The Broken SaaS Dashboard

You're running two businesses. Your P&L shows one.

I haven’t separated my AI costs from my legacy SaaS P&L yet. I’m doing it next quarter — when we roll out more AI features — because at that point I won’t have a choice.

That’s not a confession. It’s where most SaaS CEOs are right now, whether they know it or not.

Our cloud costs are up 30 to 50 percent year over year. Some of that is growth. Most of it is inference — the cost of running AI features for customers. When my PE investors asked about our AI roadmap last month, I came with three buckets: internal productivity, engineering output, and product value. They nodded. Not because the numbers were large, but because the framework was honest.

What I didn’t have was a clean answer to a simpler question: what is our AI gross margin?

I knew our blended margin. I didn’t know the split.

Here’s why that matters.

Traditional SaaS is built on one of the best business models ever invented. Software costs almost nothing to replicate. You write the code once, sell it a thousand times, and gross margin approaches 80 percent because the marginal cost of the next customer is nearly zero. Every metric we use — gross margin, NRR, Rule of 40, LTV:CAC — was built for that world.

AI broke the assumption underneath all of them.

Inference costs money every time a customer uses a feature. Not a fixed amount — a variable one that scales with usage. ICONIQ’s 2026 data puts inference at roughly 23 percent of revenue at scaling-stage AI companies. For traditional SaaS adding AI features, the margin hit is 12 to 17 points immediately, without raising prices. AI-native products run at about 52 percent gross margin in 2026, against the 70 to 80 percent traditional SaaS has always targeted.

Add an AI layer to a SaaS business without separating the P&L and you’re adding manufacturing economics to a software business. Then reporting only the software number.

Salesforce found out what this looks like at scale. Agentforce processed nearly 20 trillion tokens against $800 million in ARR. Customers ran 80 percent fewer human seats — and got billed 83 percent more. Per-seat metrics stopped working. NRR became a function of consumption patterns, not renewals. Forecasting got harder.

We’re not at that scale. But the dynamic is the same — just smaller and earlier.

The pattern shows up in aggregate too. Median Rule of 40 across public SaaS is 28 percent today. Only 20 percent clear the threshold. The floor didn’t drop because growth slowed — it dropped because AI spend is hitting margins faster than AI revenue is arriving.

Eighty percent of M&A buyers cite AI commoditization as the top risk to SaaS valuations. Twenty-five percent of CEOs do. That 55-point gap doesn’t close with better strategy. It closes when the P&L forces the conversation — at a board meeting, in an LOI, or when a quarter comes in that nobody can explain cleanly.

My forcing function is next quarter’s feature rollout. Others find out later.

What I’m doing right now is simpler than it sounds. I started tracking AI inference cost as its own line — separate from everything else. That one number, once you have it, starts to reveal what the blended dashboard hides. What it’s doing to gross margin. Whether NRR is healthy because customers value the product or because they haven’t hit their usage ceiling yet. Whether the AI layer is getting more efficient as you scale or just getting bigger.

Three numbers I didn’t have six months ago that I want every month now: AI product revenue divided by inference cost — how much revenue for every dollar of inference spend. AI COGS as a percentage of AI revenue — whether margins are improving or just growing. And gross margin by revenue stream — legacy SaaS and AI layer reported separately so you can see which business is actually healthy.

One question before the next AI feature goes in: can you draw a straight line from this spend to GRR, NRR, or gross margin? If you can’t draw it, it’s a bet. Name it as such.

None of this is complicated accounting. It’s a decision to look.

Separating the AI P&L from legacy SaaS isn’t a finance exercise. It’s the difference between managing the business and being surprised by it.

The split doesn’t create the problem. It just tells you how long it’s been there.