When a company first tries generative AI, the cost is easy to ignore: a few licenses, a pilot, a modest cloud bill. When the same company puts AI into daily operations, usage becomes a real expense that grows with every new agent, workflow and employee. At a Dell Technologies event in New York this week, Dell’s chief financial officer, David Kennedy, described how finance chiefs are responding, according to SiliconANGLE’s coverage: by treating token consumption as a budget line that has to earn its keep. This article explains what that means in practice and how to apply it.

Note that the source is a vendor’s executive speaking at a vendor-sponsored event, so treat the claims as one perspective rather than independent evidence.

What Kennedy said

According to SiliconANGLE, Kennedy argued that enterprise AI is moving from experimentation into production, and that financial services firms, especially in high-frequency trading, were early to push it to scale, with pharmaceutical companies likely next. He said most chief financial officers are now creating a dedicated line item for token spend and tracking it like any other cost.

His message was that the spend “needs to be managed,” with IT and finance putting guardrails around it, but that the real value driver is the speed and productivity a business gains. In his words, it is “not necessarily about the amount and the cost of a token” but about “the tokens you use.” Dell also said its customers run 6,500 AI factories, a figure we have not verified.

One concrete suggestion was model placement: repetitive finance workflows can run on open-weight models on-premises, turning variable token costs into fixed ones.

Why tokens need their own line

A token is the unit models use to read and write text, and most API pricing is quoted per million tokens, with different rates for input and output. You can compare current examples on vendor pages such as OpenAI’s API pricing and Anthropic’s pricing. Unlike a software license, a token bill moves with behavior: longer prompts, bigger documents, multi-step agents and retries all raise it. Our guide to how AI pricing works covers tokens, subscriptions and credits.

Because of that variability, a single “AI” line hidden in cloud or software spend makes it hard to know who is using what. A separate line lets finance forecast, compare with outcomes and spot surprises.

A simple framework for planning AI spend

  1. Tag every workload. Give each application or team its own API key, project or account so usage can be attributed. Without tags, you cannot assign cost.
  2. Measure cost per task, not per token. A cheap model that needs many retries can cost more than a stronger one that succeeds the first time; see our opinion piece on why cheap tokens do not make cheap agents.
  3. Set budgets and alerts. Choose a monthly ceiling per team and trigger alerts at fractions of it, such as 50% and 80%.
  4. Forecast with scenarios. Estimate volume (tasks per month), tokens per task and price, then model low, expected and high cases. Include growth, since adoption tends to rise.
  5. Review value, not just cost. Compare spend with hours saved, errors reduced or revenue gained, using a baseline measured before the rollout.

These steps resemble FinOps, a practice for managing variable cloud spending that the FinOps Foundation describes as bringing finance, engineering and business teams together to make spending decisions. Many organizations can extend their existing cloud cost practices to cover AI.

Levers that reduce cost

  • Right-size the model. Route simple tasks to small models and reserve large ones for hard problems.
  • Trim prompts. Remove unneeded context and boilerplate; long inputs cost money on every call.
  • Cache repeated content. Some providers discount cached input; check each vendor’s terms.
  • Batch non-urgent work. Offline jobs are sometimes priced lower than real-time calls.
  • Run steady workloads on owned or reserved capacity. This is Kennedy’s variable-to-fixed point, but it adds hardware, staffing and maintenance costs that must be counted. Our piece on building, buying or fine-tuning discusses the trade-offs.

Risks to manage

Cost control can backfire. Caps that cut off a workflow in the middle of a task may cause failures. Pushing every team to the cheapest model can reduce quality. And counting only visible API costs misses the staff time spent on setup, evaluation and oversight. Finance and engineering should agree on a small set of measures, such as cost per completed task and quality score, before adding limits.

What to watch next

  • Whether more finance leaders publish how they allocate AI spend.
  • New tooling from cloud and AI vendors for per-team usage reporting.
  • Price changes from model providers as inference gets cheaper.
  • Shifts toward on-premises open-weight models for steady workloads.

Key takeaways

  • Companies are giving token spend its own budget line and tracking it like other costs.
  • Tag usage by team and workload so you can attribute cost.
  • Measure cost per task and compare it with business value.
  • Route work to right-sized models and trim context to cut spend.
  • Owned capacity can turn variable costs into fixed ones, but it has its own costs.

Frequently asked questions

What is a token?

A token is a small piece of text, often part of a word, that models process. APIs usually bill per million input and output tokens.

Should we always pick the cheapest model?

No. Compare the total cost of completing a task, including retries and review, and the quality of the results.

Is on-premises always cheaper?

Not necessarily. It can lower costs for steady, high-volume workloads, but you must account for hardware, power and staff.