Remember when SaaS was supposed to have near-zero marginal cost per customer? When scaling meant adding seats without breaking your gross margin? That era ended in Q1 2026.

The math is brutal. Traditional B2B SaaS enjoyed 75–85% gross margins because, once built, software scales infinitely. One developer's code serves a thousand companies at virtually zero incremental cost. AI changes that equation entirely. Every inference—every query, every completion, every reasoning step—costs money. Tokens consumed become tokens paid.

AI-first SaaS is now looking at 40–60% gross margins as a ceiling, not a floor. Some vendors are worse. In early 2026, founders are discovering that pilots running on generous credits scale to production at 500–1,000% higher cost than expected. A $50K pilot deal suddenly looks like a $250K annual burn on compute alone. The unit economics blow up before the deal closes.

This isn't theoretical. Snowflake's new Cortex Code agent, launched February 2026, exemplifies the problem. It's native to Snowflake's data warehouse—brilliant positioning—but every code generation, every debug cycle, every multi-file refactor is a token burn. Snowflake hasn't disclosed margins, but industry insiders expect Cortex Code will hit 50% gross margins at maturity, not the 70%+ the core warehouse enjoys.

Meanwhile, 85% of developers now use some form of AI coding assistant. GitHub Copilot, Claude Code, Cursor, Amazon Q. They're duking it out on quality and context-awareness, not price. That's because pricing in the coding-assist market is already compressed to near-commodity levels. GitHub charges $10/month. Cursor is free for personal use. The barrier to switching is friction, not economics. Churn is a latent threat.

The forced pivot: SaaS vendors are experimenting with three escape routes, none of them comfortable.

Route One: Usage-Based Pricing. Tie revenue to tokens consumed, workflows executed, or transactions processed. This offloads compute risk to customers. If an AI SaaS vendor can't control inference costs, pass them through. But customers hate this. Unpredictable bills kill land-and-expand economics. Usage-based churn is brutal because price is variable and visible. A customer scales up usage and suddenly their software bill spikes 40%. They swap tooling overnight.

Route Two: Hybrid Seat-Plus-Usage. Reintroduce per-user licensing caps on monthly compute. You get 1M tokens per seat per month; overage is $0.50 per 100K. Vendors like Notion and Figma are testing variants. This preserves predictability for customers whilst containing vendor margin bleed. But it reintroduces the per-user friction that SaaS was designed to eliminate. In a world where one AI-augmented employee replaces five, selling seat licenses is tone-deaf.

Route Three: Outcome-Based Pricing. Charge based on measurable business impact delivered—revenue generated, cost saved, accuracy improved. This is conceptually elegant: vendor and customer align on value. In practice, it's a CFO's nightmare. Attribution is impossible. Disputes are endemic. Only deeply embedded, enterprise suites with strong customer relationships can pull it off.

The market is voting. AI-native SaaS startups captured 41% of all SaaS venture funding in Q1 2026—the highest share ever. Seed rounds for AI application startups are averaging $2–5M, with capital earmarked not for infrastructure but for proprietary data and aggressive go-to-market. The venture market believes the puck is moving toward native AI, not augmented SaaS.

But here's the twist: the $2 trillion SaaS market isn't dying. The Gartner prediction that 35% of point-product SaaS will be replaced by AI agents by 2030 is not a death sentence for software; it's triage. Weak products churn first. Deeply embedded suites expand. Accounting platforms, CRM systems, data warehouses—the systems of record—continue to add seats because they're not threatened by AI agents. What's threatened is the thin, horizontal, feature-light SaaS tool that sits between a user and a workflow.

The real disruption isn't AI replacing SaaS. It's the collapse of the unit economics that made SaaS the most profitable business model in tech. For founders still selling on a per-seat, flat-fee basis with 70%+ gross margins, the clock is ticking. For founders building inference-native products and accepting 50% margins in exchange for network effects and defensibility, the opportunity is now.

The margin-halo around SaaS is gone. What remains will be earned, not inherited.