Letter: UMass’ AI Infrastructure Push Needs A Public Value Test

Tomorrow (Sept. 17), UMass Amherst opens its Sustainable Compute & AI Infrastructure Research Symposium with a useful premise: AI infrastructure should behave as a flexible and responsible participant in the broader public-infrastructure ecosystem, respect existing resource constraints, and serve the public good.

That framing matters because arguments about AI infrastructure often collapse into two camps. One side emphasizes research, economic growth, and technological capability. The other emphasizes electricity, water, construction, and environmental burdens. A university can move the debate forward by asking a more demanding question: what evidence should an AI infrastructure project provide before the public can conclude that it earns its place?

UMass can model a public-value test with four measurable components.

First, measure grid flexibility. If computing demand can shift in time, reduce load during stress, or coordinate with energy availability, publish those results. Average efficiency is useful, but a public-infrastructure participant should also show how it behaves when the surrounding system is constrained.

Second, measure marginal resource use. A more efficient computing system can still increase total demand if use expands quickly enough. UMass is already working through that kind of systems problem in the second phase of its campus decarbonization plan. The university says more than 90% of its direct emissions come from heating, cooling and electrical systems. AI infrastructure therefore belongs in the same planning conversation about energy sources, campus systems and acceptable tradeoffs.

Third, make cost allocation transparent. If additional computing requires new electrical, cooling or other infrastructure, the accounting should make clear where those costs sit: with the project, the university, a utility,or some other party. That does not predetermine whether expansion is worthwhile. It gives decision-makers and the public a clearer picture of the tradeoff they are actually evaluating.

Fourth, measure local benefit with the same seriousness used to measure technical performance. Research output matters, but so do student training, durable workforce skills, knowledge transfer, infrastructure resilience, and partnerships that create value beyond a single project. Benefits should be specific enough to evaluate, not left as general promises.

UMass already has a useful physical example of this approach. The new Sustainable Engineering Laboratories opened this month as a research and teaching facility focused on energy, transportation, batteries and sustainable systems. Its geothermal heat-pump system and research infrastructure make the building itself part of the university’s sustainability work. The point is not that every AI facility should look like that building. The point is that infrastructure can be designed to produce observable public learning alongside its direct function.

This is partly a behavioral problem. People are more likely to trust consequential technological change when costs, responsibilities, and benefits are visible. Resistance grows when benefits are aspirational while burdens are concrete, or when accountability becomes hard to locate.

A public-value scorecard would turn those concerns into evidence. UMass could publish a small set of indicators for major AI infrastructure projects: peak-load flexibility, marginal energy and water demand, cost responsibility, research and training outputs, and community-facing benefits. The exact measures can evolve as the technology changes. The discipline of measuring them should remain.

The SCAI symposium already brings together researchers, technology companies and institutions around the relationship between AI computing and public infrastructure. Amherst is therefore a good place to move beyond abstract claims about whether AI infrastructure is good or bad. The better standard is whether a specific project can demonstrate, in public, that its benefits justify its demands.

Gleb Tsipursky, PhD, is a behavioral scientist, CEO of Disaster Avoidance Experts, and author of The Psychology of AI Adoption at Work: From Resistance to Results (Georgetown University Press, 2026).