Benedict Evans asks who will actually make money from AI

Anthropic’s IPO papers offer a snapshot of the AI boom, but technology analyst Benedict Evans argues they cannot settle its biggest investment question: where the value will end up. His answer is less a forecast than a useful challenge to the idea that the model labs must capture the biggest rewards.

Watch Desk analysis

What happened

In a discussion reported by BigGo Finance, Evans points to Anthropic’s reported revenue growth and costs as signs of an unusual business model. He says inference costs have stayed roughly flat as revenue rose, while training costs have grown alongside revenue. He also notes that about 80% of Anthropic’s revenue comes from models built in the previous six months, suggesting customers are quick to move towards newer systems.

Evans argues that AI models could become infrastructure for products built by other companies, much as browsers and spreadsheets enabled businesses whose value did not accrue to their makers. The software firms, consultancies and startups building applications on top may capture more of the lasting returns than the labs themselves. His comments appeared in a BigGo Finance account of the discussion.

Why it matters

The argument shifts attention from how much AI gets used to who gets paid when it does. Model providers may face heavy training bills and rapid model turnover, while businesses closer to customers could turn the technology into products people keep buying. Evans also allows that some tasks will need the most capable, expensive models, while others may be handled well by older, cheaper ones.

That is a more useful investment question than picking a winner from a benchmark table. It asks whether the durable business sits in the model, the chips and data centres, or the software that makes AI useful in a particular job.

Our read

Evans makes a case, not a verdict. His comparison with earlier waves of computing is a reminder that the companies enabling a technology and the companies that profit from using it are not always the same. Investors and builders should watch where customers keep paying, rather than treating model demand as proof that model makers will own the value.

What to watch

• Whether inference costs continue to fall as AI use expands.
• Whether customers keep paying a premium for the newest models.
• Which software businesses turn AI capabilities into products with lasting demand.

Discussion spark: If AI becomes a routine layer in business software, should investors expect the model labs or the companies building on top of them to capture more of the value?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.