KPI Spotlight: AI-Adjusted Gross Margin — The Metric That Broke SaaS Benchmarks on September 25, 2026

For roughly fifteen years, gross margin was the most boring number on a SaaS income statement. You hosted on AWS, you hired a support team, you amortised some capitalised software, and you landed somewhere between 75% and 85%. Investors glanced at it, nodded, and moved on to net revenue retention.

That era ended this week. GitLab crossing $1.13 billion in revenue with 24% billings growth — while absorbing 400 basis points of gross margin cost from AI — is the cleanest public confirmation yet that inference is not a transitional expense. It is a permanent line item in cost of revenue. And on the same day, OpenAI shipped GPT-6 “Astra” with AGI-era positioning, meaning the capability baseline every SaaS product benchmarks against just moved again, at a cost structure you do not control.

So today’s KPI Spotlight is gross margin — but the version that actually matters in 2026: AI-adjusted gross margin, and its operational sibling, cost of inference per active user.

Definition: What AI-Adjusted Gross Margin Actually Measures

The base formula hasn’t changed:

Gross Margin % = (Revenue − Cost of Revenue) / Revenue

What changed is what belongs inside cost of revenue. The 2026 COGS stack for a SaaS company shipping LLM features looks like this:

• Hosting and infrastructure — compute, storage, egress, CDN
• Model inference — per-token API spend, fine-tuning runs, embedding generation, vector store queries
• Third-party data and API licences — enrichment, transcription, search, verification
• Customer support and success — the portion tied to delivering the service, not expanding it
• Professional services delivery — implementation, migration, solution architecture
• Payment processing — often mis-parked in opex
• Capitalised software amortisation

The single most common reporting failure right now: teams bury inference spend inside R&D because it started as an experiment. That flatters gross margin, misleads the board, and — critically — hides which features are structurally unprofitable. If a customer action triggers a model call, that cost is COGS. Full stop.

The companion metric you need: AI Margin Drag

AI Margin Drag (bps) = (Inference + AI infra COGS / Revenue) × 10,000

GitLab’s number is roughly 400bps. That is now the public benchmark. If your AI margin drag is materially above it and you are sub-$100M ARR, you are subsidising usage you have not priced for.

Benchmarks: Where Gross Margin Sits in Late 2026

Segment your comparison set properly, because a single “SaaS” band is now useless:

• Classic multi-tenant SaaS, no heavy AI: 78–85% gross margin. Freshworks and Zoho-class horizontal software has historically operated in the low-to-mid 80s on a non-GAAP basis. This is still the gold standard.
• Enterprise SaaS shipping LLM features: 74–80%. GitLab is the reference point — strong margins, minus roughly 4 points of AI cost that is not coming back.
• AI-native application layer: 55–70%. Inference dominates COGS. Stripe’s data on AI-native companies growing 175% year over year shows the growth is real; the margin profile is simply structurally lower than 2019-vintage SaaS.
• Vertical marketplaces and booking platforms: highly variable, typically 60–75% depending on payments mix. Fresha’s $1B valuation on an $80M KKR-backed round shows growth equity still underwrites these — but payment processing sits in COGS and caps the ceiling.
• Thin AI wrappers: 35–55%, and falling. GPT-6 Astra’s launch just repriced the “we prompt better than you” moat to approximately zero. If frontier capability absorbs your feature layer, you lose pricing power and keep the inference bill.

The uncomfortable implication: reforecast your 2027 gross margin target down 300–500bps if you ship LLM features. Not as a worst case