AI Revenue is a critical Signal. But it's not SaaS Revenue.

I've spent most of my career in SaaS. I know the model well: high upfront development costs, relatively low marginal costs to serve each additional customer, and the kind of gross margins that make your bank account feel like a spreadsheet with a typo in it. Seventy, eighty, ninety percent. SaaS is a beautiful business model if you can get there.

So when I started hearing investor friends obsess over AI startup revenue growth the way they used to obsess over SaaS ARR, I was confused. Not the revenue-is-good part. That makes sense. The confusion was around the implicit assumption that the revenue number told the same story it used to.

It doesn't. Not even close.

The Margin Story Has Changed Completely

Let me start with the number that should be changing every conversation about AI startup revenue but somehow isn't: gross margin.

In SaaS, gross margins between 75% and 90% are considered standard. Companies that fall below 70% face real valuation pressure -- the median revenue multiple drops from 7.6x to 5.5x. The whole SaaS investment thesis is predicated on this: yes, you might be burning money now, but the marginal cost of that next customer is nearly zero, and once you find product-market fit, the unit economics are spectacular.

AI companies are operating in a fundamentally different world. Bessemer Venture Partners, in their State of AI 2025 report, found that the fastest-growing AI startups -- what they call "AI Supernovas" -- average about 25% gross margins early on. Many have negative gross margins. That's not a rounding error. That's a structurally different business.

The numbers get sharper when you look at specific companies. Anthropic had gross margins of negative 94% to negative 109% in 2023 and 2024 -- meaning they spent more on infrastructure than they collected in revenue. They projected 50% gross margins for 2025, then revised that down to 40% because inference costs on Google and Amazon cloud infrastructure climbed faster than expected. By 2028, they expect to reach 77% gross margins. That's a compelling trajectory, if you believe it.

Then there's Perplexity, which spent 164% of its revenue in 2024 on AWS, Anthropic, and OpenAI combined. One hundred and sixty-four percent. They were, by any reasonable definition, paying to give away their product. Yet they were valued at $9 billion in December 2024.

Here's the thing nobody says out loud often enough: a significant chunk of what AI companies report as "revenue" is token pass-through. They're charging customers for API calls to Anthropic or OpenAI, and booking the difference as top-line revenue. As Bessemer notes, AI economics mean "every AI query incurs real compute costs" -- tokens, context windows, tool calls, agent steps, eval runs, and orchestration retries. If your markup on all of that is thin (or negative), the revenue number becomes deeply misleading.

PitchBook put it bluntly: for most AI-native companies, "ARR isn't really ARR as we know it, but rather a hotchpotch of one-off, credits-based, performance-based or outcome-based revenue contracts." The predictability and stickiness that made SaaS ARR such a reliable signal? Gone.

So why are investors still using revenue as their primary gauge? I think the answer is actually correct, even if the reasoning is muddled.

Revenue Is the New Monthly Active User

Here's where I've landed: investors aren't wrong to look at revenue growth as their primary signal for early AI companies. They're just wrong about what it's a signal for.

Revenue for an AI company right now is not what revenue was for a SaaS company in 2015. It's what Monthly Active Users were for a Web 2.0 company in 2010.

Think back to Facebook, Twitter, Uber, Airbnb in their early days. None of them had great margins at the start. Uber and DoorDash were literally subsidizing rides and deliveries, burning VC money to build supply-demand flywheels. DoorDash went public in 2020 at a $72 billion valuation despite generating $2.9 billion in revenue -- investors were reaching for GMV (gross merchandise value) as a proxy metric because it made the valuation look more defensible. Airbnb was paying for professional photography for hosts to get listings to convert. Twitter was running at a loss for years. None of this was considered disqualifying. The question wasn't "what are your gross margins?" It was "how fast are you growing engagement, and how sticky is it?"

The canonical extreme case: Facebook paid $19 billion in 2014 for WhatsApp, a messaging app with essentially no revenue model -- users paid $0.99/year after a free first year, and the app served no ads. The implied price was $42 per user. WhatsApp had 450 million users and was adding one million new users per day. Its 72% DAU/MAU ratio was cited as the primary justification. NYU's Aswath Damodaran wrote at the time that the acquisition was "pricing" not "valuing" -- driven by competitive fear rather than fundamental analysis.

Sound familiar?

The shift from page views to MAU to DAU to engagement ratio was the story of the Web 2.0 era. Each generation of metric was a better proxy for something that actually mattered: is this thing becoming a habit? AI revenue is playing the same role right now. It tells you: are people using this enough to pay for it? Is there a real workflow being replaced or enhanced? Is there enough value that customers are willing to run up a usage bill?

In an early-stage market where product-market fit is the dominant risk, that's exactly the signal you want. The problem isn't that investors are tracking revenue. The problem is that they're applying the old framework's interpretation to the new framework's metric.

The Web 2.0 Parallel Is More Apt Than It Seems

There's another dimension of the analogy that I think gets overlooked: many of those companies were running at cost or at a loss because of their VC-backed growth strategy. That wasn't a bug. It was the feature.

Uber raised money to subsidize fares because the strategy was to get liquidity in every market before anyone else could. DoorDash burned cash to build the supply side of their marketplace. These weren't financially naive decisions. They were bets that the market would consolidate, that the unit economics would improve at scale, and that the category winner could eventually charge what the market would bear.

AI-native startups are following the same playbook. Bessemer found that their cohort of AI Supernovas reached approximately $40M ARR in their first year of commercialization and $125M ARR in their second year. These companies are growing at a rate that traditional SaaS essentially never achieved. And they're doing it by offering access to frontier models at prices that may not fully cover their costs, because they're betting that the category win is worth more than the near-term margin.

AI-native startups are growing roughly 2x faster than traditional SaaS at the sub-$1M ARR stage. Some are reaching $30M ARR in 20 months -- a journey that takes traditional SaaS companies 100+ months on average. That kind of growth rate only makes sense as a land-grab strategy. Investors know this. Tracking revenue is how they track the land.

The Treadmill Nobody Wants to Talk About

The bull case for AI margins improving over time relies heavily on inference costs continuing to drop. This argument is real and well-documented. A16z has coined the term "LLMflation" to describe it: when GPT-3 became publicly accessible in November 2021, it cost $60 per million tokens. By late 2024, the cheapest model to achieve the same GPT-3 benchmark score cost $0.06 per million tokens -- a 1,000x decline in three years.

Epoch AI measured the median cost decline at 50x per year, with some benchmarks falling even faster. If you believe these curves continue, the margin story for AI companies looks very different in 18-36 months.

But there's a catch that doesn't show up in the "cost per token" charts.

Frontier reasoning models -- the ones customers increasingly demand because they're better -- are getting more expensive, not cheaper. The cost declines apply primarily to older, non-reasoning models. And competitive pressure forces companies to use frontier models just to stay relevant. As SaaStr described with a specific portfolio company: at $100M ARR, they were modeling $6M in incremental inference costs over the next 12 months -- not because their current product was broken, but because they needed better models to stay ahead of competitors. That's 6 points of gross margin voluntarily surrendered to competitive pressure.

The treadmill problem: inference costs fall 50x per year on old models, while frontier models keep getting more expensive and users keep expecting frontier performance. You can run fast and still fall behind.

So What's the Problem?

The issue isn't that investors are looking at revenue. It's that the conversation often happens without acknowledging what revenue actually represents right now.

When someone says "Company X is at $50M ARR," the instinct trained by a decade of SaaS investing is to mentally apply a gross margin assumption and get excited about the implied business quality. But that assumption breaks completely if Company X is running at 25% gross margins with negative unit economics on every customer.

There's a version of this that ends well: inference costs continue to fall, companies build proprietary data moats and switching costs, pricing power materializes, and the Web 2.0 analogy proves out completely. Uber's unit economics did improve -- they turned their first profitable year in 2023. Airbnb is now quite profitable. Meta's gross margins are well north of 70%.

And there's a version that ends badly: costs plateau, commoditization sets in, and companies find themselves stuck with thin margins and customers who switch providers whenever a cheaper API becomes available. The dot-com era had plenty of those -- Pets.com, Webvan (which raised over $1 billion from Sequoia and still went bankrupt), Kozmo.com. All of them had great traffic numbers and no path to sustainable margins.

The dot-com failure wasn't that investors used proxy metrics. It was that they stopped asking whether the underlying business could ever support the proxy at positive economics. The Wikipedia post-mortem is blunt: analysts "focused on aspects of individual businesses that had nothing to do with how they generated revenue or cash flow."

The Right Way to Read AI Revenue

If I were evaluating an AI company right now, here's what I'd actually want to know alongside the revenue number:

What percentage of revenue is token pass-through? There's a meaningful difference between charging $100 and keeping $80 versus charging $100 and keeping $25. Both show up identically on a revenue slide.

Is the revenue sticky? PitchBook noted that enterprise AI customers often start with pilots and trials carrying high churn rates. A company with $30M in "ARR" that hasn't completed a single renewal cycle tells you almost nothing about long-term retention. SaaS Capital has started warning explicitly: don't count paid pilots or proofs-of-concept as ARR, since conversion rates under 50% materially inflate the number.

What's the cost trajectory for the models they need? Not models in general -- the models their customers demand. If competitive pressure forces them onto frontier reasoning models, declining cost-per-token on smaller models is irrelevant.

Consumer or enterprise? Consumer AI businesses are structurally more similar to Web 2.0 consumer businesses -- engagement, habit, and retention are the real metrics. Enterprise AI is trying to be SaaS but hasn't gotten there on margin structure yet.

Revenue in the AI era is a symptom of product-market fit, not proof of business quality. That's not a condemnation -- product-market fit is exactly what you're trying to de-risk in the early stages. But the signal it sends, and the framework you apply to interpret it, has to match the era you're actually in.

Investors aren't crazy for chasing AI revenue growth. They're just using an old label for a new kind of metric. The sooner everyone is clear on what that metric actually means, the better the decisions will be on both sides of the table.

Find me on LinkedIn if you want to argue about this. I have strong opinions, loosely held.