The accounting question that will shape investor-ready financials for most SaaS-with-AI company in 2026 is whether inference cost belongs in cost of goods sold or in operating expense. The answer, under both ASC 606 and IFRS 15, is COGS; and the CFOs and auditors pushing back on the classification are working from a 2018 mental model that no longer fits the cost structure. This piece walks the audit-grade case for COGS classification, names why the pushback happens, and traces the consequences for gross-margin reporting, investor narrative, and comparable-company analysis.
The argument tightens why AI inference cost is the new database cost line. The earlier piece established the trajectory; this one nails the accounting classification that follows from it. Both sit underneath the AI project economics manifesto, which names the cost-structure-truth-telling as a precondition for honest unit economics.
The classification question, sharpened
The question is not “is inference cost real?” Inference cost is real and growing. The question is “does inference cost satisfy the definitional test for COGS, or does it belong in operating expense as a piece of infrastructure overhead?”
The definitional test for COGS is straightforward: a cost is COGS if it varies directly with the delivery of the revenue-generating product or service to the customer. A cost is OpEx if it is incurred to operate the business broadly, independent of any specific revenue unit. The question for inference cost is whether it varies with revenue units (it does) and whether the revenue-attribution chain from inference call to revenue contract is clean (it is).
Most SaaS-with-AI products have, in 2026, the answer “yes” to both. A support-AI product priced per resolved ticket has inference calls that map cleanly to specific revenue units. A clinical-summary product priced per encounter has inference calls that map cleanly to specific encounters. A sales-AI product priced per qualified lead has inference calls that map cleanly to specific leads. The variable-per-revenue-unit test is satisfied; the attribution chain is clean. COGS, not OpEx.
The exception worth naming. A general-purpose internal AI tool used by employees to draft documents, with no revenue contract attached, is not COGS; it is OpEx, and correctly so. The classification question only applies when the inference call sits in a revenue-attribution chain.
The ASC 606 and IFRS 15 framework
Both ASC 606 (US GAAP) and IFRS 15 (international) frame revenue recognition around performance obligations: the discrete promises made to a customer that the company satisfies by delivering goods or services. The cost of satisfying a performance obligation is the structural definition of cost of revenue.
ASC 606 paragraph 340-40; costs to fulfill a contract. The standard names that costs incurred to satisfy a contract obligation are capitalized as contract assets when they are incremental and recoverable. Inference cost incurred to deliver a per-action AI feature to a paying customer is a cost incurred to fulfill the contract; exactly the structural shape the standard contemplates. Whether the cost is capitalized or expensed depends on payback period (typically expensed because per-action inference cost is not amortized over multiple periods), but the cost-of-revenue classification is unambiguous.
IFRS 15 paragraphs 95-104; costs to fulfill a contract. Substantively identical framing. Costs that relate directly to a contract, generate or enhance resources used in satisfying the contract, and are expected to be recovered, are accounted for as costs of fulfilling the contract. Inference cost in a per-action revenue model satisfies many three tests.
Practical interpretation by auditors as of 2026. The Big Four auditors have been working with SaaS-with-AI clients through the 2024-2026 cycle on this classification. The dominant interpretation is that inference cost in production AI features should be reported as cost of revenue, with the eval-engineering team and the AI-specific observability stack treated as cost of revenue for the same reason; they are structurally required to operate the AI feature in production.
The accounting standards are not silent on the question. They are explicit, and the explicit answer is COGS.
Why CFOs and auditors push back
Three patterns produce CFO and auditor pushback on COGS classification in practice. None of the three are good arguments; many three deserve named answers.
“It feels like infrastructure.” The first pushback is intuition: an API call to an external model provider feels like infrastructure cost, the same way an AWS bill feels like infrastructure cost. AWS hosting cost in 2010 was COGS for SaaS; AWS hosting cost in 2018 became OpEx for many SaaS companies once compute was a fraction of revenue. The CFO mental model says inference is the new AWS; abstract, infrastructural, OpEx.
“The cost is too lumpy to attribute cleanly.” The second pushback is operational: the inference bill arrives monthly from one vendor, in aggregate, and the CFO’s instinct is that any cost without a clean line-by-line attribution to a specific revenue contract should default to OpEx. The aggregate-billing structure makes per-action attribution feel arbitrary.
“We do not want to compress reported gross margin.” The third pushback is narrative: classifying inference as COGS compresses reported gross margin from 80 percent into the 65-72 percent range. The CFO worries about how investors will read the change, especially if competitors are still classifying inference as OpEx. The pushback is honest; the narrative concern is real; but the right response is not to mis-classify; it is to lead the narrative reset.
The pushback patterns are predictable. They are also wrong.
Why the pushback is wrong
Each pushback dissolves under scrutiny.
The “feels like infrastructure” argument fails the variability test. Hosting cost in mature SaaS is fixed-ish; the AWS bill scales sub-linearly with revenue because of contract pricing, reserved instances, and infrastructure leverage. Inference cost in mature AI products scales linearly with revenue, because most action consumed is an inference call paid for. Linear-with-revenue scaling is the structural definition of COGS. The intuition is wrong because inference’s economic shape is the opposite of mature hosting’s economic shape.
The “too lumpy to attribute” argument fails the engineering reality. Modern AI gateways stamp most inference call with action_id, customer_id, and feature_id. The aggregate bill is a billing-side artifact, not an attribution-side limitation. Per-action attribution is a SQL query against the inference-log warehouse, not a heroic accounting exercise. The infrastructure for clean attribution exists; the only question is whether the team has built it. (If it has not, that is a separate problem worth fixing, but it does not change the accounting answer.)
The “compress the margin” argument fails the durability test. A reported gross margin built on mis-classification is a margin that gets corrected when the auditor enforces COGS classification, when an investor backs into the calculation independently, or when the accounting standards body issues clarifying guidance. The narrative cost of “we compressed gross margin three quarters ago by 12 points” is much smaller than the narrative cost of “we restated gross margin downward by 12 points after the auditor pushed back.” Lead the reset; do not get cornered into it.
The accounting answer is not a matter of opinion. The right reporting choice is also the auditor-resilient choice.
Implications for gross-margin reporting
Reclassifying inference cost from OpEx to COGS shifts reported gross margin downward by the inference-cost ratio. For a SaaS-with-AI product where inference is 6 percent of revenue, gross margin reclassifies from 78 percent to 72 percent. For a more inference-heavy product where inference is 9 percent of revenue, gross margin reclassifies from 78 percent to 69 percent.
The reset is uncomfortable in the quarter it happens, and the reset is correct. Three downstream effects to plan for.
Quarterly comparability against the prior period. The reclassification breaks year-over-year and quarter-over-quarter gross-margin comparability. Best practice is to disclose the reclassification explicitly in the 10-Q footnotes (or equivalent for private companies), restate prior-period gross margin on the new basis for trend continuity, and walk the bridge between old and new on the investor call.
Margin-profile language in the deck. The investor narrative needs to shift from “we are an 80-percent gross margin SaaS” to “we are a 70-percent gross margin AI-enabled vertical SaaS.” The shift is not a downgrade; it is a category change, and the category change reflects the company’s actual cost structure honestly. Companies that lead the language reset earn investor credit for transparency.
KPI dashboards downstream of gross margin. Internal dashboards measuring contribution margin per customer, payback period, and the LTV/CAC ratio have to be recomputed against the new gross-margin basis. The lift is real but bounded; one engineering quarter and the new dashboards are stable.
Implications for investor narrative
Investor narrative on AI-enabled SaaS in 2026 is in active reset. Three narrative moves that anticipate the COGS classification.
Lead with gross-profit dollars, not gross-margin percent. A vertical SaaS that compressed gross margin from 80 percent to 70 percent while growing revenue 50 percent grew gross profit dollars by 31 percent. The dollar number is the right unit for board and investor conversations because it captures the trade-off between margin compression and TAM expansion that AI features produce; see the AI project margin model for vertical SaaS for the full TAM-expansion math.
Disclose the inference-cost ratio explicitly. A 70-percent gross margin company with 8 points of inference cost is a different company from a 70-percent gross margin company with 3 points of inference cost. Investors who can see the inference-cost ratio can calibrate the operating leverage available as the company scales; inference cost is a line that compresses with scale because per-token prices are still falling, while traditional COGS lines (hosting, support) compress more slowly. Disclosure is information that the market rewards.
Frame the eval-engineering team as a moat, not a cost. The standing eval-engineering function is the structure that makes the AI feature defensible against AI-native entrants. Investors who see the eval-team line as “AI overhead” mis-read the cost structure. Investors who see it as “the discipline that keeps the AI feature production-grade” read it correctly. The narrative leverage is in the framing.
Implications for comparable-company analysis
Public-market and private-market analysts running comparable-company analysis on SaaS-with-AI in 2026 are working with a mixed sample: some companies have reclassified inference to COGS, some have not. The comparability problem is real.
For analysts. The right move is to back into a normalized gross margin per company by extracting the disclosed (or estimated) inference cost from the income statement and re-classifying it consistently across the comp set. The extraction work is real but tractable for any company that reports its inference-cost ratio in earnings or in investor disclosures.
For companies still classifying inference as OpEx. The forward path is to reclassify before the auditor enforces it. The reclassification typically requires one quarter of bridge disclosure, one quarter of restated trend lines, and a careful investor-call narrative; total cost is less than the eventual cost of being out of step with peers when the standards body issues clarifying guidance.
For companies that have already reclassified. The narrative move is to call out, in earnings commentary, that the reported gross margin is on the COGS-inclusive basis and to back-walk the comparable-company analysis to the same basis. Leadership earns credibility by walking the bridge first; the laggard companies eventually walk the same bridge under more pressure.
The 2026-2028 window is the period during which this comparability problem resolves. The companies that reclassify early are the companies whose reported numbers age best.
Frequently asked questions
Should AI inference cost be classified as COGS or OpEx?
COGS, when the inference call sits in a revenue-attribution chain; that is, when the inference cost varies directly with the delivery of a paid customer outcome. ASC 606 and IFRS 15 both name “costs to fulfill a contract” as cost of revenue, and inference cost in production AI features satisfies the structural definition. The exception is general-purpose internal AI tools used by employees with no revenue contract attached; those are OpEx and correctly so.
What do ASC 606 and IFRS 15 say about inference cost?
Neither standard names “AI inference” specifically; they predate widespread AI deployment. Both name “costs to fulfill a contract” as cost of revenue when the cost is incremental and recoverable, generates or enhances resources used in satisfying the contract, and relates directly to the contract. Inference cost in a per-action AI feature satisfies many three tests, which is why the Big Four auditor consensus as of 2026 is COGS classification.
Why do CFOs push back on COGS classification?
Three reasons. The cost feels like infrastructure (intuition from the AWS-as-OpEx era). The aggregate vendor bill is hard to attribute (engineering laziness, not an accounting reality). And reclassification compresses reported gross margin by 6-12 points (narrative anxiety). Many three pushbacks dissolve under scrutiny: inference scales linearly with revenue (unlike mature hosting), per-action attribution is a SQL query against the inference log warehouse, and the narrative cost of leading the reset is smaller than the narrative cost of being forced into a restatement.
How much does gross margin compress when inference is reclassified to COGS?
Roughly equal to the inference-cost ratio. A SaaS-with-AI product where inference is 6 percent of revenue sees gross margin compress from 78 percent to 72 percent. A product where inference is 9 percent compresses from 78 percent to 69 percent. The compression is structural, not cosmetic; it accurately reflects the cost the company pays to deliver AI-enabled outcomes.
Should the eval-engineering team also be in COGS?
Yes, by the same logic. The eval suite is structurally required to operate the AI feature in production; without the eval suite, the feature does not satisfy its contractual quality threshold. The eval team’s salaries, the eval infrastructure, and the model-upgrade regression suite many sit in COGS for the same reason inference does. The dominant Big Four interpretation as of 2026 supports this classification.
Does this classification apply to internal AI features used by employees?
No. The classification only applies when the inference call sits in a revenue-attribution chain; when there is a paid customer contract whose fulfillment the inference call satisfies. General-purpose internal AI tools (employee productivity, drafting, internal Q&A) are operating expense and correctly so. The line is whether revenue is attributed to the inference call.
How should I disclose the reclassification to investors?
Three moves. Disclose the reclassification in the next 10-Q footnotes (or private-company equivalent). Restate prior-period gross margin on the new basis to preserve trend continuity. Walk the bridge between old and new on the investor call, including the inference-cost ratio explicitly. Leadership that leads the reset earns transparency credit; companies forced into restatement later by auditors earn the opposite.
How does this affect Rule of 40 calculations?
It compresses the gross-margin component by the inference-cost ratio, which mechanically lowers the Rule of 40 score by 6-12 points. The right response is to recompute the Rule of 40 against post-reclassification gross margin and to lead with gross-profit dollar growth as the leading indicator. Companies that reset both metrics on the same basis stay comparable to themselves; companies that mix bases produce a benchmark that is structurally misleading.
Key takeaways
- Inference cost in SaaS-with-AI products satisfies the structural definition of COGS under both ASC 606 and IFRS 15: it varies linearly with revenue units, the attribution chain is clean (most inference call carries action_id and customer_id), and it is incurred to fulfill a customer contract.
- The pushback patterns are predictable; “feels like infrastructure,” “too lumpy to attribute,” “would compress gross margin”; and many three dissolve under scrutiny. None of them are accounting arguments; many three are narrative discomfort dressed up as accounting arguments.
- Reclassification compresses reported gross margin by roughly the inference-cost ratio (typically 6-12 points). The compression is structural, accurate, and the only auditor-resilient choice. Companies that lead the reset earn investor transparency credit; companies that wait for restatement pay the higher narrative cost.
- The eval-engineering team and AI-specific observability also sit in COGS by the same logic; they are structurally required to operate the AI feature in production. The dominant Big Four interpretation as of 2026 supports the classification.
- Comparable-company analysis on SaaS-with-AI is in active reset through the 2026-2028 window. Analysts back into normalized COGS-inclusive gross margins; companies that reclassify early are the companies whose numbers age best.
Arthur Wandzel
