Rent the Model, Own the Workflow: Valuing AI Without a Comparable
In late May 2026, the leading AI labs were valued at roughly 20 to 35 times sales, according to Research Affiliates - Anthropic at the low end, OpenAI near the top. None of those figures come from audited accounts.
That uncertainty doesn't stay with investors. Every company buying, building or budgeting for AI is betting on the same unknown: what a model is worth, and for how long.
What exactly is the asset?
A model isn't a patent. Flagship models typically become economically obsolete within 12 months. The model still works. Its scarcity value doesn't. Unlike classic software, models keep changing after launch as users interact with them.
So what's being valued - this year's model, the ability to build the next one, the data behind it, or the distribution that sells it? Businesses increasingly put a separate value on proprietary datasets, weighing uniqueness, quality, scalability and regulatory limits. Specialists describe a "valuation vacuum": the standard tools assume stable assets, and AI isn't one.
Who keeps the margin?
Even if a third of revenue became after-tax profit, 20 to 35 times sales means about 61 to 106 times earnings. That only works if margins hold for years.
They're squeezed from three sides. From below, some open-weight models are approaching frontier performance at much lower cost per token. From above, suppliers hold the scarce inputs - NVIDIA's near-75% gross margin is a line in a lab's cost of revenue. From outside, regulation can widen or narrow the moat by decree.
Spending runs ahead of returns. The FACT Coalition notes hyperscalers keep pouring money into data centers for a technology that has so far largely failed to produce meaningful profit.
What it means if you're the customer
Enterprises are learning to route tasks between providers by price and performance. Once a business owns the user-facing application, the model underneath starts to look like a commodity.
The counter-case is lock-in through workflows. When a model is wired into compliance reviews, developer tools and security processes, customers stop buying tokens and start buying reliability and trust. Re-platforming isn't cheap.
Practical read: treat model spend as rented capability, and put the durable investment into the data, workflows and integration you own.
Research Affiliates frames the investor question as whether you're buying a durable stock of profits or a transient flow of competitive advantage. Every AI adopter faces a smaller version of that bet. Which part of your AI stack will still be worth something when the next model ships?
Sources: Research Affiliates, "Moat or Mirage?" (July 2026); Mondaq reporting (Apr 2026); Financier Worldwide, Grant Thornton roundtable (Nov 2026); FACT Coalition (Aug 3, 2026).