The Problem With AI Valuations Nobody Is Talking About

Anthropic filed for its IPO on June 1 at $965B on roughly $47B in annualized revenue — about 21x. OpenAI filed confidentially targeting $730-850B on a ~$25B run-rate — roughly 30x. xAI raised its January Series E at $230B. Eighteen months ago these same companies were priced at 80-100x revenue; the revenue caught up, but the absolute numbers now assume trillion-dollar outcomes for everyone. This is not a market. This is a faith-based asset class.

The Revenue Multiple Problem

The last time multiples were this distorted was 2021, when SaaS companies traded at 40-60x ARR. We know what happened next: 70-80% valuation crashes across the board within 18 months. AI infrastructure companies today are trading at multiples that make those 2021 peaks look rational.

Consider the math: to justify a 10x forward multiple at its targeted $730-850B IPO valuation, OpenAI needs roughly $75-85B in revenue — triple its current ~$25B run-rate — in the next 5-7 years, while Anthropic has already passed it in revenue ($30B vs $25B run-rate as of April 2026) and free alternatives from Meta, Google, and the open-source community keep intensifying.

The difference between now and 2021? In 2021, the revenue was real — SaaS companies had 120%+ NRR, multi-year enterprise contracts, and true expansion revenue. Today, a significant portion of AI "ARR" is pilot revenue, month-to-month API usage, or heavily discounted design partnerships that don't reflect sustainable unit economics.

The Competitive Moat Question

I've done 65+ investments. The first question I ask every AI company is: "What's your moat?" The honest answer from most of them is: the model. That's a problem.

Foundation model pricing has dropped 100x in 18 months. GPT-4o API pricing is down 60% since launch. Claude's inference costs have fallen faster than almost any infrastructure input in tech history. That's great for application developers — terrible for any company whose primary value claim is "we use advanced AI."

When your moat is renting someone else's model and fine-tuning on top of it, you don't have a moat. You have a margin structure that compresses the moment the underlying API cuts prices — which they will, because their own investors demand volume growth above all else.

Red Flags Hiding in Plain Sight

• •Revenue that is pilot ARR, not contracted multi-year ARR — enterprise pilots cancel at 70-80% rates when budget cycles tighten
• •Valuation premiums driven entirely by brand affiliation with OpenAI, Anthropic, or Google — "backed by" is not a moat
• •Gross margins below 60% for software companies — AI inference costs are destroying the economics that made SaaS so attractive
• •Customer concentration above 30% in a single enterprise account — that's a business development deal, not a repeatable product
• •No disclosed net revenue retention — the single most important metric in software, conspicuously absent from most AI company pitch decks

What the Smart Money Is Actually Doing

The sophisticated LPs and family offices I talk to are not chasing the headline AI valuations. They're looking at three things: contracted revenue (not pilots), gross margin trajectory, and whether the company owns anything proprietary — data, distribution, or regulatory positioning — that a free model can't replicate.

The companies that will survive this correction are the ones charging for outcomes, not seats. Outcome-based pricing tied to measurable enterprise value — hours saved, revenue generated, errors prevented — is the only pricing model that holds when underlying model costs approach zero. Everything else is renting a house on someone else's land.

I'm not saying AI is overhyped as a technology. It isn't. I'm saying the current valuation environment is pricing in a winner-take-most outcome for dozens of companies simultaneously — and that is mathematically impossible. A repricing from 80x to 15x revenue still destroys 80% of investor value, even if the underlying technology is transformative.