What a data center community benefit agreement should contain, and who makes sure it is kept

Version 1.1, September 5, 2026. Adds what a community does and does not give up by signing, narrow-release drafting, capital and refresh-cycle contributions, participation funding, writing terms into the permit, and a decommissioning bond, drawn from the CURE webinar with the Better Data Center Project. about 4,000 words. A checklist is at the end.

The Short Version

In 2026 the community benefit agreement became the industry's answer to local opposition. Brookings called them necessary in January. Pennsylvania's governor made good-faith negotiation of one a condition of state support in May. Frederick County signed a $110 million one in September. The NAACP published a template. The Federation of American Scientists went looking for executed data center agreements and found eight, only one of them posted in public.

Almost all of the attention is on the negotiation: what to ask for, what to concede, what the document should say. Almost none is on what happens the day after it is signed. On a Brookings panel in February, three audience questions asked what enforcing one of these agreements looks like. The supervisor from the most mature data center county in the country said his county had none. The panel's authors said it was "another blog we need to write."

This piece is about the day after. It sets out what a data center agreement should contain, drawing on the executed agreements and templates now public, and then it argues that the contents matter less than four structural choices: where the money sits, who decides, who reports, and who checks. Most agreements get the contents partly right and the structure entirely wrong. That is why residents in Loudoun County, with 200 facilities next door, have never seen one enforced.

Baseload Commons designs and runs the after-signing layer of these agreements, and the founder convenes an independent standard that audits them. Both are disclosed at the end. Read the argument with that in mind.

1. What Is Actually In The Agreements

Four sources let us describe the state of the art with some confidence rather than from principle.

Lancaster, Pennsylvania (November 2025). The only executed data center agreement posted in full. The City of Lancaster and Chirisa Technology Parks, developer of two campuses leased to CoreWeave. Twenty years. The owners contribute $10 million to each of two funds: an Economic Development Fund held at the Lancaster County Community Foundation and administered with the Lancaster Economic Development Company, and a Sustainable Development and Clean Energy Fund held by the City. Payments arrive in $2.5 million increments at construction financing and within 30 days of operations, phased per building, and are secured by a letter of credit. Water is capped at 20,000 gallons a day per campus, down from the 100,000 to 125,000 the previous printing plant used. Noise may not exceed pre-construction ambient levels at nearby homes, measured before and after. All buildings run on 100 percent clean electricity, with a $10 million contingent fund if they fall short. Generators test only in daytime, never on Sundays. The owners must keep a public website with permits and a complaint portal, present a report to City Council every April on clean energy compliance, plan for e-waste and decommissioning, and bind any buyer of the property to every obligation. The City can seek an injunction or, as a last resort, discontinuance of operations. No tax abatement was granted.

Brookings (January 2026). Nicol Turner Lee and Darrell West list twelve elements a data center agreement should include: direct payments to a locally administered fund, with one proposal at 0.5 percent of gross revenue; realistic and public tax revenue projections; infrastructure commitments with exit fees if a project closes early; enumerated jobs with abatements tied to them; electricity cost sharing; water consumption and closed-loop cooling; noise and light studies; environmental monitoring; workforce training; health services; digital access and AI literacy; and a public dashboard tracking jobs, taxes, electricity, water, and workforce results. They reject nondisclosure agreements outright and note that abatements of 10 to 50 years can outlast the facility itself.

The NAACP template (2026), hosted by Columbia's Sabin Center. A community-side model agreement. Its center of gravity is a Community Advisory Board, majority residents and workers, on which the developer has no vote and no appointment power, and which has "sole authority to administer, allocate, and disburse" a Community Benefit Fund held in a dedicated auditable account. Annual contributions are indexed to inflation and rise if energy or water use grows. The developer funds a public Community Reporting Dashboard whose metrics the board sets, an independent annual audit, a staff coordinator housed at a community organization, and stipends for board members. Payments are additional to, not a substitute for, legally required mitigation. The board can issue a Notice of Concern that triggers audit, mediation, and arbitration.

The WRI Community Benefits Frameworks Database. Seventy-two agreements across two decades and many sectors. Four in five contain some reporting language. Fewer than half have an enforcement clause. Fewer than half name a community committee. Fewer than one in four combine all three. Only ten address renewal.

Read together, these sources agree on the contents to a surprising degree. Water caps, noise limits measured against ambient, clean energy with a real definition, jobs with numbers rather than goals, a fund, a dashboard, a decommissioning plan, no NDAs. Where they differ, and where most executed agreements fail, is on structure.

1.5 What A Community Gives, And What It Does Not

Before the contents, the trade. A community benefit agreement has the structure of a settlement: the developer commits to benefits, and the signing community groups commit not to challenge or oppose the project as approved, sometimes agreeing to support it. Sarah Friedman of the Better Data Center Project, a lawyer and former Department of Energy official, made the point on a CURE webinar in 2026 that this is the same shape as a pre-settlement agreement, and that a community should go in knowing it.

It is worth being precise about what that does and does not mean, because the shorthand version, that a community gives up its right to sue, is wider than the reality and can frighten a township away from a tool that may serve it well.

What a release typically covers. Challenges to the approvals for this project as permitted: the land use approval, the zoning decision, the environmental review. That is the thing the developer is buying, and it is a reasonable thing to sell in exchange for enforceable commitments.

What it does not cover, and should never be drafted to cover. Claims for breach of the agreement itself, which is the whole point of making it enforceable. Permit violations, environmental violations, and workforce violations that happen later. Statutory rights that generally cannot be waived. Harms that arise from operation rather than from approval. And anything at all involving people and organizations who did not sign; a release binds its parties, not a town.

What it does not cover unless it is badly drafted. Expansions and additions. Data center campuses grow, and a Microsoft campus in the Des Moines area has kept adding buildings since it opened. Draft the release to the project as described and permitted, and treat an expansion as a new project requiring a new negotiation. This is the clause most worth a lawyer's time.

Read that way, the practical effect is narrow and defensible. It becomes harder to relitigate what you agreed to, for as long as the agreement is being administered as agreed. If it is not being administered as agreed, the community's remedies are the ones the agreement itself creates, which is why the enforcement and reporting sections below matter more than the dollar figure.

Two other consequences shape the drafting.

It is not a tool to stop a project, and the benefits usually flow only once the facility is operating. A community whose goal is to prevent a project should not treat an agreement as the route to it, and a community that wants a strong agreement should understand that its leverage comes from the same organizing, engagement, and legal exposure a fight would require.

One group cannot sign for everyone. Because the impacts run across water, air, energy, schools, and the tax base, a single organization settling on its own can undercut everyone else's position, even though it cannot legally bind them. The counterparty should be a coalition, and every affected interest should be consulted before negotiation opens.

For communities that would rather not negotiate a release at all, a community benefit ordinance is the alternative: a local law applying to any development over a defined threshold, requiring benefits as a condition of approval, with nothing given up by anyone. Our companion piece on the instruments covers that choice.

2. What Should Be In It

The following is the union of what the sources require and what the executed agreements have taught. It is organized in the order a township will ask about it.

2.1 Disclosure before negotiation

Every serious source now agrees that an agreement negotiated on undisclosed numbers is not an agreement. Before the first term is discussed, the developer publishes:

Acreage, square footage, and peak megawatts at full build-out, not phase one
The load ramp: contracted capacity by year until full build-out, because utilities in constrained markets now quote four to seven years to full power and service agreements are otherwise proprietary
Water demand and source at full build-out, with the cooling design
Generator count, permitted run hours, and emissions controls
Realistic tax revenue projections with the abatement modeled in, and the assumptions published
Projected water and peak power demand at least 90 days before the first approval hearing

Lancaster's water cap only means something because the previous user's consumption was known. Frederick County's agreement reduced potable water use by 80 percent, which is only measurable because a baseline existed.

2.2 Mitigation with numbers

The Columbia critique of the Lancaster agreement is that too much of it relies on "reasonably maximize" and "where feasible." Its noise standard is relative to ambient with no decibel ceiling. Its energy commitment has no megawatt figure per building. The lesson is not that Lancaster is a bad agreement. It is the most complete one publicly available. The lesson is that even the most complete one leaves room, and not everyone rates it as highly: Friedman's assessment is that Lancaster is not a bad example but that its finances do not look strong for a project of that size, which she attributes to the city's limited leverage rather than to poor drafting. Both readings can hold. A community should read the agreement itself and decide whether $20 million against that capital investment is proportional in its own case.

Each mitigation term needs a number, a measurement method, a cadence, and a place the result is posted:

Noise: a decibel limit at the property line, measured quarterly, posted. Pre- and post-construction baselines as Lancaster requires.
Air: Tier 4 controls specified, caps on generator hours per year, per day, and by time of day, published run-hour and emissions data, and an explicit prohibition on running backup generation for demand response. Scale makes this urgent: Virginia has permitted enough data center backup generation to power roughly 20 million homes in a state with about 4 million, and generators must run 10 to 30 hours a year for testing alone, which at that scale approaches the emissions of a major power plant.
Water: a daily cap, metered, posted monthly. Closed-loop cooling specified by design, not by intent.
Energy: a definition of clean that excludes unbundled certificates, an hourly or at least annual matching requirement, and a consequence for falling short. Lancaster's $10 million contingent fund is the right shape. The St. Louis agreement's "endeavor to purchase" credits is the wrong one.
Light, traffic, buffers, screening: as Lancaster: dark-sky lighting, 100-foot buffers on streets and residential edges, screened plant.

2.3 Jobs with counts, not goals

Construction employment is large and temporary. Operating employment is small and permanent. Every source says the same thing: name the numbers. Cedar Rapids ties a 70 percent, 20-year abatement to a minimum of 31 full-time jobs at $26.20 an hour. The NAACP template asks for 40 to 50 percent local hiring at prevailing wage. Lancaster has a "Local Hiring Plan" with percentage goals, which Columbia rightly flags as the weakest part of the agreement. The fix is a number, a reporting cadence, and funded apprenticeships so the number is reachable.

2.4 The contribution, tied to the facility

The Brookings proposal is 0.5 percent of gross revenue. The NAACP template is an upfront payment plus indexed annual payments for the life of the project including decommissioning. Lancaster is $20 million in fixed increments tied to financing and operations milestones. Older agreements in the WRI database show the cleanest formulas: the Eight Point Wind host community agreement pays $3,000 per megawatt per year for the life of the project; South Fork Wind pays 25 annual installments with a 2 percent escalator.

For a data center the defensible structure is a per-megawatt annual contribution on energized capacity, escalating with inflation, paid quarterly, for the operating life and through decommissioning. It scales with impact, it survives a sale, and it does not depend on anyone's revenue accounting. Check the mechanism against the total. Whatever formula you choose, do the arithmetic that tells you whether the number is proportionate: add up everything the agreement commits over its full term, including refresh triggers, and express it as a percentage of the facility's total capital investment. A per-megawatt formula is the cleanest mechanism, because it scales as the campus energizes and it can be audited against public interconnection data. But a mechanism is not a size. Two percent of capital investment over the term is a defensible target and one percent is a floor; below that, a community should ask why. For reference, the most complete publicly posted data center agreement in the country commits somewhere around one to two percent, and an experienced practitioner has publicly called its finances not strong for a project of that scale.

Two further refinements from practitioners who negotiate these. First, tie at least part of the contribution to capital investment rather than revenue. Nobody outside an operator has reliable insight into a facility's annual revenue, and a formula the community cannot audit is one it cannot enforce; capital investment appears in filings and financings. Second, and less obvious, capture the refresh cycle. Every three to seven years the chips, servers, and electrical equipment inside the building are replaced, and at hyperscale that reinvestment can run into the billions. An agreement written only against the original build silently exempts every dollar spent after it. Name reinvestment as a triggering event.

What to avoid: a single up-front gift, a step-down after year five, and any clause that lets the developer direct where its own contribution goes. The Kahana Solar agreement in Hawaii reserved part of each year's payment for the developer's discretion, and it is the clause residents cite when explaining why they distrust these documents.

Two further lines that almost no agreement carries and every one should:

Capacity funding, in two places. A Maine host community agreement funded a $100,000 account used solely for the town's own planning and decision making. A township facing its first gigawatt campus has no data center expertise and no budget to buy it, and funding that at arm's length is the cheapest credibility a developer can purchase. The second place is the one almost every agreement misses: money for residents to participate, before the agreement is signed and for as long as it runs. A council on a twenty-five year facility asks unpaid people to keep showing up for twenty-five years. Stipends, a paid coordinator housed at a community organization, and a standing budget line for the council's own advisors are what make resident governance real rather than nominal. The NAACP template requires all three; most executed agreements require none.
Digital access. Brookings puts it on the list. The fiber a data center pulls along its route can become public middle-mile service for the towns it passes. A broadband line in the fund gives the council a benefit residents will actually feel.

2.5 Durability clauses

Transfer: every obligation binds the buyer. Lancaster has this.
Write the terms into the permit. The most useful enforcement move is also the least discussed: incorporate the agreement's commitments into the development agreement, the conditional use permit, or the land use approval itself. The local government can then enforce them as permit conditions, with ordinary municipal remedies, instead of a community group having to fund litigation against a party with far deeper pockets. An agreement enforceable only by a coalition is enforceable only as far as that coalition's legal budget.

Enforcement: notice, a cure period, and a consequence. Lancaster's ladder runs from injunction to discontinuance of operations, with a letter of credit behind the payments. Yale's Sonnenfeld, Siegel, and Henriques add that pledges should be backed by "letters of credit, escrow reserves, or parent-company guarantees" and include a proportional clawback.

Scope the whole agreement to the project as permitted. Not just the covenant not to challenge: the contribution formula, the mitigation limits, and the reporting obligations should all be written against a defined project, described by megawatts, acreage, buildings, water demand, and use. An agreement that says "the facility" without defining it silently absorbs whatever the facility becomes.
Renegotiation on expansion or material change. The companion to a narrow scope is an explicit trigger. A new building, an increase in contracted capacity or water demand beyond a stated threshold, a change of use, a change in cooling technology, or a sale should each reopen the terms rather than extend the old ones by default. This is the clause that decides whether a community negotiates once or negotiates every time the project grows, and it is the difference between a $20 million agreement on a 240 MW campus and the same $20 million on the 900 MW campus it becomes. Define the thresholds numerically so nobody argues later about what counts as material.
Renewal: a scheduled re-opener at year five and every five years, independent of the event triggers above, so priorities can be revisited even when nothing changes. Ten of 72 frameworks in the WRI database have any renewal language.
Decommissioning: a site restoration plan and a decommissioning bond proportional to the project's value rather than a nominal figure, so the community is not left with the cleanup as it was after mining and logging, plus continuity of contributions through closure. Lancaster requires a decommissioning plan. The NAACP template runs payments through decommissioning. Almost nothing else does.
No NDAs: the agreement, the incentive terms, and the compliance reports are public. Microsoft ended NDAs with local governments worldwide in 2026. Pennsylvania's governor cited projects "shrouded in secrecy" when announcing GRID. This is no longer a negotiable term.

3. The Four Structural Choices That Decide Whether Any Of It Happens

Everything in Section 2 can be written into a document and still never happen. The WRI numbers say so: most agreements promise reporting, few structure it, and enforcement is the exception. The difference between an agreement that is kept and one that is not comes down to four choices about structure.

3.1 Where the money sits

Three options appear in practice. The developer holds it and pays grants (the corporate grant model; Meta's Community Action Grants work this way). The municipality holds it (Lancaster's Sustainable Development Fund). A community foundation holds it under a fund agreement (Lancaster's Economic Development Fund at the Lancaster County Community Foundation; the NAACP's dedicated auditable account).

The community foundation is the right answer for most places. It already operates under charitable rules, publishes audited statements, is known to local officials, and cannot be captured by either the developer or a single election. A municipal fund is defensible but exposes the money to the next budget crisis. A developer-held fund is a grant program, not a community program, no matter what the agreement calls it.

3.2 Who decides

The NAACP template's most important sentence is that the developer "shall have no role in approving, vetoing, directing, or otherwise influencing" distributions once the money is deposited. Lancaster routes decisions through the Economic Development Company and City Council. Brookings recommends "community advisory boards for project management."

A resident-majority council with an independent facilitator is the structure that makes the fund a community's rather than a company's. Residents hold the majority. The township, the school district, and the developer may sit, but the developer does not vote on distributions. Members are paid stipends, as the NAACP template requires, because an unpaid board is a board of people who can afford to attend. The council sets priorities from a published needs assessment and recommends grants; the foundation confirms they are within the fund rules and pays them.

3.3 Who reports, how often, and to whom

Lancaster reports once a year to City Council on clean energy. The NAACP template reports quarterly on a dashboard whose metrics the community board sets. Brookings asks for a public dashboard tracking five things. Turner Lee, on the February panel, called for "a uniform standard of transparency" and "public dashboards."

The right cadence is quarterly, published within 30 days of quarter end, with the fund's statement from the foundation attached, every grant named, every mitigation reading posted, and the agreement itself alongside so that any reader can check performance against every commitment. The report is one document for four audiences: the township, the developer's board, the state's annual compliance review, and the neighbors. Separate reports for separate audiences are how promises diverge.

The dashboard must outlive the contract. If it disappears when the consultant leaves or the retainer lapses, it was marketing. It stays online, with its history, for the operating life of the facility and through decommissioning, and that is written into the agreement.

3.4 Who checks

This is the choice no executed agreement has made, and it is the one that answers the February panel's question.

Every mechanism in Sections 3.1 to 3.3 is self-reported. The developer reports, the council reviews, the foundation accounts for the money. That is far better than nothing, but it is the developer grading its own homework with a community proofreader. The NAACP template gets closest with an independent annual audit selected in consultation with the community board and paid by the developer.

What is missing is an independent third party, paid by no one in the transaction, assessing whether the agreement is being kept against a published standard, with a public result that the developer cannot edit and a consequence for misstatement that does not require a lawsuit. That is what the Compute Stewardship Council's draft standard is designed to be for the community pillar: an annual assessment of whether contributions arrived on schedule, whether the council actually decided, whether the reports were published and accurate, and whether every commitment in the executed agreement is being honored, with a public scorecard, and suspension for material misstatement. The scorecard is the enforcement mechanism that does not need a court. A developer's incentive to comply with an agreement nobody publishes is weak. Its incentive to keep a public scorecard from reading "suspended" is strong.

Loudoun County's vice chair said on that panel that enforcement might turn out to be "a nonrelevant factor" because operators now want good community relationships and it is in their market interest to keep their word. He is right, but only if reputation is measured somewhere the market can see it. That is what an independent check does. It turns the operator's self-interest into a mechanism.

4. Our Role, Stated Plainly

Baseload Commons does not negotiate or draft community benefit agreements. Township counsel and the developer's counsel do that, and the templates from the NAACP and, soon, FAS are better starting points than anything a consultant would write. We design what goes in it and we run what it promises: the needs assessment and listening record before signing, the contribution schedule and its sizing rationale, the fund structure, the resident-majority council and its charter, the fund agreement with a community foundation, the quarterly report, and the public dashboard that outlives our contract. We are paid a fixed fee by the developer or the corporate fund. We take no fee tied to a permit outcome, no percentage of the fund, and no role in deciding where the money goes.

The Compute Stewardship Council is a separate, independent nonprofit in formation, convened by the same person who founded Baseload Commons. Its draft standard, in public comment through November 3, 2026, is the "who checks" layer described above. The Council does not design programs, and its assessors may not audit a facility whose program Baseload Commons designed within three years. The founder takes no part in any certification decision involving such a facility. Every scorecard names the program designer a facility used. The full policy is at computestewardship.com/governance, and the standard is at computestewardship.com/standard. We would rather you read the conflict there than discover it later.

5. What A Township Should Ask For, In One Page

Before negotiating

Recognised screening tool run for the site's census tract: air quality, groundwater, hazardous waste, published
Footprint at full build-out: acres, square feet, peak MW, water demand and source, generators
Load ramp by year; projected water and power published 90 days before the first hearing
Tax projections with the abatement modeled and the assumptions public
Funded technical assistance for the township's own advisors

In the agreement

Noise: decibel limit at the property line, quarterly readings, pre- and post-construction baselines
Water: daily cap, metered, posted monthly; closed-loop specified by design
Energy: definition of clean that excludes unbundled certificates; a funded consequence for shortfall
Air: generator run-hour caps and emissions data
Jobs: counts and wages, not percentage goals; funded apprenticeships
Contribution: per-megawatt annual payment on energized capacity, indexed, paid quarterly, for the operating life and through decommissioning; no developer discretion over allocation
Total committed over the term checked against capital investment: 2% a target, 1% a floor
Contributions secured by letter of credit, escrow, or parent-company guarantee
Operator obligations listed separately from fund-eligible priorities, so the community never pays for what it is already owed
Digital access named as an eligible fund priority
Every obligation binds any buyer of the facility
Notice, cure period, remedy ladder, letter of credit or escrow behind the payments
Any release scoped to the project as permitted: no expansions, no waiver of claims for breach, permit violations, or later harms, and no effect on non-signatories
Agreement scoped to the project as permitted, defined by MW, acreage, buildings, water, and use
Renegotiation triggered by expansion, material increase in capacity or water, change of use, change of cooling, or sale, with numeric thresholds
Scheduled re-opener at year five and every five years
Decommissioning and site restoration plan with a bond proportional to project value
No NDAs; the agreement and the incentive terms are public

The structure

Fund held at a community foundation under a written fund agreement
Resident-majority council, independent facilitator, paid members, developer holds no vote on distributions
Quarterly public report within 30 days, foundation statement attached, every grant named, the agreement published alongside
A public dashboard that stays online for the life of the facility
Independent annual assessment against a published standard, with a public result and a consequence for misstatement

Sources

City of Lancaster, Pennsylvania, Data Center Community Benefits Agreement with Chirisa Technology Parks, executed November 2025, and the City's published summary. cityoflancasterpa.gov/data-center/community-benefits-agreement
Nicol Turner Lee and Darrell M. West, "Why community benefit agreements are necessary for data centers," Brookings Institution, January 29, 2026; and the Brookings Center for Technology Innovation webinar of February 17, 2026 with Mike Turner (Loudoun County) and Lily Gangas (Kapor Foundation).
NAACP, Community Benefits Agreement Template for Data Centers, 2026, hosted by the Sabin Center for Climate Change Law, Columbia University; and NAACP, "A Data Center Is Moving In: Now What?"
Sabin Center Climate Law Blog, "Community Benefits Agreements and Data Center Development," May 28, 2026.
Federation of American Scientists, "How We Found Eight Data Center Community Benefit Agreements and Why It Matters," Kate Garman Burns and Stephen Badea, August 2026.
Jeffrey Sonnenfeld, David Siegel, and Stephen Henriques, "How Communities Can Get a Better Deal from Data Centers," Yale Insights, 2026.
World Resources Institute and Data for Progress, Community Benefits Frameworks Database, 72 frameworks, March 2025 edition.
Maryland Association of Counties, Frederick County community benefits agreement with Catellus, announced September 1, 2026.
Commonwealth of Pennsylvania, Governor's Responsible Infrastructure Development (GRID) Standards, May 27, 2026.
Clean Up the River Environment (CURE), "Data Centers: Community Benefit Agreements" webinar, 2026, with Sarah Friedman (Better Data Center Project), Hudson Kingston and Sarah Meridian (CURE).
AI Now Institute, North Star Data Center Policy Toolkit, datacenters.ainowinstitute.org.
The Compute Stewardship Standard, Version 1.0 Public Comment Draft, revised September 5, 2026. computestewardship.com/standard

This guide may be reproduced with attribution. Corrections to hello@baseloadcommons.com.