Eighteen months ago, inference cost sat inside a blended cloud bill, unattributed and unexamined. Today it appears in the board pack as a named line, even where nobody can yet reconstruct the number behind it. The AI budget line item has arrived as a formal artefact with an owner, a forecast method and a variance review attached to it. You have watched this line appear in your own board pack, and few people can say what entitled it to that seat. You will need to hold this argument with your CFO on financial mechanics rather than technical instinct.
Why is AI spend suddenly a board-level line item when it was a rounding error eighteen months ago?
AI spend became a board-level line item the moment someone was required to explain variance against it. The underlying reason they could not explain it was that inference turned into a variable, usage-linked cost with no clear owner, no meter and no reliable forecast. On the survey evidence, most boards are now interrogating a number their own organisations cannot reconstruct.
Ramp spent US$1.5 million on tokens during a single week in May 2026. Co-CEO Eric Glyman disclosed the figure in an announcement of their AI spend tool only after the figure had circulated internally for five months. The number was known internally for five months before it was speakable publicly.
Usage shifted quickly from feature garnish to core product workflows. Spend that finance had absorbed at around 2% of run-rate became material once it scaled with workload volume rather than headcount. Yet according to a global CIO survey by Dataiku, 21% of CIOs have full near-real-time visibility into AI costs by team or use case, while 81% lack complete oversight of agents built outside IT. Finance arrived at the number before engineering had installed a meter.
What is tokenomics, and is it a real discipline?
Token economics is the study of how inference cost is generated, allocated and priced across the lifecycle of AI production. The Linux Foundation launched the Tokenomics Foundation in August 2026 to standardise token cost telemetry inside the FOCUS specification. Until those standards land, anyone claiming precision on AI token spend is improvising, so what you can honestly tell your CFO today is a general shape (growing, usage-linked and poorly attributed) alongside a dated plan to instrument it. If the obligation to explain variance is what created the line, the first question is whether the instrument finance already owns can carry it. It cannot.
Why don’t token costs behave like per-seat software licensing?
Seat licensing is a predictable step function of headcount, allocated by directory and forecast directly from the hiring plan. AI token spend is continuous and usage-derived: non-linear in task complexity, invisible to directory groups, and driven by decisions engineers make at the keyboard. The forecast model and the variance explanation both break down.
The mismatch comes down to basic mechanics:
- The unit of cost is a token rather than an employee.
- The cost driver is execution rather than a contract signing.
- The variance story in SaaS (Software as a Service) is joiners and leavers, whereas token variance is driven by prompt context length, model selection and agentic AI loops.
Your hiring plan is the only forecast input you have, and it contains no term for agent steps. As Gartner reported on developer token consumption, consumption-based pricing creates volatile cost structures that standard procurement cannot govern.
Cursor serves as the archetype here. The invoice line is flat and familiar, but heavy frontier-model use or extended context calls fall back to consumption rates. An annual budget with a quarterly reforecast is too slow for spend that can spike inside a single week. As explored in our analysis of why bills climb while token unit prices fall, the per-seat model has no term for how many autonomous steps an agent decides to execute. Because the bill sits in a directory-allocated budget, whoever manages the directory is presumed to own the cost, but that presumption is usually wrong.
What are the three ways token spend lands in the P&L — capex, opex, COGS — and why does the classification change how it gets managed?
AI token spend lands in the profit and loss statement (P&L) as COGS (cost of goods sold) when tokens serve customer-facing features, as opex (operating expenses) when they power internal workflows, and as capex (capital expenditure) when organisations purchase reserved capacity or build self-hosted infrastructure. Classification dictates the owner, the forecast method and the control lever, meaning the exact same token can change executive hands without changing a line of code.
According to analysis by Boston Consulting Group, treating token spend as a single IT bucket makes mismanagement likely. The accounting treatment shapes the operational reality:
- COGS: When inference powers a customer-facing product, it reduces gross margin. Industry margin data from SaaSMag indicates that for every US$1 million in AI product revenue booked in 2026, roughly US$230,000 leaves as inference cost. Broader benchmarks show AI product gross margins compressing from historical 80% levels down to between 50% and 70%. In most organisations, the VP of Engineering typically manages this line like cloud hosting.
- Opex: Internal developer tooling and workplace assistants belong here. Opex is discretionary, tied to departmental budgets and governed by licence permissions. The functional department head or CFO typically oversees this spend.
- Capex: Reserved capacity, multi-year minimum commitments and private model clusters fit capital allocation rules. The platform team and CFO typically manage this through utilisation metrics and depreciation schedules.
The danger is the reclassification hazard. When an internal experiment graduates into a shipped customer feature, the underlying tokens shift from opex to COGS. The owner changes mid-budget. Understanding which AI cost levers a board will accept requires settling this classification first. Routing a workload to a cheaper model now changes gross margin, which makes it a decision finance has a say in. Once classification decides who answers for the line, the next question is what the business does with the cost it now owns.
Should you absorb inference cost inside gross margin or pass it through to customers?
Absorbing inference cost inside gross margin makes sense when AI is a supplementary feature, usage is vendor-triggered, and market pricing norms are flat. Passing costs through suits products where AI is the core deliverable, usage is customer-directed, and consumption varies significantly across accounts. Neither approach is defensible without granular, per-customer cost attribution.
Both strategies are commercial bets with distinct risks. Absorption bets that near-term gross margin compression will be offset by pricing power, retention or falling model prices. Pass-through protects margin integrity but transfers financial volatility to the customer. As monetisation research from Zuora outlines, purely passing through compute cost can anchor willingness-to-pay to raw compute, leaving product value on the table.
In enterprise and government deals governed by fixed-price procurement panels, metered consumption-based pricing often collides with the contracting norms buyers already work under. A model that thrives in self-serve software can stall a panel contract. On the other side of the same bet, absorbing unmetered usage creates exposure: a five- to ten-fold swing in inference between your heaviest and lightest accounts will not show up as an incident, but it shows up inside an averaged gross margin line.
Many businesses adopt middle paths, such as capped tiers or hybrid subscription-plus-overage models. Understanding what a normal token bill looks like is the prerequisite for calculating true cost per outcome.
Ultimately, finance owns the budget line, engineering owns the architectural levers, product owns the monetisation model, and the board owns the variance conversation. Recognising why finance now owns an AI line allows you to turn a contested board topic into a clear operational baseline. The real question is not who owns the line today, but who notices when it shifts from opex to COGS mid-year.
Frequently Asked Questions
Is tokenomics a real discipline or just a vendor buzzword?
Token economics is a real discipline wearing a vendor’s coat. It studies how inference cost is generated, allocated and priced, and every question it asks is one a board is already asking. The label has been cheapened by observability vendors selling dashboards, so treat the term with suspicion and the underlying questions with respect. The Tokenomics Foundation’s work on telemetry standards is where the discipline is currently being made concrete.
My CFO just asked me to explain our AI spend. What do I actually tell her?
Give her a shape and a date, not a false precision. Tell her the line is growing, usage-linked and currently attributed by inference rather than by evidence, then commit to a specific measurement plan with a delivery date. Anyone presenting an exact per-team number this quarter is improvising, and a CFO usually finds that out. The credibility you gain from conceding the gap is worth more than the number you invent.
Where do I follow the standardisation work on AI cost measurement?
Follow the Tokenomics Foundation, launched by the Linux Foundation in August 2026, and its work on token cost telemetry inside the FOCUS specification. That is where the definitions for allocation and attribution are being argued, and where your finance team will eventually get a shared vocabulary. Until it lands, your internal cost model is a local dialect, so keep it documented and portable.
Does a cost have to be large before it earns a board line item?
No. Size gets a cost noticed; unexplained variance gets it a seat. A line item exists when someone must account for movement against a forecast, which is why a mid-sized, volatile cost can outrank a larger predictable one on the board agenda. Plenty of organisations carry large cloud bills that are never interrogated line by line, and much smaller token bills that are.
Why did our AI bill climb when headcount didn’t move and per-token prices fell?
Because token spend tracks work, not people. Three things move it without a single hire: agents taking more steps per task, prompts carrying longer context, and workloads shifting to larger models for quality. The agentic multiplier is the usual culprit, since a workflow that used one model call can quietly become twenty. Falling unit prices are real, but they set the rate, not the volume.
Is AI token spend part of our cloud bill, or should it be reported separately?
Report it separately, even while it is still invoiced inside a cloud commitment. The invoice groups by vendor and contract; the board needs it grouped by who consumes it and what it serves. Separating the line forces the attribution work a blended cloud bill lets you avoid, and it stops inference growth hiding inside a stable-looking infrastructure number.
Can we capitalise token spend as software development, or is it always opex?
Mostly opex or COGS, and only rarely capital. Under AASB 138 and IAS 38, capitalisation depends on a recognisable intangible asset and control of future economic benefits, which fits self-hosted model infrastructure and committed capacity far better than it fits metered consumption. Tokens consumed running a product are a cost of delivery. Get accounting advice early, because a treatment that changes at year end reads as a restatement.
What happens to ownership when a feature moves from an internal tool to a shipped product?
The token moves from opex to COGS, and with it the owner, the forecast method and the control lever, usually without anyone announcing it. An internal assistant that becomes a customer-facing capability changes the line it lands in and the executive who answers for it. Ask at every product gate whether the workload is now sold, and reclassify deliberately rather than discovering it at the next board pack.
Is it true that routing to a cheaper model fixes the margin problem?
Rarely on its own. Routing a workload to a cheaper tier reduces unit cost, but it is a decision with a quality consequence and a finance-visible outcome, so it belongs in a margin conversation rather than a sprint. If the workload is genuinely commodity work, the saving is real. If the cheaper tier increases retries, escalations or agent loops, the bill can rise while the price per token falls.
What is cost per outcome, and why do boards ask for it instead of cost per token?
Cost per outcome divides token spend by the unit of value actually delivered, such as a resolved ticket, a completed claim or a closed sale. Boards ask for it because cost per token says nothing about whether the work was worth doing, and it is the only figure that lets them compare an AI investment against the labour it replaced. Cost per token is a supplier metric; cost per outcome is a management one.
How much month-to-month variance in token spend is normal?
There is no benchmark worth quoting yet, which is the honest answer. Variance depends on workload mix, context length and how many agent loops your workflows run, so a 30 per cent swing can be perfectly ordinary in one organisation and a signal of leakage in another. Establish your own baseline first, then treat deviation from it as the trigger for questions. Published averages are too coarse to defend in a board pack.
