AI Unit Economics and Gross Margins

What exactly is being measured?#

The accounting definition is fixed:

gross margin = (revenue - cost of revenue) / revenue

For operating decisions, two managerial measures do more work:

contribution per unit  = unit revenue - unit-variable delivery cost

customer contribution  = customer revenue
                       - directly attributable delivery cost
                       - customer-specific support or success cost

Gross margin follows your accounting policy and is auditable. Contribution is a management number: it is not GAAP gross profit, it must be labelled as such, and it must reconcile back to the income statement. Mixing the two — comparing a customer contribution margin against a company gross margin, for instance — produces a comparison with different numerators and different denominators, which is worse than no comparison at all.

The cost boundary is a judgement call, and public filings show how much room there is inside it. Palantir includes third-party cloud hosting, hardware, subcontractors, field-service representatives, and the payroll and stock-based compensation of delivery personnel inside cost of revenue (Palantir Q2 2025 Form 10-Q). On $4.475 billion of 2025 revenue that produced $789.2 million of cost of revenue and an 82% GAAP gross margin, of which $64.6 million was stock-based compensation (Palantir Q4 and FY2025 earnings release). Those are examples of one company's classification, not a rule. Document your own policy with an accountant and keep it stable across periods.

Why does this matter to founders?#

Pricing can conceal a cost mismatch. Seat pricing makes revenue predictable and usage unbounded. Outcome pricing aligns with value but leaves you paying for attempts the customer never accepts. Usage pricing passes variability through and is harder for a buyer to budget.

Growth does not automatically improve economics. Large customers bring longer contexts, premium model routes, dedicated capacity, data licences, implementation, and human assurance. Expansion that lowers blended margin can still be the right trade — but only when it is visible and argued for, not discovered a year later.

The chosen unit writes your roadmap. Cost per API call rewards making calls smaller. Cost per successful task rewards better prompts, routing, caching, and fewer retries. Cost per retained customer additionally asks whether the output was good enough to keep the account.

Investors will interrogate the bridge. "Margins will look like SaaS" is not a plan. A credible bridge names the baseline, each lever, its measured effect, its timing, and the quality guardrail attached to it.

How do you build the margin stack?#

1. Choose the economic unit#

Use the narrowest unit tied to customer value. Define what counts as success, and decide how retries are treated, before you calculate any cost. A unit that silently excludes failed attempts is not measuring delivery.

2. Build cost of revenue from the workload#

Shared cost needs an explicit driver, and allocated cost should be labelled separately from observed cost so a reader can see how much of the answer is measurement and how much is assumption. Cloud providers support this directly through cost allocation tags, which must be activated in the billing console before they appear in reports and can take up to 24 hours to become active (AWS Billing, cost allocation tags).

3. Report three views, not one#

task contribution margin     = (task revenue - task-variable cost) / task revenue
gross margin                 = (revenue - cost of revenue) / revenue
customer contribution margin = (customer revenue
                                - customer-attributable delivery and service cost)
                               / customer revenue

They answer different questions: variable product efficiency, accounting delivery economics, and customer quality. Compare each only against its own prior period or its own peer set.

4. Build a driver bridge with counter-metrics#

Model margin as a function of price and packaging, model and tool rates, input/output size, routing mix, cache hit rate, attempts per success, capacity utilisation, human escalation rate, fixed delivery cost absorbed across volume, and revenue mix. Every proposed lever gets a counter-metric attached — task success, quality score, latency, retention, or loss rate. A lever without a counter-metric is a way of moving cost somewhere the dashboard cannot see.

Worked example: where did the 7.9 points come from?#

An AI workflow company completes one million successful tasks per month. It earns $60,000 of usage revenue ($0.06 per successful task) and $20,000 of platform fees.

gross profit  = $80,000 - $40,000 = $40,000
gross margin  = $40,000 / $80,000 = 50.0%
variable cost = $31,000, or $0.031 per successful task

After routing, caching, and workflow changes, volume reaches 1.6 million successful tasks at the same $0.06 price. Variable cost falls to $0.023 per success; the delivery platform steps up to $12,000.

gross profit = $116,000 - $48,800 = $67,200
gross margin = $67,200 / $116,000 = 57.9%

Now decompose the 7.9 points, because the two causes are worth very different amounts:

Roughly seven of the eight points are durable engineering and under one point is volume absorbing a fixed platform. That distinction is the whole substance of a margin bridge: absorption reverses the moment growth slows, and it can be manufactured simply by growing.

Caveats this model hides. The $0.023 variable cost is asserted, not derived — in a real bridge it decomposes further into route mix, cache hit rate, and attempts per success, each with its own evidence. Revenue per task also fell from $0.080 to $0.0725, because the platform fee did not grow with volume; a fixed fee that stops scaling is a slow drag on the same percentage the engineering work is improving. And none of this is valid unless success rate, quality, and retention held. Show those three next to the margin, or the bridge is not checkable.

Now a single enterprise customer paying $10,000 a month, consuming $2,600 of metered delivery cost and $1,900 of attributable review and support:

customer contribution margin = ($10,000 - $4,500) / $10,000 = 55.0%

Resist comparing that 55% against the company's 57.9% gross margin — they are different measures. Compare it against the customer-contribution distribution across your other accounts. Then the options are workflow redesign, a service tier, usage limits, a price change, or deliberate acceptance because the account earns its keep in product feedback or distribution. The number informs the decision; it does not make it.

Related concepts#

• Token Economics and Metering — the usage ledger that makes cost attributable to a task or customer.
• GPU and Compute Economics — capacity, utilisation, and cost per accepted output.
• Inference vs. Training Costs — separating recurring delivery from model development.
• AI-as-a-Service — packaging and service levels around the delivery stack.
• Gross Margin — the accounting measure this page decomposes.
• Contribution Margin — the managerial companion to gross margin.
• Usage-Based Pricing — aligning variable revenue with variable cost.
• Pricing Metric / Value Metric — choosing a unit the customer can control.

Sources#

1. Snowflake Inc., Snowflake Reports Financial Results for the Fourth Quarter and Full-Year of Fiscal 2026, 25 February 2026. GAAP product gross margin of 72% and non-GAAP product gross margin of 76% for fiscal 2026; GAAP professional-services-and-other gross margin of −31% (FY2026) and −36% (FY2025); GAAP total gross profit of $3.146 billion at 67% and non-GAAP gross profit of $3.388 billion at 72%; stock-based-compensation and acquired-intangible adjustments of 3% and 1% of product revenue.
2. Palantir Technologies Inc., Q4 and FY2025 earnings release (Form 8-K, Exhibit 99.1), 2 February 2026. FY2025 revenue of $4,475,446 thousand, cost of revenue of $789,177 thousand, gross profit of $3,686,269 thousand (82% GAAP gross margin), and $64,555 thousand of stock-based compensation within cost of revenue.
3. Palantir Technologies Inc., Form 10-Q for the quarter ended 30 June 2025, August 2025. Cost of revenue composition: delivery payroll and stock-based compensation, subcontractor expenses, field-service representatives, third-party cloud hosting, hardware, travel, allocated overhead, and other direct costs.
4. C3.ai, Inc., C3 AI Announces Fiscal Fourth Quarter and Full Fiscal Year 2026 Results, 3 June 2026. FY2026 total revenue of $250.3 million with GAAP gross profit of $77.4 million (31%) and non-GAAP gross profit of $116.2 million (46%); subscription revenue of $227.1 million against cost of subscription revenue of $166.291 million; prior-year gross profit of $235.863 million on $389.056 million of revenue; cost increase attributed primarily to payroll and contractor costs.
5. Amazon Web Services, Organizing and tracking costs using AWS cost allocation tags, accessed 15 August 2026. AWS-generated and user-defined cost allocation tags, activation in the Billing and Cost Management console, and the delay before activated tag keys appear in cost reports.

Note: This page is educational and is not accounting, tax, or investment advice. Where cost of revenue is drawn, and how contribution measures are presented alongside GAAP figures, are decisions to make with a qualified accountant. Company figures are as reported on the dates cited and are not restated here; verify against the current filing before using any of them in a board pack or a fundraise.