Vijayanta Gupta

Operator note

Token Budget Versus Headcount. The New Planning Trade-Off Every CFO Will Run in 2027.

.6 min read

The New Planning Trade-Off Every CFO Will Run in 2027

Next year’s operating plan will carry a line the SaaS-era plan never had. Beneath headcount, beside cloud and tooling, sits a token budget and an inference cost line, broken out by function. Most finance teams will book it as procurement and move on. That is the mistake. The number is small today and easy to file under cloud spend. The decision behind it is not small. It is the first time a leadership team can convert operating capacity into compute instead of people, function by function, and decide the mix on purpose.

The surface question is what the token line equals in dollars. The real question is the trade. For a given function, the choice is no longer only how many people to hire. It is whether the next unit of capacity comes as a person or as a block of tokens for the people already there. In field conversations this year, operators are already running that math at unit level: roughly thirty thousand dollars of annual token budget against a developer who costs well over a hundred thousand fully loaded. The comparison is crude, and it is exactly the comparison the plan now has to make explicit.

The usual diagnosis treats inference cost as a volatile input to be controlled. Finance asks engineering to cap usage, sets a budget, and watches the line. That framing is wrong twice. It is wrong on direction of cost, and it is wrong on what the line is for.

Start with cost. The instinct is to fear a runaway compute bill. The longer trend points the other way. For a model of equivalent performance, inference cost has fallen by roughly an order of magnitude a year since 2021, from sixty dollars per million tokens to a few cents over three years (Andreessen Horowitz, November 2024). Independent measurement puts the decline between nine and nine hundred times per year depending on the task, with a median near fifty times and the fastest drops arriving after January 2024 (Epoch AI, March 2025). At current list prices a capable model runs at three dollars per million input tokens and fifteen on output, with a cheaper tier near one dollar (Anthropic, 2026). The unit is cheap and getting cheaper. What grows is consumption, not price. Agentic and reasoning systems generate far more tokens per task than a single chatbot call, so the line rises even as each token falls (Epoch AI, March 2025). A budget set to suppress that line is fighting the wrong variable.

Now the purpose. Capping usage treats tokens as a cost to minimise. Framed correctly, the token line is the cheaper side of a substitution. The expensive side is headcount, and headcount in software is already moving. United States programmer employment fell by about a quarter between 2023 and 2025, while the more design-oriented software developer category barely moved (IEEE Spectrum, December 2025). Entry-level hiring at the fifteen largest tech firms fell twenty-five percent from 2023 to 2024 (IEEE Spectrum, December 2025). The capacity that used to arrive as a junior hire is arriving, in part, as compute attached to a senior one. A plan that holds headcount flat and treats the token line as overhead has described that shift without deciding anything about it.

The bottleneck is not the compute bill. It is that no one owns the trade. Headcount sits with the functional leader and HR. Compute sits with engineering or finance. The substitution between them sits nowhere, so it never gets priced. The result is a plan that funds both sides independently and captures the leverage of neither.

The intervention is to make the trade an explicit planning input, owned at the centre, run for every function. For each function, the plan should state three things: the work to be done, the headcount it would take under the old model, and the token budget that would do part of that work against the people already in seat. That turns an invisible substitution into a visible allocation decision. It also exposes the quality of the trade. Tokens substitute well where work is structured and reviewable, and badly where the constraint is judgment, taste, or relationship. The capital-efficiency case is real where the work suits it: the leading AI-native companies already run at three to four and a half million dollars of revenue per employee, above Microsoft and Meta and near Nvidia (CB Insights, July 2025). That number is the upside of the trade made well, not a law that applies to every function.

What changes when a leadership team runs this is the conversation, not just the spreadsheet. Hiring requests stop being binary. A function asking for three people can be met with one hire and a token budget, or two hires and less compute, and the leader has to argue the mix. Compute stops being an engineering footnote and becomes a lever every function holds. The CFO stops being the person who caps the AI bill and becomes the person who frames the substitution. That is a better job, and a more powerful one.

The risk on the other side is real and worth naming. Compute does not carry institutional knowledge, mentor the next cohort, or hold a customer relationship through a bad quarter. A plan that over-rotates into tokens to flatter near-term cost can hollow out the bench that produces tomorrow’s senior people, the same myopia training leaders already warn about in engineering (IEEE Spectrum, December 2025). The token budget is a lever, not a target. The point is to run the trade deliberately, not to win it by default.

The test for next year’s plan is simple. Open it to any function and look for two numbers next to each other: the headcount it assumes and the token budget it funds. If only one is there, the plan was written for the old cost structure. The discipline is to size them together, decide the mix on purpose, and revisit it each cycle as the price of compute keeps falling and the price of people does not.

Sources

Back to writing