The Compute Stewardship Standard

How To Read This Draft

This is a scored framework released for public comment and founding member review. It is not yet a certification instrument. No facility has been assessed against it, and none will be until the Council exists as an incorporated nonprofit with an independent board and a seated Certification Committee (see the Governance and Independence Policy).

Thresholds in this draft are set where a well-run facility can reach them today with real effort. They are expected to tighten on the two-year revision cycle, and Section 10 lists the specific numbers we most want challenged. A standard that no facility can meet is a manifesto. A standard that every facility can meet is a directory. This draft aims between those, and we would rather be told where it lands wrong than have it politely ignored.

1. Purpose

To provide independent, audited verification that an AI compute facility operates without depleting the resources of its host region and delivers measurable, ongoing, funded benefit to that region.

Certification produces a public scorecard that enterprises may cite in voluntary sustainability reporting and that host communities may use to hold operators to account.

2. Scope

Applies to facilities at or above 500 kW of IT load hosting AI training or inference workloads, whether colocation, dedicated, or operator-owned.

Certification is granted per facility, never per company. No company-wide claim may be made on the basis of a single certified site. In colocation, the facility operator holds the certification; tenants may cite it for workloads physically located there.

3. Definitions

4. Scoring Model

Each pillar is scored out of 20 points, for a total of 100. Tiers are set by total score and by minimum floors in individual pillars, so a facility cannot compensate for ignoring one pillar by excelling in another.

Lower tiers are not failing grades. They are published as a position on a path, and the scorecard shows where the points were lost.

Note on the floors. Siting and Local Environment and Fiscal Accountability carry floors at every tier, including Entry. A facility that keeps its neighbors awake, runs uncontrolled generators, or will not publish what its tax abatement is worth is not responsibly operated whatever its water and energy numbers say, and should not carry any certification issued by this Council.

5. Pillar One: Water and Resource Efficiency (20 points)

Beneficial heat reuse includes district heating, agriculture, aquaculture, industrial process heat, and municipal facilities. Reuse inside the facility's own envelope does not count.

Evidence: cooling schematics and as-built drawings; water meter data for the trailing 12 months; water withdrawal and discharge permits; water authority correspondence; heat export metering and offtake agreements; site inspection of loop integrity, makeup water metering, any evaporative equipment, and the heat export tie-in.

Note on the weighting of heat recovery. Earlier drafts of this standard gave heat recovery a full quarter of the total score. That weighting was a bet that heat reuse would become the durable differentiator once closed-loop cooling went mainstream. It may still, and the item bands are set so that a facility exporting heat is clearly rewarded. But in the public record of what host communities actually raise, heat recovery is close to absent while noise, air quality, siting, and fiscal impact are near universal. Six points reflects that evidence without abandoning the forward position, and the bands are expected to tighten as US reuse becomes achievable.

6. Pillar Two: Additive Energy (20 points)

Evidence: power purchase agreements and their commercial operation dates; hourly consumption and generation data or granular certificates; interconnection agreement and utility filings; the ratepayer impact analysis and its author's independence statement; grid program enrollment records.

Note on the load ramp: electricity service agreements are typically proprietary, and utilities now quote four to seven years to deliver full contracted power in constrained markets. A community cannot assess a facility's impact from its phase-one number. The ramp schedule is the disclosure that matters.

Note on E4 and generators: demand response earns a point only where the flexibility comes from the load itself or from storage. Meeting a demand response call by starting diesel engines moves an emission from a power plant to a neighborhood, and is scored under item L3, not here.

Assessor note: the distinction between additive and reallocated supply is the single easiest claim to overstate. The Energy Analyst verifies commercial operation dates against interconnection queue records and utility filings, not against the operator's summary.

Context for comment: the largest and best-resourced operator in the world reports roughly 66% hourly matching globally and about 70% in North America as of 2024. The 85% band is intended to be reached by only a small number of facilities in this revision cycle.

7. Pillar Three: Siting and Local Environment (20 points)

What this pillar assesses. Where the facility is, and what a resident standing at the property line experiences: what they hear at night, what they breathe when the generators run, what the sky looks like, and what the site looks like from the road. These are the impacts residents raise first and that most facility standards score last or not at all. An elected official responsible for more than two hundred data centers described noise as the single largest community complaint in the sector.

Evidence: the screening tool output for the site's census tract with the assessment date; environmental site assessment and remediation records; parcel and land use records showing prior classification; the pre-construction ambient survey and every subsequent monitoring report, with the monitoring contractor's engagement letter; air permits and engine tier certifications; generator run-hour logs reconciled against permit limits; the photometric plan and a night-time site inspection; as-built site plans and a boundary walk.

Note on S1. One state's analysis found that more than eight in ten of its data centers sit in areas already carrying poor air quality, and roughly eight in ten in the top quintile for hazardous waste burden. A standard that scores only what a facility emits, and never where it chose to emit it, would certify the concentration of burden it exists to prevent. Item S1 does not forbid siting in a burdened community; it requires that the choice be examined in public and mitigated with the people who live there.

Note on generators. The scale of permitted backup generation is now a public health question in its own right rather than a technicality: one state has permitted enough data center backup capacity to serve several times its own housing stock. Engines must be run periodically to remain reliable, so even a facility that never suffers an outage produces emissions from testing alone, and at campus scale those testing hours approach the output of a small power plant. Item L3 therefore scores the cap and the prohibition, not only the equipment.

Note on L1. The limit is expressed against pre-construction ambient because a rural night is quiet, and an absolute limit alone would let a facility raise a 30 dBA night to 50 dBA and call it compliance. The absolute ceiling is there because an already-noisy parcel should not license an unlimited addition. A facility is measured against whichever of the two is stricter.

8. Pillar Four: Fiscal Accountability (20 points)

What this pillar assesses. Whether the public can see what the facility costs the public, and whether the promises made to the community are backed by anything. Roughly thirty-five states exempt data center equipment from sales and use tax; one state forwent an estimated $1.6 billion in a single year; another state's audit found a net fiscal loss of $433 million. A facility can hold a thirty-year abatement, publish nothing about its value, and still describe itself as an economic benefit. This pillar makes the arithmetic public and requires the commitments to be secured.

Evidence: executed abatement, exemption, and PILOT agreements; the fiscal impact analysis and its author's independence statement; the security instrument and its issuing bank or guarantor; the emergency services agreement and the responding authority's acknowledgment; the traffic study and the road improvement agreement.

Note on F3. A community benefit agreement is a promise to pay for twenty-five years, and it is usually made by a project company created for the purpose. Lancaster, Pennsylvania secured its contributions with a letter of credit, and independent analysts have argued that letters of credit, escrow reserves, or parent-company guarantees should be a universal condition. An unsecured promise from a special-purpose entity is worth what the entity is worth on the day it is called.

Note on F1 and NDAs. This item follows Microsoft's 2026 decision to end nondisclosure agreements with local governments and the emerging consensus that subsidy arrangements must be public. It sits in this pillar rather than the community pillar because secrecy about money is a fiscal failure before it is a relationship failure.

9. Pillar Five: Verified Community Benefit (20 points)

What this pillar assesses. Whether a facility's commitments to its host community are real, funded, proportionate, governed by residents, and kept. Where a community benefit agreement, host community agreement, or community benefit plan exists, this pillar does not judge how it was negotiated; it verifies that it is being carried out: contributions received on the agreed schedule, decisions made by the body the agreement names, results reported in public, and every commitment traceable to a published record. The Council does not draft or negotiate agreements. It is the independent check that the agreement is being honored, which is the function most agreements lack.

Note on C4, proportionality. This is the item that decides whether the rest of the pillar means anything. A twenty billion dollar campus contributing fifty thousand dollars a year would, without it, score exactly what one contributing two million scores. The measure is expressed against capital investment rather than revenue because facility-level revenue is not observable from outside, while capital investment appears in filings, financings, and the incentive applications item F1 requires to be published. The 2.0% and 1.0% bands are the least tested numbers in this draft and are specifically put out for comment. For reference, the most complete publicly posted data center agreement commits roughly one to two percent of estimated capital investment over its term, and an experienced practitioner has publicly described its finances as not strong for a project of that size.

Note on C8. Item C3 treats equipment refresh as a contribution trigger, which is an acknowledgment that servers, chips, and electrical plant turn over every three to seven years at very large scale. A standard that captures the money from that cycle and says nothing about where the hardware goes would be incoherent.

Explicitly insufficient: one-time philanthropy scores zero. A scholarship or a fire department donation is welcome and irrelevant to this pillar. Emergency services support is scored under item F4 as an obligation, not here as a benefit.

Fund structure: contributions are held in a restricted fund at a community foundation or equivalent fiscal agent in the host region, under a fund agreement with the Council. The advisory council recommends distributions; the Council confirms they fall within the fund rules; the fiscal agent disburses. The Council's Community Assessor audits this pillar and has no role in facilitating the council or administering the fund.

10. Maintenance, Suspension, and Publication

• Annual re-audit, with a full reassessment every third year
• Public scorecard for every certified facility, showing points by item, tier, and the name of any recognized program designer used
• Material misstatement triggers immediate suspension, not a remediation period; the scorecard shows the suspension
• A facility that drops below its tier floor at re-audit is re-tiered, not suspended
• Complaints from residents or local officials are accepted through a public channel and may trigger an unscheduled review
• Decommissioning and continuity (condition of certification at Provisional tier and above): the operator files a decommissioning and site restoration plan with a bond or equivalent financial assurance proportional to the project's capital value rather than a nominal figure, and the fund agreement provides for committed contributions to continue through closure and for the fund's remaining balance to stay in the host region under the advisory council's direction. A facility without a filed plan is capped at Entry tier.
• Permanence of the record: every scorecard, every quarterly fund statement, and every assessment summary remains published after a facility is re-tiered, suspended, sold, or decommissioned. The record belongs to the host region, not to the operator or to the Council's contract with it, and it is never removed at an operator's request.

11. Open Questions for Comment

These are the numbers and definitions we expect to be challenged, and want to be.

1. Is 85% hourly matched CFE the right Full-tier band for this revision cycle, given current North American performance?
2. Is 36 months the right window for additivity, and should the definition require that the facility's load was named in the generation project's financing?
3. Is ERF 0.15 reachable in the US within three years by any facility not adjacent to a district heating network? If not, is export readiness (H3) rewarded enough?
4. Is a resident majority the right composition for the advisory council in regions where the facility is the dominant employer?
5. Is 1.0% from each party the right matched contribution level? What level makes an enterprise decline to enroll?
6. Should colocation tenants be able to earn a tenant-level citation, and on what basis?
7. Is the 500 kW scope floor correct, or should it follow the EU's delegated regulation threshold exactly?
8. Should decommissioning be a scored item rather than a condition of certification, and what form of financial assurance is realistic for a facility whose operating life is uncertain?
9. Should replenishment or watershed restoration outside the host region earn any credit in Pillar One, or does crediting it repeat the substitution problem that Pillar Two rejects for unbundled certificates?
10. Is 45 dBA at night, referenced to pre-construction ambient plus 5, the right band for item L1 in a rural setting, and is it the right test in an already-industrial one?
11. Item C4 asks whether committed community contributions reach 2.0% of total capital investment, with a lower band at 1.0%. These are the least tested numbers in this draft. Are they the right order of magnitude, should the measure be capital investment at all rather than a per-megawatt figure or a share of revenue, and how should a facility be scored where the operator will not disclose capital investment?
12. Should item S1 forbid certification outright for a facility sited in a highest-quintile burden tract, rather than allowing 2 points where a cumulative impact assessment is published and mitigation agreed? A standard that can certify a facility in an already-overburdened community may be legitimising the concentration it exists to prevent.
13. Is two years of committed payments the right security requirement in item F3, and should the requirement scale with the length of the agreement?
14. Heat recovery now carries 6 of 100 points, down from 25 in the first draft. Is that too far? The evidence for the reduction is that host communities raise it almost never while raising noise, air, siting, and fiscal impact constantly. The argument against is that a standard should lead rather than follow.
15. Where a community benefit agreement exists and includes a covenant not to challenge the approvals, should the standard require that covenant to be narrowly scoped: limited to the project as permitted, excluding expansions, and preserving claims for breach of the agreement and for later permit, environmental, or workforce violations? A facility whose agreement swept more broadly than that would score the same today as one whose agreement did not.

12. Governance

The standard is maintained by the Compute Stewardship Council, an independent nonprofit in formation. Governance, board composition, conflict of interest rules, and the founder's disclosed relationship with a recognized program designer are set out in the Governance and Independence Policy. In summary:

• Founding operator members hold advisory seats on the Standards Advisory Committee and have no deciding vote
• Certification decisions are made by an independent Certification Committee
• Criteria are revised on a two-year cycle with a 60-day public comment period
• No assessor may audit a facility they helped design or advise

Version History