Framework by Arie Trouw (@arietrouw). Licensed CC BY 4.0.
A 400 MW data center has been floated for the Navy’s Old Town campus in San Diego, about 70 acres between Pacific Highway and I-5. I live in San Diego County. When the developer was asked what the community would get, the answer in its emails was “TBD.” No tenant is signed, and nobody has said who pays for the grid.
I build things for a living. I run XY Labs and I started XYO. So I do not have a problem with big infrastructure. I have a problem with big infrastructure that cannot fill in its own scorecard.
Most of the fight over data centers is the wrong fight. People argue about AI. I think the better move is to take the decision apart, the way you would any large decision with a lot of moving pieces, and ask five smaller questions.
The five questions
1. Define “community,” and who decides. The answer changes with the size of the circle. The neighbors hear the fans. The city gets the tax base. The county gets the rest of the bill. The state writes the rules. The utility’s other customers may pay for the power plants and may not have been asked. Before anything else, write down who is in each circle and who actually has the vote.
2. Run the net-benefit test. The costs should be a fraction of the benefits, and they should be written down. Power rates, grid upgrades, water, air, noise, tax breaks given up. Against that, tax revenue, jobs, construction spending. Signed numbers count. Press releases do not.
3. Weight the benefits by need. A job in a place with high unemployment is worth more than a job in a place with none. A tax windfall is worth more to a parish with a 24% poverty rate than to a county that is already rich.
4. Count the opportunity cost. Land, power, and water can only be used once. What else could they do?
5. Whether you are pro-AI or anti-AI is irrelevant. Swap in a steel mill with the same load and the same footprint. Would you answer differently? If so, you are arguing about AI and not about the deal.
I ran it on seven real projects
Each one has a full write-up with tables, sources, and the questions nobody has answered yet. The leans below are my positions on the evidence so far, and I am open to changing them.
- Mount Pleasant, WI (Microsoft). The best-documented deal. A pre-built tax district, a guaranteed minimum value, and a power tariff that makes Microsoft pay 100% of the generation built for it. The open items are the ones the village does not control: transmission costs, the school district’s share, and a neighboring town that had no vote. Yes, if I lived there.
- Saline Township, MI (Oracle and OpenAI). A tight power contract, but the township said no, got sued, and then settled in a consent judgment. The tax abatement is still being fought. Conditional yes on the power deal, no on the abatement as written.
- Prince William vs. Loudoun, VA. Loudoun is the fiscal best case, with about $1.3B a year in data center taxes and a county fact sheet that now warns against dependence. Prince William lost its mega-project on a notice failure before anyone tested the deal. Yes where the land is already industrial and there is a plan for the plateau.
- Meta Hyperion, Richland Parish, LA. A poor parish got a windfall. The bill for a $20 billion power buildout may land on 1.1 million Entergy customers who were not at the table. Yes for the parish, no on the process.
- Project Blue, Pima County, AZ. The city said no, the county sold the land anyway under an NDA, and the project got better after the fight. There is still no confirmed tenant. Conditional.
- xAI Colossus, Memphis and Southaven. SELC and the NAACP lawsuit say turbines ran before permits. The benefits are real but modest, and the emissions land in a metro area that the NAACP says fails the smog standard. No on the process.
- NAVWAR, San Diego, CA. My local one. On the public record it cannot yet be scored. For my street, no. For the county, only if the numbers are written down.
What separates a good deal from a bad one
After seven cases, the pattern is not about the size of the project. It is about four tests.
Was it binding before the shovel? Mount Pleasant had a tax district built in advance, a guaranteed minimum value, and a regulator-written tariff. Saline’s power contract is strong (19 years, an 80% minimum bill, a termination payment), but the community terms were set in a consent judgment after the township said no and was sued, and that judgment bound the township to the tax abatement. In Memphis and Southaven, SELC and the NAACP lawsuit say turbines ran before a permit covered them. Terms won up front are worth more than terms won after.
Did the people who bear the cost get a vote? In Hyperion, the cost risk sits with Entergy customers across the whole territory. In Mount Pleasant, the neighbor town that hears the fans had no vote. In Pima County, Tucson’s council voted unanimously to end its involvement, and the county sold the land 3-2 anyway. The gap between who decides and who pays is where bad deals live.
Do the company numbers match the signed numbers? The Mississippi state release says more than $20 billion and hundreds of jobs in Southaven. The tax agreement, as quoted by an advocacy site, says $10 billion and 20 jobs. In Saline, the project value went from $4.8B at the settlement to more than $43B by July 2026, and that matters for the abatement. In Louisiana, Entergy’s customer-savings claim went from $2.65B to $4B at a hearing, and its own witness said the $4B is not guaranteed.
What is the reliance risk? Loudoun gets about 46% of its local tax funding from data centers and has cut homeowner tax rates for a decade. Its own budget fact sheet warns against building recurring spending on it. A windfall that becomes the budget is a dependency.
There is a fifth pattern that runs through all of them: secrecy. An NDA in Pima County, confidential contracts and a subpoena quashed 3-1 in Louisiana, and “TBD” in San Diego. A proposal that will not fill in its own scorecard has not passed the test.
Good deals write the protections down before the vote. Bad deals ask you to trust the press release.
Use the template on your own town
I put the framework in a blank template with a prompt you can paste into an AI tool along with a project name. It tells the AI to open the actual documents, separate company claims from binding terms and actuals, mark unknowns as questions, and cite everything.
Get the template. It is free to share and adapt under CC BY 4.0. Credit is all I ask.
It will not be perfect. The point is that you can check it. If it breaks on your town, tell me where.
What comes next
I am publishing one case a day, a weekday at a time:
| Date | Case |
|---|---|
| Mon Nov 2 | This post, the template, and Case #1: Mount Pleasant, WI |
| Tue Nov 3 | Case #2: Saline Township, MI |
| Wed Nov 4 | Case #3: Prince William vs. Loudoun, VA |
| Thu Nov 5 | Case #4: Meta Hyperion, LA |
| Fri Nov 6 | Case #5: Project Blue, AZ |
| Mon Nov 9 | Case #6: xAI Colossus, TN and MS |
| Tue Nov 10 | Case #7: NAVWAR, San Diego, CA |
Send me a project
Is there a data center proposal near you that you want run through this? Reply to my post on X (@arietrouw) with the project name and location. I will pick some and publish them here, and the best replies will show up in 🎤 Open Mic.
What I am not saying
I am not saying data centers are good or bad. I am not saying yes or no for your town. The leans above are drafts, my positions on the evidence so far, and the questions at the bottom of each write-up are the things nobody has published yet. Many of the numbers are company, county, or advocate claims, and I label them that way. If I got something wrong, I want to know.
Here is the framework. Where does it break? Change my mind.
How this was made
I used an AI research assistant to gather and cross-check sources, then checked each write-up line by line against the sources it opened and removed anything I could not support. Every number links to a numbered source list. Where I could not verify something, I wrote it down as an unknown.
Framework by Arie Trouw (@arietrouw). Licensed CC BY 4.0.