AI Companies Are Priced Like SaaS - Their Costs Say Otherwise

For most of the last decade, tech investors have used a simple mental model: software is a fixed-cost business. You spend a fortune building the product once, then every extra customer is almost pure profit. Netflix doesn't spend meaningfully more serving subscriber 300 million than it did serving subscriber 1 million. That's what let SaaS companies trade at 5-10x revenue and still look reasonable; the market could see the margin expanding as they grew.

AI companies are being priced the same way. OpenAI trades at roughly 38 times revenue. Anthropic has bounced between 20 and 60 times. Those are software multiples. The problem is AI isn't running software economics.

The catch: usage doesn't get cheaper, it gets more expensive

Every time someone sends a prompt, a model has to actually run; real GPU time, real electricity, real cost, every single query. Unlike a traditional app, there's no point where an AI company's cost per user flattens out. Industry data backs this up starkly: AI companies are running gross margins around 50-60%, compared to 80–90% for traditional SaaS. And it's getting worse with scale, not better; research from ICONIQ found that as AI products mature, the share of costs going to compute actually climbs (from around 20% to 23%), while the AI companies' own internal cost management (like headcount) shrinks as a share of spend. Normally, scale is what fixes your margins. In AI, scale is what stresses them.

The $20-a-month problem, with an actual number on it

This is the part that's easy to say anecdotally - now there's a hard figure. SemiAnalysis, an independent research firm, didn't guess at this, they tested it: they bought a $20 ChatGPT Plus plan, a $200 ChatGPT Pro plan, and a $200 Claude Max plan, then ran them as hard as possible until they hit the usage caps. The results:

- A heavily-used $20/month ChatGPT Plus plan works out to roughly $700 worth of compute at normal pricing

- A maxed-out $200/month ChatGPT Pro plan can cost OpenAI up to around $14,000 a month to actually deliver

- A maxed-out $200/month Claude Max plan costs Anthropic somewhere around $8,000 a month

The entire model survives because most subscribers don't use anywhere near what they're paying for. It's the same logic as an all-you-can-eat buffet, profitable specifically because most people don't eat $50 worth of food. The difference is a buffet's marginal cost is a few extra prawns. AI's marginal cost is a fully-loaded server. And there will always be people who try to eat their value out of it.

This isn't just a theory anymore - it already moved markets

In early June 2026, this exact concern helped trigger the worst tech sell-off since the previous October, wiping hundreds of billions off AI-linked stocks in a single session, followed by a broader correction that erased over $1.3 trillion in semiconductor value within a week. Sequoia Capital's David Cahn has been tracking this gap for two years; what he first called AI's “$200 billion question” in 2023 has become the “$1.5 trillion question” in 2026: essentially, how much real revenue is actually needed to justify what's being spent building all this compute capacity, and whether AI companies can generate it fast enough.

To be fair, there's a genuine counter-argument

Two things are working in the industry's favour. First, the cost of running a model is falling fast, often cited around 10x a year for equivalent performance, so today's expensive query is tomorrow's cheap one. Second, pricing is already shifting: more AI companies are moving from flat subscriptions toward usage-based or outcome-based pricing, which is a direct admission that the flat $20-a-month model doesn't hold up long-term. Goldman Sachs and J.P. Morgan have both argued the spending is fundamentally justified by real demand, not just hype.

The takeaway

None of this means AI is a bubble that's about to pop overnight. But it does mean the “SaaS-style” valuations being applied to these companies assume a margin profile they haven't earned yet, and won't, unless one of two things happens: compute costs keep falling faster than usage grows, or the $20-a-month era quietly ends in favour of usage caps and higher prices. Worth remembering next time a headline calls a normal repricing a “bloodbath.”

Funded Futures Financial Services Socials