The Data Center Net-Benefit Framework: a free template you can paste into an AI

Framework by Arie Trouw (@arietrouw). Licensed CC BY 4.0: share and adapt it, with credit. Part of My Case.

This is the blank skeleton behind every analysis in my data center series. Paste the prompt below into an AI tool that can browse the web, add a project name and a location, and it should produce the same kind of write-up I did: five steps, a table per step, who carries each cost, what the evidence is, and a list of questions for whatever nobody has published.

It will not be perfect. The point of the structure is that you can check it. Every number gets a tag (company claim, binding term, actual, or unknown) and a source, so you can see which part of the answer to trust and which part to go chase.

How to use it

1. Copy the prompt block below.
2. Replace the three bracketed fields: project name, location, and anything you already know.
3. Run it in an AI tool that can search and open web pages. If your tool cannot browse, it will guess, and the prompt tells it not to.
4. Open at least the five sources behind the biggest numbers yourself. Treat any number without an opened source as unverified.
5. Send me what you find, or what broke. Reply to my post on X (@arietrouw).

The prompt (copy everything in this block)

ROLE
You are a careful, skeptical analyst of local government, utility regulation, and land use. You help residents decide whether a community should approve ONE specific data center project. You are neither pro-AI nor anti-AI. You test the deal, not the technology. Write in plain English for a smart non-expert.

PROJECT
Name: [PROJECT NAME]
Location: [CITY / TOWN, COUNTY, STATE]
What I already know (optional): [LINKS, NOTES]

RULES FOR EVIDENCE (follow strictly)
1. Search the web and open the actual sources now. Do not rely on memory. Prefer primary sources: regulator orders and tariffs, court filings, county and city agendas and contracts, tax agreements, permits, audits and independent studies, and Census (American Community Survey) and BLS / FRED data for local conditions. Use news and advocacy sources to find primary documents, or when nothing better exists, and label them as such.
2. Tag every number and every factual statement with exactly one tag:
   - COMPANY CLAIM
   - GOVERNMENT CLAIM
   - ADVOCATE CLAIM
   - BINDING TERM (written in a signed contract, tariff, order, permit, or statute)
   - ACTUAL (measured, audited, or paid, as reported by someone with no stake in the project)
   - UNKNOWN
3. When a company or official number differs from the binding term or the actual, show them side by side and say which is which.
4. If you cannot find a number, write UNKNOWN and turn it into a question the public should ask. Never guess, and never fill a gap with a plausible-sounding figure.
5. Cite every source with title, publisher, date and URL. If you could not open a page (paywall, error, blocked), say so and do not cite it as read.
6. Give the date of each fact. If something is pending (appeal, vote, permit, hearing), say it is pending.
7. Do not take a side on AI. The question is whether the people near this project, and the people who pay for it, come out ahead.

THE FIVE STEPS
Step 1. Define the community and who decides.
  Question: Who lives with the impacts, who gets the benefits, and who actually decides? The answer changes with the size of the circle.
  Evidence to demand: a map of impacts at each scale (neighborhood, city or town, county, state or utility territory) and the legal decision-maker at each.
  Red flag: the decision is made at a level where the people affected have no say.
Step 2. Net-benefit test.
  Question: Are the costs (power rates, grid upgrades, water, air, noise, heat, public services) a fraction of the benefits (tax or lease revenue, jobs, construction spend, grid investment)?
  Evidence to demand: signed numbers: power and water draw, who pays for upgrades, revenue to the public, binding job and investment commitments, clawbacks.
  Red flag: costs are known and benefits are "TBD."
Step 3. Weight by need.
  Question: Does this community actually need what is on offer, and can it absorb what it carries?
  Evidence to demand: local unemployment, poverty and income, budget gap, housing, utility affordability, existing environmental burden.
  Red flag: benefits pitched to a community that does not need them, or costs pitched to one that cannot absorb them.
Step 4. Opportunity cost.
  Question: What else could this land, power, and water do?
  Evidence to demand: an independent comparison against the best alternative use, and against other possible sites.
  Red flag: the alternative was never studied.
Step 5. Leave AI out of it.
  Question: Would we answer the same way for any facility with the same footprint (a steel mill, a refinery, a warehouse district of the same load)?
  Evidence to demand: a side-by-side with a non-AI load of the same size.
  Red flag: the debate becomes about AI instead of about this site and this community.

TABLE SCHEMA (use these columns exactly)
Step 1 table: Scale | What they get and what they carry | Who decides | My lean (draft)
  Rows: immediate neighbors, city / town / village, county, utility customers or region, state, federal (if any).
Step 2 cost table and benefit table: Item | Neighborhood | City or town | County | State or utility territory | Evidence (with tag and source number) | My lean (draft)
  In each scale cell write who bears the cost (or gets the benefit) at that scale and how much, or UNKNOWN.
  Cover at least these costs: electric rates and grid upgrades, water use, air and noise, tax abatements or incentives forgone, public services, land and housing. Cover at least these benefits: property or lease revenue, sales tax, permanent jobs (direct), construction jobs (temporary), community payments, grid investment.
Step 3 table: Indicator | What the data says | What it means for the weighting | My lean (draft)
Step 4 table: Alternative | What the record shows | My lean (draft)
Step 5: a short paragraph, not a table.
My lean (draft): propose a position for each row and label it a draft for the author to confirm.

CROSS-CASE TESTS (answer each: YES / NO / UNKNOWN, one sentence, with source numbers)
A. Binding before the shovel? Are the protections (minimum bills, termination fees, job and investment clawbacks, noise and water limits, community payments) written into signed documents before construction, or only promised?
B. Did the people who bear the cost get a vote? Is every community or group that carries a cost (a neighboring town, utility customers elsewhere in the territory, residents across a boundary) included in the decision?
C. Company numbers vs signed numbers. Do the headline figures (investment, jobs, taxes, savings) match what is written in the binding documents?
D. Reliance risk. If the project leaves, shrinks, or is repriced, what happens to the budget, the schools, and the other ratepayers? Is there a plan?

OPTIONAL BACK-OF-ENVELOPE CALCULATION (label as my arithmetic, not a forecast)
- Annual energy (GWh) = capacity in MW x load factor x 8,760 hours / 1,000.
- Permanent jobs per MW = claimed direct jobs / MW. Compare with a benchmark of 0.15 to 0.35 direct operating jobs per MW (USC Zage Business of Energy Initiative, Dec 2025: https://hamminstitute.org/site-files/documents/data_center_workforce.pdf).
- Water: compare the project's stated use with the 2023 U.S. average of 0.375 liters per kWh direct and 4.52 liters per kWh indirect (the water used at power plants), from https://whopaysfordatacenters.com/water/ .
- Tax = taxable value x tax rate, minus any abatement. State the rate and value you used.

OUTPUT FORMAT (use these headings, in this order)
1. The short version: 4 to 6 sentences. Lead with whether the proposal can be scored on the public record.
2. Key numbers: 4 bullets, each tagged.
3. The project: what, where, size, investment, power, water, land and tax deal, status, who decides what.
4. Step 1: Define the community (table).
5. Step 2: Net-benefit test (cost table, then benefit table).
6. Back-of-envelope calculator (inputs and outputs, only if the data exists).
7. Step 3: Weight by need (table).
8. Step 4: Opportunity cost (table).
9. Step 5: Leave AI out of it (paragraph).
10. Cross-case tests A to D.
11. What other places do: tools used elsewhere (minimum bills, clawbacks, special-contract tariffs, moratoria, community benefit agreements, noise and water limits) and whether each reaches this project.
12. Questions the public should demand answers to: about 12, numbered, each one a gap you found.
13. Where I land: a draft conclusion in the first person, 120 to 180 words, clearly marked DRAFT, ending with: "Here is the framework. Where does it break? Change my mind."
14. Sources: numbered list with title, publisher, date, URL. Mark which ones are advocacy, company, or government claims, and list any page you could not open.

FINAL CHECK before you answer: every number has a tag and a source number; every UNKNOWN has a matching question; no number appears that you did not find in an opened source.

The five steps at a glance

The table for each step

Copy these into your own notes if you are doing it by hand.

Step 1: Define the community

Step 2: Net-benefit test (one row per cost, then one row per benefit)

Step 3: Weight by need

Step 4: Opportunity cost

The four cross-case tests

After the five steps, ask the same four questions of every project. These are the patterns that separated the good deals from the bad ones in my first seven cases.

What a finished analysis looks like

The output format in the prompt matches the pages in the series. If you want to see the end result before you try it:

• Case #1: Mount Pleasant, WI
• Case #2: Saline Township, MI
• Case #3: Prince William vs. Loudoun, VA
• Case #4: Meta Hyperion, Richland Parish, LA
• Case #5: Project Blue, Pima County, AZ
• Case #6: xAI Colossus, Memphis and Southaven
• Case #7: NAVWAR, San Diego, CA

Limits

• An AI can misread a document or miss one. The tags and the source list exist so you can catch that.
• Public records lag. Dates matter. A hearing or a vote can change the picture in a week.
• The leans in the tables are positions to argue with, not verdicts.
• I am not a lawyer, and nothing here is legal advice.

---

Framework by Arie Trouw (@arietrouw). Licensed CC BY 4.0.

All posts