The Margin Inversion: Why AI-Native Software Spends 40% of Revenue on Inference

A four-person startup ran a $113,000 monthly bill from a single AI provider. Uber’s CTO burned the company’s entire 2026 AI budget in four months. Neither of these companies made a pricing mistake. They ran into a structural fact that most finance teams have yet to price into their models: the per-token cost of AI is collapsing, and total AI bills are climbing anyway.

Epoch AI tracked the price of a million tokens from roughly $60 in 2021 to about $0.60 in 2024. A 100x reduction, halving on average every two months. Any CFO reading that number would reasonably expect AI to become a rounding error on the P&L. The opposite is happening. The volume of tokens consumed per unit of work is rising faster than the price is falling, and the result lands on the income statement as a new, large, growing cost of goods sold.

The number the board should be tracking

Cost per seat is the metric most finance teams inherited from the SaaS era. You license software per user, you forecast headcount, you multiply. It worked because the marginal cost of one more user logging in was close to zero.

Agentic AI breaks that arithmetic. When an employee runs an agent instead of typing a prompt, the cost of that action is no longer fixed. It scales with how hard the agent works: how many times it loops, how many tools it calls, how much context it reloads on each pass. A single-shot reply might cost a fraction of a cent. An agentic loop that plans, retrieves, executes, checks its own work, and retries consumes 60 to 140 times the tokens of that single reply. Goldman Sachs forecasts token consumption growing 24x by 2030.

The governable unit is cost per completed task. Not cost per seat, not cost per API call, not monthly spend in aggregate. What did it cost to close one ticket, draft one contract, reconcile one account, resolve one customer case? That number is the one a board can hold a business accountable to, because it ties spend directly to output. Everything else hides the variance that is about to define your margins.

Why the old cost structure does not transfer

Legacy SaaS ran inference-free. Its cost of goods sold was hosting, storage, bandwidth, and support, and it landed at roughly 10 to 20 percent of revenue. That is why software carried 70 to 80 percent gross margins and why investors paid the multiples they did. The economics of the entire category rested on the marginal cost of serving one more customer being trivial.

AI-native software lacks that property. When inference is the product, every unit of output carries a compute cost that persists at scale. Current AI-native companies are running inference at 40 to 50 percent of cost of goods sold. That is a different gross margin, a different valuation logic, and a different question at every board meeting. Those economics belong to a different kind of business, one that happens to wear the same category label as SaaS.

For a finance leader, the implication is direct. If you are modeling an AI product line with SaaS-era gross margins, your model is wrong by 30 to 40 points. If you are underwriting an acquisition or an internal build on the assumption that unit costs fall to zero as you scale, you are underwriting the wrong company.

The budgets are already being redrawn

The organizations closest to the frontier have stopped treating AI spend as an IT line item. JPMorgan sets per-analyst token budgets ranging from $10,000 to $100,000 and above. That functions as a compute allowance attached to a person, governed like a trading limit, because the downside of an ungoverned agent loop is measured in dollars per hour of runtime.

This is what the Uber budget story actually signals. The CTO did not overspend through negligence. The budget was built on last year’s mental model, where AI cost scaled with users and use cases, not with the depth of agent reasoning. Four months in, agentic workloads consumed what a year of the old model predicted. The forecast was conservative. It was based on the wrong unit.

Per-token deflation makes this worse, not better, and the mechanism is worth stating plainly. Cheaper tokens make more use cases economically viable. More viable use cases mean more agents deployed. More agents mean more loops, more retries, more autonomous reasoning burning tokens without a human in the loop to notice. The savings on the price of a token get spent, many times over, on the volume of tokens consumed. We covered the compounding version of this dynamic in the economics of AI governance.

What finance needs before the next budget cycle

Three moves, in order of urgency.

First, instrument cost per completed task for every AI workload in production. Not aggregate spend, not spend per team. The unit cost of a finished output, tracked over time. If a workflow’s cost per completed task is rising while its output quality is flat, you have found margin erosion before it reaches the quarterly numbers. This is the metric that makes AI spend a managed line rather than a surprise. The mechanics of attributing agent cost to specific work are covered in per-use agent cost oversight.

Second, set per-workload and per-person budgets with hard ceilings, the way JPMorgan does. An agent without a spend ceiling is a financial exposure with no stop-loss. The ceiling can be loose. It needs to exist, so a runaway loop trips a limit instead of a monthly invoice.

Third, re-underwrite every AI product line and AI-heavy acquisition at 40 to 50 percent inference-of-COGS, not SaaS-era margins. If the business still works at that cost structure, you have a real business. If it only works when you assume inference trends to zero, you are pricing a hope.

Do this now

Ask your finance team one question at the next review: what does it cost us to complete one unit of work with AI, and is that number going up or down? If nobody can answer, the metric is not instrumented, and the exposure is unmanaged. The token price is going to keep falling. Your bill is going to keep climbing. The only defense is knowing, per completed task, what you are actually paying, and governing to that number before the board asks why gross margin moved.

This analysis synthesizes Lex’s “AI tokens are cheaper but AI bills” (Lex on Substack, July 2026), whose reported figures on provider bills and per-analyst budgets we present as cited rather than independently verified.

Victorino Group helps finance and operations leaders instrument cost per completed task and govern AI spend as a managed line, not a quarterly surprise. Let’s talk.