Securing Infrastructure Before You Need It
Negotiate compute when leverage is real — before inference costs compound indefinitely.
The Economic Case
The true cost of AI in government is not model development — it's inference. A model trained once can serve a city for decades, but every query to a commercial API compounds that dependency indefinitely. The infrastructure buildout has accelerated: 2026 estimates now range from roughly $785 billion in hyperscaler capex to more than $1 trillion in global AI investment, with analyst projections pushing toward trillion-dollar annual data-center spending in 2027. By negotiating compute set-asides during data center permitting — while municipalities hold rare leverage over power-hungry facilities — cities can lock in the infrastructure they'll need before the models are even built.
The sequencing matters: secure the compute now, develop the models on whatever timeline funding allows, and deploy into an infrastructure commitment already in place. The marginal cost of inference is locked in at the moment of greatest negotiating power.
What to Ask For in the Next Permit
Municipalities currently negotiate traffic studies, affordable housing set-asides, and stormwater infrastructure as permit conditions. Compute access is no different in principle.
Early data-center Community Benefits Agreements are beginning to prove the leverage exists, but they usually still stop at money, jobs, water, power, transparency, and mitigation. Lancaster, Pennsylvania's AI Hub CBA secured $20.25 million in community and clean-energy contributions, water-use limits, public-records treatment, and operating conditions. St. Louis approved a 2026 data-center permit with conditions and a community-benefit framework. Those are important precedents. They are not yet transferable public GPU-hour allocations.
"As a condition of approval, Applicant shall reserve not less than [X]% of total GPU compute capacity — or an equivalent allocation of [Y] GPU-hours per month — for use by [Municipality] for public-interest AI workloads. This allocation shall be subject to an SLA guaranteeing [Z]% uptime, shall survive any change of operator or ownership of the facility, and shall be transferable to a successor public entity. Specific workload types, access protocols, and performance standards shall be defined in a Compute Access Agreement to be executed prior to certificate of occupancy."
This is a starting point for negotiation, not legal advice. A city attorney should adapt it to local permitting law. The distinctive ask is the durable compute right: a public allocation that survives ownership changes and can be used by the municipality or a successor public entity.
The Cost You're Not Counting
The true long-term cost of commercial AI is not the contract or the integration — it's every inference call, forever. Frame this as a present-value problem: a municipality that secures compute access now, when leverage is highest, is hedging decades of operational cost at today's negotiating position.
Commercial API dependency remains structurally volatile. Models retire, pricing changes, access terms shift, and public agencies are left to migrate critical workflows on vendor timelines. OpenAI's 2026 retirement notices and Assistants API sunset are normal platform lifecycle management for a private vendor; for a public service, they are also a reminder that continuity cannot depend entirely on another party's roadmap.
Sources: Moody's / Data Center Dynamics, May 2026; Goldman Sachs Research, August 2026; City of Lancaster data-center CBA materials; Columbia Climate Law Blog, May 2026; City of St. Louis, April 2026; OpenAI model and Assistants API retirement notices.