The Rule of 40; growth rate plus operating margin must clear 40; is the SaaS industry’s single most-quoted health metric, and it is mispriced for AI-native software. The metric assumes COGS is dominated by infrastructure that scales sublinearly with revenue. AI-native SaaS has inference COGS that scales roughly linearly with active usage, plus eval-engineering opex that scales with model lifecycle rather than with revenue. Applying the same 40 bar to both produces inflated comfort about AI-native economics and unfair penalty against the trajectory’s compounding tail.
The recalibration argued here: Rule of 35 for AI-native SaaS, with three named offsets; a 5-point credit for compounding capability, a 3-point debit for eval-engineering rigidity, and a contextual 2 to 5 point band for inference COGS shape. An AI-native SaaS clearing Rule of 35 is performing at parity with a legacy SaaS clearing Rule of 40. One clearing 40 outright is genuinely outperforming.
This is a spoke under the AI project economics manifesto, which argues that legacy SaaS economic templates mis-price AI work. Rule of 40 is the cleanest example of that mispricing because the metric is so structurally embedded in board-level conversation that recalibrating it requires explicit argument.
Why classic Rule of 40 mis-prices AI-native SaaS
The Rule of 40, as it took shape in the 2015 to 2018 SaaS investing literature, encodes an implicit cost structure: COGS dominated by hosting and database infrastructure that scales sublinearly with revenue, with the residual operating expense going to S&M, R&D, and G&A. Gross margin in this model sits comfortably above 70 percent, and the Rule of 40 lever is a tradeoff between growth rate and the OpEx side of the income statement.
Three structural assumptions break for AI-native SaaS.
Assumption one: COGS scales sublinearly with revenue. False for AI-native. Inference cost is roughly linear with usage. Each query costs tokens; tokens cost money; that cost lands in COGS. A growing AI-native SaaS does not see its COGS shrink as a percent of revenue the way a legacy SaaS does; it sees the line stay flat or grow if usage intensity per dollar rises.
Assumption two: technology cost is a fixed-ratio category. False. Inference cost moves with model selection, prompt design, retrieval architecture, and caching strategy. It is engineered, not purchased. A team that does not actively manage it watches it climb; a team that does manage it captures the year-over-year decay detailed in the AI project compounding return.
Assumption three: software has reached operational steady-state at GA. False. AI software at GA is mid-lifecycle. Eval engineering, model upgrades, and threshold re-locking continue at meaningful expense for years post-GA. The OpEx line that funds this work does not scale down with margin pressure the way headcount in a stable engineering org does.
The Rule of 40 metric does not see any of these. Applied to an AI-native SaaS, it produces a number that mixes the linear-in-usage cost shape of inference with the sublinear-in-usage cost shape of legacy infrastructure. The mix obscures the underlying trajectory. A board that scores AI-native SaaS by classic Rule of 40 is scoring it on the wrong instrument.
The structural decomposition
Three line items distinguish AI-native economics from legacy SaaS economics. Each enters the recalibration.
| Line item | Classic SaaS | AI-native SaaS | Driver |
|---|---|---|---|
| Inference COGS | Not present | 8 to 22 percent of revenue | Linear in active usage |
| Eval-engineering opex | Not present | 4 to 9 percent of revenue | Driven by model lifecycle |
| Observability COGS | 1 to 3 percent of revenue | 3 to 6 percent of revenue | 15-25% of inference spend |
Inference COGS. Eight to twenty-two percent of revenue at 2026 model pricing, depending on workload shape. Chat-shaped products with controllable token counts run 8 to 12 percent. Agent and reasoning-heavy products run 14 to 22 percent. Long-context document analysis and multimodal products can spike past 25 percent on heavy users without per-seat caps. Why inference belongs in COGS rather than OpEx is argued in why AI inference belongs in COGS not OpEx.
Eval-engineering opex. Four to nine percent of revenue at steady state. Decomposes into eval-set maintenance, regression triage, model-upgrade re-evaluation, and threshold re-locking. It is OpEx, not COGS, but it does not scale down with margin pressure the way other OpEx categories do. A finance team that tries to cut it during a margin tightening will discover that cutting it produces capability stagnation that costs more than it saves. The 3-point debit in the Rule of 35 reflects this structural rigidity.
Observability COGS. Sized at 15 to 25 percent of inference spend, observability adds 3 to 6 percent of revenue to the COGS line for AI-native SaaS, against 1 to 3 percent for legacy. The increase reflects the fact that AI products require observability to know whether they are working, not just to know whether they are up.
These three items together account for the gross margin compression that distinguishes AI-native from legacy SaaS. A product that would be a 78 percent gross margin SaaS company on legacy economics becomes a 60 to 68 percent gross margin SaaS company on AI-native economics, with the lost margin distributed across the three lines above.
The proposed Rule of 35
Rule of 35 = growth rate + operating margin >= 35, applied to AI-native SaaS workload segments. Three offsets sit alongside.
Offset one: +5 compounding capability credit. AI-native SaaS accumulates capability through eval libraries, prompt registries, agent skills, and observability stack; assets that compound across product lines and reduce time-to-value on subsequent capability launches. Legacy SaaS has no equivalent at the same intensity. The 5-point credit recognizes that an AI-native SaaS clearing Rule of 35 carries strategic option value beyond its current margin profile.
Offset two: −3 eval-rigidity debit. Eval-engineering opex does not flex down. The 3-point debit reflects that an AI-native SaaS reading a particular operating margin has less near-term flexibility than a legacy SaaS reading the same margin. The debit is structural, not punitive.
Offset three: ±2 to 5 point inference-COGS band. A chat-shaped product running at 9 percent inference COGS sits in the favorable end of the band; an agent-heavy product running at 19 percent sits in the unfavorable end. The band moves the bar 2 to 5 points based on workload shape.
The summary table:
| Metric | Bar | Notes |
|---|---|---|
| Classic Rule of 40 (legacy SaaS) | 40 | Original |
| Adjusted Rule of 35 (AI-native, base) | 35 | Inference COGS structural |
| Compounding capability credit | +5 | Asset accumulation |
| Eval-rigidity debit | −3 | OpEx inflexibility |
| Net adjusted bar | 37 | For comparison to legacy 40 |
| Inference-COGS band | ±2 to 5 | Workload shape |
The net adjusted bar of 37 against legacy 40 captures the position cleanly: an AI-native SaaS at growth-plus-margin 37 is approximately as healthy as a legacy SaaS at 40, accounting for the cost-shape difference and the compounding asset accumulation.
Worked example: a Series B AI-native SaaS
Take an AI-native SaaS with the following profile:
- Growth rate: 45 percent
- Inference COGS: 18 percent of revenue (agent-heavy product)
- Eval-engineering opex: 4 percent
- Observability COGS: 4 percent
- S&M: 22 percent
- G&A: 10 percent
- R&D: 10 percent
- Other COGS: 10 percent
- Operating margin: −4 percent
Classic Rule of 40 reads: 45 + (−4) = 41. Comfortably above the 40 bar. The board feels good.
Adjusted Rule of 35 reads: 41 against a 35 bar. Clearing by 6 points. With the +5 compounding capability credit applied to the trajectory, true performance reads 46; the company is genuinely outperforming for its category.
The inference-COGS band check: at 18 percent, the workload is in the unfavorable end of the band. Apply a 3-point inference-COGS debit. Adjusted reading becomes 38 against the 35 bar; still clearing, but with less cushion.
The compounding return overlay: if the company budgets and captures the 30 to 50 percent year-2 inference savings, inference COGS lands at roughly 12 percent in year two. Operating margin lifts by 5 to 6 points without growth slowing. The Rule of 35 reading rises to 46 by year two. The trajectory clears classic Rule of 40 by year three even with no further growth acceleration.
This is the example that makes the case for the recalibration. Classic Rule of 40 looked at this company and said “fine.” Adjusted Rule of 35, applied with the offsets and the compounding overlay, says “fine today, healthy tomorrow, possibly outstanding by year three.” That is more information, not less.
How the compounding return interacts
The Rule of 35 is a static metric; the AI-native cost trajectory is dynamic. Year-over-year compounding return on inference cost; detailed in the year-2 savings article linked above; is what makes the adjusted rule a forward-looking instrument.
A company at Rule of 35 in year one, with compounding capture properly budgeted, typically reaches Rule of 38 to 40 by year two and Rule of 40 to 43 by year three on stable growth. The path is real and reproducible across mature 2026 AI-native operators. The path requires that the budget structure capture the savings rather than absorb them; without explicit governance, the compounding return is silently spent on capability expansion and the metric does not move.
Reporting both the Rule of 35 reading and the compounding capture rate gives investors the trajectory. A company reporting Rule of 35 at 38 with year-over-year inference COGS shrinking 12 percent is a different investment than one reporting Rule of 35 at 38 with inference COGS flat. The static metric alone cannot distinguish them; the metric plus the trajectory can.
Reporting the metric to investors
A board reporting AI-native SaaS metrics in 2026 should consider three additions to the standard package.
Addition one: workload-segmented Rule reporting. Pure-AI revenue lines apply Rule of 35; pure-legacy revenue lines apply Rule of 40. Mixed revenue weights the rule by the AI revenue fraction. A SaaS company that is 30 percent AI-revenue runs an effective rule of 38.5 on a blended basis. Reporting the segmentation is more useful than the blended number.
Addition two: inference COGS as percent of revenue, with year-over-year delta. A standalone line. The trajectory is the point. A company holding 12 percent inference COGS while growing at 45 percent is doing different and harder work than one absorbing inference savings into capability without the line moving.
Addition three: eval-engineering opex as a separate line, not bucketed into R&D. Treating eval-engineering opex as an undifferentiated R&D expense hides the structural rigidity that the Rule of 35 debit reflects. Breaking it out is a small reporting change with a meaningful clarity gain.
The first investor or board member who asks for these breakdowns will find that the engineering team already has the data; the data lives in the same systems that produce inference cost dashboards. The breakdown is a reporting change, not a measurement change.
Frequently asked questions
Is the Rule of 35 framework specific to 2026 model pricing? The framework is general; the specific bar reflects 2026 inference economics. As inference cost compresses further, the bar may rise toward 38 to 40 over the 2027 to 2028 window if compounding capture continues to outpace usage growth.
Does the rule apply to AI inference platforms (LLM API providers)? No. Platform economics are different; they sell the inference, they do not buy it. Platform metrics need their own recalibration centered on tokens served per dollar of compute infrastructure.
How does the rule interact with PLG-driven AI products? PLG AI products tend to have higher inference COGS as a percent of revenue early in the funnel because free-tier inference is real cost without offsetting revenue. The inference-COGS band shifts toward the unfavorable end during PLG expansion and recovers when monetization closes.
Does the rule penalize companies investing heavily in eval engineering? The 3-point debit captures structural rigidity, not strategic investment. A company that is investing temporarily above 9 percent eval opex to build a competitive moat in evaluation rigor should treat that as a capacity build, not steady-state opex.
What about open-source self-hosted inference? Self-hosted inference shifts the cost from variable per-token to fixed per-GPU-hour. The Rule of 35 still applies; the band reads differently because COGS includes amortized GPU lease cost and the variability lives in utilization rate.
How does the rule treat AI features bolted onto a legacy SaaS product? Workload-segmented reporting handles this cleanly. The legacy product runs Rule of 40; the AI feature runs Rule of 35; the blended reading weights by revenue.
Should the rule apply to AI services and agencies, not just SaaS products? The rule is product-shaped. Service businesses (agencies, consultancies) have different unit economics dominated by gross margin per engagement. The economics manifesto’s principles apply; Rule of 35 specifically does not translate cleanly.
Is there a Rule of 50 for outperforming AI-native SaaS? Some operators frame Rule of 50 as the next-tier bar. The argument here would price it as Rule of 45 for AI-native, applying the same offsets; meaning an AI-native SaaS clearing growth-plus-margin 45 is performing at a Rule-of-50 trajectory.
How sensitive is the rule to the inference cost trajectory next year? Highly sensitive. A continued 35 percent annual decline in flagship inference cost compresses the inference-COGS band toward the favorable end and may eliminate the 3-point structural debit by 2028. The rule is a 2026-period instrument; it should be re-examined annually.
Key takeaways
- Classic Rule of 40 mis-prices AI-native SaaS because inference cost is linear in usage and eval-engineering opex is rigid against margin pressure.
- The recommended adjustment is Rule of 35 base, with three offsets: +5 compounding capability credit, −3 eval-rigidity debit, ±2 to 5 inference-COGS band.
- Net adjusted bar of 37 puts an AI-native SaaS clearing 37 in approximate parity with a legacy SaaS clearing 40.
- Inference COGS runs 8 to 22 percent of revenue depending on workload shape; eval-engineering opex runs 4 to 9 percent; observability COGS runs 3 to 6 percent.
- The compounding return on inference cost is what makes adjusted Rule of 40 territory achievable by year two to three, on stable growth and budgeted capture.
- Investors should see workload-segmented Rule reporting, inference COGS with YoY delta, and eval-engineering opex as a separate line.
- Worked example: a 45 percent grower with 18 percent inference COGS and −4 percent operating margin reads 41 on classic Rule of 40, 46 with the compounding credit on adjusted Rule of 35.
- The framework is 2026-specific; re-examine annually as inference economics continue to compress.
Arthur Wandzel
