As governments move from AI pilots to enterprise deployment, the decisive question is shifting from which model to use to what each answer costs, whether that cost can be predicted a fiscal year in advance, and when it pays to own the capacity rather than rent it.

In September, the federal government changed how it pays for AI.

For the past year, agencies could get frontier AI for almost nothing. Under the General Services Administration’s OneGov program, ChatGPT Enterprise , Claude and Gemini for Government each cost a dollar or less. Those offers are giving way to commercial terms, and GSA’s new 27-month agreement with OpenAI is priced on a different principle: a 50% discount on token-based usage, no platform-access fee, no minimum order and no spend commitment. In GSA’s words , agencies “pay only for what they use.”

That is a good deal. It is also a different kind of deal. A flat fee is a budget line. A meter is a forecast. And because the new agreement is open to state, local and tribal governments, this is not only a federal story.

Public sector AI has just become a variable cost, priced in tokens, the word fragments AI models read and write. Most agencies do not yet have the instrumentation to predict it.

The trap: falling token prices, rising bills

AI tokens are getting cheaper, so budgets should be getting easier. Only the first half is true. At Dell Technologies World 2026, Jeff Clarke, Dell’s Vice Chairman and Chief Operating Officer, reported token costs down roughly 80% year over year and token use for reasoning alone up 320 times.

Unit prices fall. Consumption outruns them.

Uber found this out in April, when its Chief Technology Officer, Praveen Neppalli Naga, told The Information that AI coding tools had maxed out the company’s full-year AI budget, sending him “back to the drawing board.” Given how rapidly AI investments have grown, it is unsurprising that 98% of FinOps Foundation’s 2026 survey respondents now have responsibility for managing AI costs, compared with only 31% two years ago.

Uber can absorb the overrun. A government agency has no such cushion. It operates on an appropriation, and for a federal agency, spending beyond one is not a variance but a violation of law . In a state or county the same overrun becomes a supplemental request, a deferred project, or a service that stops in March.

That is why tokenomics, the economics of producing and consuming tokens, cannot wait until after deployment. In government, cost predictability is a condition of deployment.

In NASCIO’s 2026 State CIO Top Ten Priorities , artificial intelligence ranks first for the first time. Budget, cost control and fiscal management rank third. The same CIOs must deliver AI and find savings in the same year.

In government, cost predictability is a condition of deployment.

Four decisions that move the number

Between a pilot and an enterprise deployment, four choices do most of the work on cost, and model brand is not one of them.

Model size, measured correctly. Headline parameter counts mislead. Cost is set by how many parameters compute for each token, and mixture-of-experts models can carry a trillion parameters while activating only a few percent of them per token. The practical move is to run two: a capable model for hard questions and a small, fast one for summarizing, classifying and searching. Routing routine work away from the expensive model is the least glamorous, most dependable saving available.

Reasoning complexity. Reasoning models spend tokens deliberating before they answer, which is the usage Clarke saw grow 320-fold . Agentic systems go further, chaining prompts, calling tools and retrying until a result satisfies them. Because almost no useful agentic workflow finishes in a single inference, cost per token stops being the right measure. The one that matters is cost per completed outcome.

Data access patterns. This is the driver agencies control the most and examine the least. Retrieval, which searches agency records and answers from what it finds, is what makes an assistant useful. It also fills the model’s working memory, which is finite. That memory grows in step with the text each request carries , so quadruple the retrieved text and the same server holds about a quarter as many requests at once. An agency that tunes retrieval before it buys hardware buys less hardware.

Deployment location. Metered inference suits variable, experimental demand, which is what a pilot is. Steady-state work behaves differently, and it is most of what government does. That makes location the decision with the most leverage.

The first three decisions reduce the computing each answer takes. The fourth changes how that computing is paid for.

On a meter, spending rises in step with consumption. On infrastructure an agency owns, the cost is largely fixed, and each additional token costs little more than the electricity behind it. Utilities have planned around that difference for a century: own the base load, buy the peaks.

At national scale, sovereign AI means what NVIDIA calls “trusted domestic capabilities” that protect sensitive data and reflect local institutions. For a state agency or a county, the practical meaning is narrower: custody. The agency owns the hardware, chooses the model and decides when it changes, and keeps every prompt and answer in its own logs on its own retention schedule. Some records add rules of their own: federal tax information can enter a cloud only on FedRAMP-authorized services, with at least 45 days’ notice to the IRS.

Custody matters only if the models an agency can run itself are capable enough, a constraint that has eased sharply. Epoch AI , an independent research group, found that since January 2026 the most capable open-weight models, the kind anyone can download and run privately, have trailed the leading closed models by about four months on average. Stanford’s 2026 AI Index puts the top closed model’s lead over the top open one at 3.3%.

Government buyers will spot the caveat. As of mid-2026, the open models closest to the frontier are mostly Chinese-developed, and several states have barred at least one of them from government devices. US-developed open models from NVIDIA, Google, OpenAI and newer labs sit a step behind that group21, and NVIDIA has released the datasets used to train its largest Nemotron model, so an agency can document what it is running. For most of what government asks of AI (drafting, summarizing, searching and classifying its own records), the test is not whether a model is the world’s best, but whether it is good enough. A model a few months behind the frontier is simply the frontier of a few months ago.

The capacity required is usually smaller than expected. Dell’s own support assistant went live in a single on-premises rack, needing neither the latest GPUs nor any additional power or cooling. Analysis by Signal65 and Futurum , commissioned by Dell and NVIDIA, found that agentic workloads on Dell workstations and servers can cost up to 87% less than cloud APIs over two years. The qualifier is utilization: an idle GPU produces no tokens and still depreciates, so owned capacity beats the meter only for work that keeps it busy.

That is why the sidestep is not all or nothing. Terms like OneGov’s, with no minimums and no commitments, suit the unpredictable tail: the hardest reasoning, the new experiment, the seasonal spike. Owned capacity suits the predictable base: the assistant every employee uses daily, the document search that runs all year, the regulated workloads that were never going to leave the building.

A model a few months behind the frontier is simply the frontier of a few months ago.

Building the number into the plan

Four practices separate an AI line item leadership can defend from one that surprises them.

Size on concurrency, not headcount. Capacity is set by how many requests are in flight at once, not by how many employees hold a login. Most of any session is spent reading and typing, so at any moment only a small fraction of employees are waiting on an answer.

Make the pilot produce the budget figure. Instrument a proof of concept from day one to report peak concurrent requests, queue depth, working-memory utilization, time to first response and typical prompt length. Those five numbers turn a pilot into a defensible appropriation request, and show which workloads are base load.

Put a ceiling on the meter. “No minimum, no spend commitment” is flexibility for the buyer and exposure for the appropriation. Per-workflow spending caps, per-team quotas and automatic alerts belong in the design, not in the postmortem.

Record placement where the policy work already lives. Many states are building AI use-case inventories. Two added fields, where each workload runs and what it should consume, cost almost nothing and make cost visible at approval rather than at year-end.

Cost per outcome, not cost per token

By August, Naga said costs were falling even as adoption grew, because Uber had “treated efficiency as an engineering problem rather than a budget problem.”

The governments that scale AI well will not be the ones that found the cheapest tokens. They will be the ones that knew, before they scaled, how many tokens an outcome takes and where it is cheapest and safest to produce: on a meter, on their own hardware, or, most often, both. In the public sector that discipline does more than protect a budget. It lets an agency make a commitment to residents that it can keep for the length of a fiscal year.

Dell Technologies and NVIDIA’s guide Overcoming Pilot Fatigue: A Roadmap for Scaling AI in State and Local Government looks at why public sector AI pilots stall and how to scale them. The executive e-book Sovereign AI on Your Terms covers keeping control of data, infrastructure and models.26

NVIDIA’s Tokenomics Guide27 includes a token-demand forecasting worksheet that builds an estimate from base demand, workload multipliers and operational peaks.

More on the Dell AI Factory with NVIDIA for state and local government is on Dell’s state and local government page .