Investor note

Where AI creates margin expansion and where it destroys it

May 2026.7 min read

Audience: Public-equity software investors, PE software underwriters, and CFOs of B2B software businesses.

The market claim

The market is still treating AI revenue as fungible with SaaS revenue at the gross margin line. The same dollar of recurring revenue is being priced through the same multiple regardless of the cost structure delivering it. By Q1 2026 that reading had become indefensible. The public software basket erased about $2 trillion in market capitalisation as investors began pricing AI into the cost side rather than only the demand side (FinancialContent, The SaaSpocalypse: AI Agents Trigger a Massive Repricing). ICONIQ’s January 2026 snapshot put average AI product gross margin at 52 percent, up from 41 in 2024 and 45 in 2025, but still well below the 75-to-80-percent SaaS norm (Monetizely, The Economics of AI-First B2B SaaS). Public software is now disclosing 60 to 70 percent as the realistic AI-adjusted gross margin range (SaaSMag, The AI COGS Problem). The eighty-point standard is structurally weaker than it looks on a current earnings deck.

What the market is mispricing

AI does two things to the income statement that are usually conflated and shouldn’t be. It can raise gross margin, and it can destroy it. Which one happens is not a function of the strength of the AI feature. It is a function of the pricing model, the architecture, and where the inference cost lands inside the contract envelope. This is the AI margin illusion in its operating form. Revenue rises with feature adoption, but COGS rises faster if any one of three configuration choices is wrong.

A SaaS product generating $100 of revenue with $20 of traditional COGS gives an 80-percent gross margin. Adding AI features with $15 of direct variable cost changes COGS to $35 and drops gross margin from 80 to 65 (The SaaS CFO, Your AI Feature Is Quietly Destroying Your Gross Margin). Inference now averages 23 percent of total revenue at scaling-stage AI B2B companies, which means $230,000 of every $1 million of AI revenue is consumed by inference before engineering, sales, or marketing get paid (SoftwareSeni, Why AI Gross Margins Are So Much Lower Than SaaS).

The question is not whether the company has AI revenue. It is which side of the margin equation the AI revenue sits on.

The configurations that raise gross margin

Three configurations raise gross margin under AI economics. Each rests on the same mechanic: keeping variable cost off the revenue line, or capturing value faster than the cost grows.

The first is proprietary inference. Owning the model, hosting it, and routing customer requests against it removes the marginal-token cost that erodes wrapper businesses. Bessemer notes that hybrid pricing built around proprietary inference and outcome-tied charging is the dominant transition state in 2026 enterprise renewals (Bessemer Venture Partners, The AI Pricing and Monetization Playbook).

The second is value-based or outcome-based pricing. The unit being sold is the customer outcome (a closed support ticket, a generated approval, a completed transaction), not the seat. When pricing is anchored to value created rather than seat consumed, inference cost can rise inside the cost structure without compressing gross margin, because the unit price moves with it. ServiceNow Now Assist is tracking to $1.5 billion in ACV for 2026 with 50 percent of net new business already on non-seat pricing tied to tokens and consumption.

The third is infrastructure gravity. Companies positioned at the data and runtime layer underneath the agent layer earn margin on every agent built on top of them. Snowflake, Datadog, and Cloudflare strengthened against the Q1 2026 basket selloff on this logic; each owns infrastructure depth that agents need, not replaces (FinancialContent, Why B2B Software Stocks Are Plunging 20%).

The configurations that destroy gross margin

Three configurations destroy gross margin under AI economics. Each rests on the same mechanic: variable cost growing inside a fixed-price contract.

The first is per-seat pricing with usage-based inference cost. The seat price is fixed at renewal. The inference cost moves with adoption. Higher adoption then compresses gross margin even when revenue is flat, and any agent-driven seat compression compounds the problem on the revenue side. Atlassian disclosed its first ever decline in enterprise seat counts in 2026 from AI agent adoption replacing human-driven tasks, and the gross margin profile suffered through both pricing and cost (The SaaS CFO, Death of Per-Seat Pricing).

The second is undifferentiated AI wrappers. A copilot built on a third-party model, sold as a feature flag inside a SaaS contract, with no proprietary data layer underneath, has every dollar of inference cost passing through to the vendor’s COGS while the customer pays the original SaaS list price. The wrapper category compresses fastest because there is no defensible mechanism to reprice up the curve. As feature parity becomes reachable in a quarter using AI coding agents, the wrapper has no time to amortise the cost.

The third is customer-borne API costs hidden inside the contract. When the contract obliges the vendor to provide the AI capability and the inference cost is paid out of vendor margin, every customer that succeeds with the feature reduces vendor profitability. This is the worst configuration of the three because it creates a perverse incentive: deeper customer success worsens the P&L.

Upside and risk

Upside. A software business sitting on proprietary inference, value-based pricing, or infrastructure gravity is mispriced cheap on a current multiple, because the gross margin profile is durable through the transition and the AI revenue line is an actual margin expansion event. Thoma Bravo and Vista are openly arguing this case in their LP communications, with portfolio statements showing the majority of their software companies are not experiencing meaningful churn or losing customers to AI alternatives (Bloomberg, Thoma Bravo and Vista on AI; Yahoo Finance coverage).

Downside. A software business sitting in any of the three margin-destroying configurations is mispriced rich because the gross margin in the current model is overstating the durable margin. The path from 80-point gross margin to 60-point gross margin can run in two to three renewal cycles when the company holds per-seat pricing against rising customer adoption of AI features. The eighty-point standard the position was underwritten against does not hold. When the disclosure adjusts, both the multiple and the multiplicand fall at the same time.

Watch this, not that

Watch the cost-side breakdown alongside the revenue side. Specifically, watch disclosed inference cost as a percentage of AI product revenue, the share of net new business on non-seat pricing, and the gross margin published on a constant pricing-mix basis. Stop watching headline AI revenue growth without the cost configuration behind it. Growth in any of the three margin-destroying configurations is the wrong thing to underwrite.

Three diligence anchors for any 2026 software position:

A configuration map of the AI revenue line: what share is on proprietary inference, value-based pricing, or infrastructure gravity, versus per-seat with usage-based inference, undifferentiated wrappers, or customer-borne API.
A gross margin sensitivity at year three under a forced shift to hybrid pricing, with inference cost curves from the company’s own data or a credible third-party benchmark.
A read of contract structure on the top ten customer accounts to test where inference cost actually lands inside the envelope.

The eighty-point standard is no longer the default. The configuration is.

Sources

margin, pricing, investors, diligence, multiples

Back to writing