You're not a SaaS Company, You're a Compute Company.

If I had a dollar for every time I say this to founders.

Investor: “You cannot run an AI company like a SaaS company. Your biggest cost isn’t headcount anymore, it’s intelligence. And every time, your model improves, your P&L changes.”

For the past decade and a half, founders and investors have spoken a common financial language: ARR, CAC:LTV, gross margin, burn multiple - the standard startup scoreboard. The playbook was - if you could grow recurring revenue, manage burn, and keep gross margins above 70%, it was “working”.

That would make sense when software had almost zero marginal cost.

You built the product once, hosted it cheaply in the cloud, and sold it infinitely - the standard SaaS model. But with AI: every query is burning compute and every new feature changes your unit economics.

And the most important metric is how efficiently you turn computation into value, which (surprise!) doesn’t appear on any SaaS dashboard.

Traditional software metrics rest on three quiet assumptions:

1. The cost of serving one more user is basically zero

2. The cloud costs scale predictably with usage

3. Your unit economics are stable regardless of how your product evolves

But, with AI:

1. The cost of serving one more user is basically zero → Your marginal cost now depends on how long a user’s context window is, which model tier you route to, and how efficiently your caching works.

2. The cloud costs scale predictably with usage → Your cloud cost graph no longer rises smoothly - it jumps, dips, and occasionally explodes.

3. Your unit economics are stable regardless of how your product evolves → Improve quality too aggressively and you might double your inference bill overnight.

While ARR and gross margin still matter, they are backward-looking. They tell what you sold, not how efficiently your system learns.

AI products aren’t static software. They are living systems that learn, infer and adapt. Every inference burns compute and every learning cycle refines performance that costs money. That means the real discipline in building an AI company is compute discipline: how efficiently you convert intelligence into margin.

In SaaS era, investors underwrite market size.

In AI era, they are beginning to underwrite compute efficiency - the rate at which your system gets cheaper as it gets smarter.

If SaaS rewarded top-line velocity, AI will reward “learning efficiency”.

The three numbers that actually matter when it comes to looking at AI metrics:

1. Contribution Margin per Task, or CMPT

   Forget company-level gross margin for a moment. Let’s assume that the atomic unit of your business is a completed task - basically a query answered a code snippet generated, a workflow executed. CMPT measures what you earn on that event, i.e., revenue minus the variable costs of producing it (inc. tokens, infrastructure, evaluations, and any human review that closes the loop).

   CMPT = Price per task - (variable COGS per task)

   This tells you whether each customer interaction is profitable.

   Founders: This is your true unit economics - the foundation for pricing, comp plans, and board conversations.

2. Token ROI

   Every time you upgrade a model or expand context length, your cost curve shifts. Token ROI captures whether those improvements are worth it. It’s the ratio of improvement to cost - the modern equivalent of margin leverage, a metric that investors will eventually demand.

   Token ROI = % gain in accuracy / % increase in token cost

   High Token ROI = You are getting smarter faster than you are getting expensive

   Low Token ROI = Vanity upgrades

3. System Learning Rate, or SLR

   Your model learns from data and feedback. Track how efficiently it converts new experience into performance.

   SLR = Delta (success rate or automation rate) / Delta (data volume or usage)

   A rising SLR means your system is compounding intelligence.

   A flat SLR means you are spending compute just to stand still.

Growth in SaaS was linear: sell more, onboard more, expand accounts and repeat.

Growth in AI is circular: A better model creates better outcomes, which create more users, which generates more data, which reduces cost and improves quality, which creates better outcomes again. It’s a learning flywheel, not a sales funnel.

So, your job as a founder is to keep that loop spinning.

You’ll know it’s working when task success rates climb, token cost per task falls, contribution margins improve, and human-in-the-loop intervention steadily declines.

In SaaS, architecture affected the UX, but in AI, it basically defines your balance sheet. There is no surprise that every technical decision in AI is a financial one in disguise.

• If you choose a larger model → your inference costs rise

• You shorten your context length or improve caching → your margins improve

• Fine-tune instead of prompt-engineer → you shift spend from OpEx to CapEx

That’s why the best AI founders treat their infrastructure dashboards like financial statements. They know their cost per thousand tokens the way a SaaS founder knows CAC. They understand that routing logic is as strategic as pricing, and that latency targets double as margin levers.

SaaS founders could chase top-line growth and clean up margins later, but AI founders don’t have that luxury.

Your cost base scales with every user interaction, so efficiency has to be designed in from day one.

And the good news is that founders who internalize this early will have an enormous edge. They will know how to talk to investors in a language that reflects how their product actually works. They can manage compute like cash and can design decisions that compound learning and margin together.

The views expressed are those of the author and do not necessarily reflect the views of any investment firm or portfolio company.