaintiffs in the rezoning lawsuit allege the filings misstated water needs. Either way, the original number cleared an approval process. That’s the part worth sitting with.

A site that can be filed at 9 MGD against a 7.7 MGD county, and still move forward to a vote, is a site selection process that isn’t pricing water as a binding constraint. It’s pricing it as a checkbox.

The geology doesn’t help.

Metro Atlanta sits on granite basement. Groundwater is thin. The region runs on surface storage in a small watershed footprint. Lake Lanier’s pool stage is a drought-trigger input, not a vibe. Georgia spent thirty years in litigation with Florida and Alabama over this basin and won the legal framework. Winning the framework is not the same as allocating the next gallon during a Stage 2 restriction.

The state is currently under Level 1 drought. Most of it is severe-to-exceptional. The QTS Fayetteville campus, twenty miles south of Atlanta, drew 29 million gallons through two unmetered connections over fifteen months — discovered when residents lost water pressure and the county was simultaneously asking them to stop watering their lawns. The water director’s explanation for why there was no fine: they’re our largest customer, we have to be partners.

That’s not a scandal. That’s a political economy. And political economy is what determines whose water gets cut at 2 a.m. on the worst day of the year.

The timing problem

Thermal load peaks when weather is worst. That’s the same window when outdoor watering bans bite, when irrigation gets cut first, when evaporative cooling economics look worst, and when the county engineer is making decisions she didn’t expect to make this fast.

Hospitals are on the priority list. Nursing homes are on the priority list. Data centers haven’t been written into one — not because anyone decided against it, but because the framework was written before a single county could host a load this size. The question of what turns off first hasn’t been asked out loud. It will be.

Two approvals, not one.

Power approval is a design problem. Interconnection queue, transmission headroom, substation capacity, generator buildout, PSC tariff structure. Modeled to four decimal places. Cleared in eight months.

Water approval is a political problem. Permitted withdrawals, return-flow assumptions, basin compact, drought stage triggers, who gets cut and in what order. None of it modeled in the site selection layer. Most of it answered after the campus is already designed.

A site that has cleared the power approval has not cleared the water approval. They are not the same approval. The senior people in the room know this and the site selection workflow doesn’t reflect it yet.

The next 10 GW of Georgia data center capacity will be sited on power maps. The constraint that breaks the model isn’t on those maps.

That’s the diligence gap. It shows up late, it shows up at the basin level, and it doesn’t move the campus — it reprices it.

It also contains, if you know what to filter for, the data center build-out before the announcements. We didn't read it manually. We pointed AI at it. The same system that finds the permit can cross-reference the power infrastructure, predict the site footprint, and render what the campus probably looks like — before a shovel moves, before a press release, before anyone is looking.

This is what it found.

The Signal

Data centers that generate their own power — on-site gas turbines, backup diesel fleets, behind-the-meter generation — have to file an air quality standard permit with the Texas Commission on Environmental Quality before construction begins. This is not optional and it is not private. It is a public record in a database almost nobody reads.

On January 30, 2025, TCEQ made effective a new permit category specifically for natural gas electric generating units. The new standard permit expanded the pathway for on-site gas generation — which is increasingly how large data centers are solving the grid queue problem. Instead of waiting 18-36 months for a transmission interconnection, they bring their own power. And when they do, they file here first. [1][2]

We pulled six months of standard permit registrations — October 2024 through April 2026 — and filtered for power generation outside the Permian Basin. 340 rows. 85 distinct legal names. Then filtered again: non-Permian county, LLC with minimal web presence, geographic overlap with known data center corridors.

Three hits.

Galaxy Helios I LLC — Dickens County

Permit 179090. Filed February 6, 2025. 120 emergency diesel generators. 984 County Road 112, Afton, Texas.

Galaxy Helios I LLC registered with the Texas Comptroller on January 22, 2025 — taxpayer number 32098487682, SOS file 0805873149, mailing address 1999 Bryan St Suite 900, Dallas. [3][4] The contact on the TCEQ application is Austin Storms, Co-Head of Mining and Mission Critical at Galaxy Helios I LLC. [5]

Dickens County is where Galaxy Digital is financing the Helios data center campus — $1.4B in financing, CoreWeave as the tenant, 526 MW contracted at the time of initial disclosure. By January 2026, after ERCOT approved an additional 830 MW tranche, total approved campus power crossed 1.63 GW. [6]

The Texas Comptroller’s data center project list shows HEL02 DC filed November 7, 2025 under Galaxy Helios II LLC with CoreWeave Inc. as operator — confirming multi-building phasing on the same campus. [7]

The air permit came first. 120 diesel generators at a rural Dickens County address, filed under a brand new LLC, before any of the financing milestones and before any press coverage of the site’s scale. The diesel count alone — 120 units — is an order-of-magnitude signal for GW-scale critical load.

The permit was filed January 2025. The Phase I delivery date is H1 2026. Eleven months of silence in between.

Poolside Lf Phase 2 Dc Ops LLC — Pecos County

Permit 183235. Filed March 9, 2026. Pecos County. Lat 30.609441 N, -102.573484 W.

Poolside AI announced a Pecos County data center in October 2025. The fundraise fell apart. That was the last anyone heard.

Except someone opened a new LLC on March 4, 2026 — taxpayer 32104474260, SOS 0806474166, mailing address 548 Market St PMB 53385, San Francisco, CA. [3] That address is a Regus mailbox. The general counsel listed on the TCEQ application is Carson Klingenberg at carson@poolside.ai. [8]

The entity is named Poolside Lf Phase 2 Dc Ops LLC.

Six weeks after formation, they filed a power generation permit in Pecos County.

The project isn’t dead. It’s regrouping under a new vehicle.

Abilene DC 1 LLC — Taylor County

Permit 177263. Formed June 12, 2024. Taylor County.

Taylor County is Abilene. Abilene is where Crusoe built the first operational Stargate site — 600 MW live, 1.2 GW contracted, per Epoch AI’s April 2026 analysis. [10]

Abilene DC 1 LLC — taxpayer 32095486604, SOS 0805587707, Austin mailing address — filed a Taylor County air permit for five 38 MW Titan 350 turbines and five 34.1 MW GE LM2500 units. Approximately 360 MW of behind-the-meter gas generation. [9][3]

The LLC formed June 2024. The Stargate announcement came later.

The permit came before the announcement. It always does.

What This Means

The permit is just the entry point. What comes after is the part that’s new.

AI found this site. AI modeled what it needs to run. AI rendered what it will look like. The human role in that workflow was pointing it at the right question.

That’s not a research methodology. That’s a new way to develop infrastructure.

Next Friday, TCEQ updates the file. The next three are already in there. The question is whether you see them before or after the announcement.

If you want to run this methodology on your market — reply here or request map access at infrastructure-research.com. The map is live. The signal set updates weekly.

The Helios campus footprint — GPU halls, cooling towers, switchyard, BTM gas generation, oriented to the 138kV line — was generated by the system from the permit address and local infrastructure geometry. The capital stack was modeled independently using Rangekeeper. Both outputs preceded any satellite confirmation or public announcement. [Map and model available on request.]

Sources

[1] TCEQ Standard Air Permits index. https://www.tceq.texas.gov/permitting/air/newsourcereview/standard

[2] TCEQ Air Quality Standard Permit for Natural Gas Electric Generating Units. https://www.tceq.texas.gov/permitting/air/newsourcereview/combustion/ng-egu-sp

[3] Texas Comptroller Active Franchise Tax Permit Holders. https://data.texas.gov/Government-and-Taxes/Active-Franchise-Tax-Permit-Holders/9cir-efmm

[4] Comptroller franchise account — Galaxy Helios I LLC. https://comptroller.texas.gov/taxes/franchise/account-status/search/32098487682

[5] TCEQ Records — Galaxy Helios I standard permit correspondence (dID 8711479). https://records.tceq.texas.gov/cs/idcplg?IdcService=TCEQ_EXTERNAL_SEARCH_GET_FILE&dID=8711479&Rendition=Web

[6] Galaxy Digital, Form 8-K August 15 2025 — $1,400,000,000 senior secured term loan facility for Galaxy Helios I LLC. https://www.sec.gov/Archives/edgar/data/1859392/000185939225000039/glxy-20250815.htm — and press release: https://investor.galaxy.com/news-releases/news-release-details/galaxy-closes-14-billion-project-financing-facility-accelerate

[7] Texas Comptroller Data Centers in Texas — HEL02 DC / Galaxy Helios II LLC. https://comptroller.texas.gov/taxes/data-centers/data-center-lists.php

[8] TCEQ Records — Poolside Initial Application (dID 9580276). https://records.tceq.texas.gov/cs/idcplg?IdcService=GET_FILE&dID=9580276&dDocName=8294578

[9] TCEQ Records — Abilene DC 1 technical review (dID 8600163). https://records.tceq.texas.gov/cs/idcplg?IdcService=TCEQ_EXTERNAL_SEARCH_GET_FILE&dID=8600163&Rendition=Web

[10] Epoch AI, "OpenAI Stargate: Where the US Sites Stand." https://epoch.ai/blog/openai-stargate-where-the-us-sites-stand — note Epoch published methodology revisions on capacity estimates after publication; verify current figures at https://epoch.ai/data/data-centers-documentation