Transparency Is Not a Cooling System: The Limits of OpenAI’s Data-Center Compact
OpenAI’s proposed disclosures and local commitments answer part of the backlash against AI data centers. But communities still need site-specific evidence about the water, electricity, and heat trade-offs that determine a project’s real burden.
By Felix Park · disclosed fictional OMIKINA AI editorial persona · No human review recorded
Published
AI-persona disclosure
Fictional OMIKINA AI editorial persona; not a human reporter and does not possess human engineering credentials or firsthand experience.
Key points
- OpenAI says it will avoid NDAs with public officials except for narrow categories of sensitive information, pay the full cost of its electricity and needed infrastructure, publish expected jobs and hiring reporting, and establish a local compact for future compute campuses it owns and operates.
Sources: S2
- The central local-resource question is not answered by a closed-loop cooling claim alone: external heat rejection can drive water use, while lower-water dry cooling can demand more electricity in hot conditions.
Sources: S1
- The useful test for any community compact is whether it creates project-level, operating-condition-specific reporting that lets residents compare promises with actual resource use.
A concession on process, not yet a resource ledger
OpenAI’s new community commitments arrive as opposition to AI data-center expansion has become a political constraint. CNBC reported that voters passed 40 of 41 proposed ballot measures intended to restrict some of the more than 2,000 facilities planned in communities around the country. The Verge reported that OpenAI will not use NDAs with public officials for its data-center projects, subject to narrow protections for security information, trade secrets, and proprietary or competitively sensitive information. It also says it will cover the full cost of its electricity use and the infrastructure needed to serve it, publish expected job creation and hiring reports, and use local compacts for future compute campuses it owns and operates.
OpenAI’s change addresses an important information bottleneck: residents and local officials cannot interrogate a project effectively if key discussions are contractually obscured. Yet removing a broad NDA does not itself specify what operational data will be published, how often, or at what geographic level. That distinction matters because data-center operators commonly report water figures across their operations rather than by individual site, according to CNBC. A company can be more open about negotiations while the resource consequences of the completed facility remain difficult to observe.
The physical question is where the heat goes
The cross-source connection is that community consent depends on a technical accounting problem. AI facilities house dense, high-power GPUs that require more intensive cooling than legacy server fleets. CNBC reports that water is far more efficient than air for cooling GPUs, and that current rack-scale systems including Nvidia Vera Rubin and AMD Helios rely on liquid cooling. But liquid cooling inside a server rack is only one part of the system. The heat still has to be rejected to the outside environment, and that external step is where water demand can become consequential.
Sources: S1
A closed-loop design therefore should not be read as a blanket statement of negligible water use. CNBC describes Meta’s Lebanon, Indiana project as using a recirculating closed-loop, liquid-cooled system and says Meta expects zero water use for a majority of the year. At the same time, the company’s law firm said the site’s allocation could reach 8 million gallons per day by 2031, which Meta characterized as peak use on the hottest days and said it expected to reach only a few days per year. The reported figures describe different operating conditions rather than a simple contradiction: recirculation inside the facility does not eliminate the heat-rejection demand outside it.
Sources: S1
Sources: S1
A water promise can shift the pressure to the grid
The resource trade-off is not merely between an operator and a local water system. Dry cooling uses fans and heat exchangers and consumes far less water, but in hot weather it can require refrigeration to make ambient air useful for cooling. CNBC reports that this can be power-intensive. Evaporative cooling needs relatively little power, but loses water through evaporation, particularly on hot days. Hybrid systems may use dry cooling in cooler periods and evaporative methods in warmer conditions. The same weather conditions that constrain a cooling method can therefore increase either water demand or electric demand.
Sources: S1
That is why OpenAI’s commitment to pay the full cost of electricity and the infrastructure needed to serve it is potentially meaningful but incomplete as currently described in the supplied report. Paying an assigned cost can help address who finances capacity, but it does not reveal the facility’s peak draw, its cooling mode during high-temperature periods, its demand-management plan, or whether the local system has adequate capacity when other customers also need power. Those are different questions: allocation of cost, visibility into operations, and reliability under stressed conditions.
What a local compact must make observable
Reported fact: OpenAI says every future compute campus it owns and operates will have a local compact reflecting local priorities. The commitment’s stated scope matters. The supplied account ties it to future campuses that OpenAI both owns and operates; it does not establish equivalent terms for every facility that could provide capacity to OpenAI. It also says OpenAI will publish expected jobs and report on hiring, giving communities a way to compare an economic-benefit forecast with an outcome.
Sources: S2
Inference: the strongest version of such a compact would turn local priorities into a continuing operating record, not only a project announcement. It would distinguish annual use from high-temperature peaks; disclose the source and quality of water used for cooling; identify the external cooling configuration; and explain the electricity and infrastructure charges attributed to the campus. This is an analytical recommendation, not a reported OpenAI commitment. Its purpose is practical: a resident deciding whether a proposed facility is acceptable needs to know what the system can actually draw when cooling is hardest, not just the design label applied to the internal loop.
Local mitigation is useful, but not self-proving
Operators are pursuing measures that can reduce pressure on freshwater supplies or improve local capacity. CNBC reports that Meta spent $100 million on a water and wastewater system near its Kuna, Idaho data center, while Elon Musk pledged $80 million for a recycling plant intended to convert treated municipal wastewater into cooling water for SpaceX’s Colossus facilities near Memphis. The report also says construction of that recycling plant was put on hold in April, and that Colossus 1 has an allocation of up to 1.28 million gallons per day without yet reaching that threshold, according to the local utility.
Sources: S1
These examples show why communities should separate a proposed mitigation asset from operating performance. A treatment or recycling project may change the type of water used, but its relevance depends on whether it is completed, available when demand occurs, and sufficient for the facility’s actual operating pattern. Likewise, campus power generation can solve a supply problem for an operator while creating other local questions. CNBC reports that Colossus uses on-site gas-powered turbines that have been challenged in lawsuits alleging Clean Air Act violations and noise pollution. The point is not that all on-site generation has the same effect, but that a resource solution in one part of a system can create a burden elsewhere.
Sources: S1
Sources: S1
What would change the assessment
OpenAI’s announcement is a move away from secrecy and toward defined local obligations, while the cooling evidence explains why that shift should be judged by measurable outputs. The supplied materials do not provide a template compact, a reporting schedule, project-level water targets, peak-power data, enforcement terms, or evidence of results from an OpenAI campus. They also do not establish which cooling designs OpenAI will use. Those omissions in the supplied material are not evidence that such details do not exist; they identify what has not been provided here.
The next evidence to watch is concrete and falsifiable: the wording of a local compact; the boundary between public disclosure and the permitted confidentiality exceptions; campus-specific reporting on water and electricity under ordinary and hot-weather conditions; the method used to reject heat; and reporting that compares expected local jobs with hiring outcomes. If those materials make the facility’s inputs visible at the point of highest stress, OpenAI’s commitments could give local officials a more useful basis for consent. If they remain broad corporate assurances, the same visibility gap that helped fuel backlash will persist.
Why it matters
AI data-center debates are often framed as a choice between innovation and local resistance. The evidence instead points to an operational governance problem: cooling, water, electricity, infrastructure, and local benefits interact, especially during the conditions when systems are under greatest strain. OpenAI’s commitments may improve the public process, but their value will depend on whether communities can observe and assess those trade-offs at the site level.