View / The mistrust of AI labs bubbles over

A few weeks ago, an Anthropic executive was asked a question on every CEO’s mind: Why should we trust that you won’t steal our business?

“It’s a fair question,” Syed Mohiuddin, Anthropic’s head of healthcare, said at a venture-capital event I attended last month on Cape Cod. “Because we are both a frontier lab that’s building models and a product company that has applications.”

The edgy mistrust between companies and the frontier labs they are deeply in bed with is bubbling over into a fury channeled by Palantir’s Alex Karp in a chaotic appearance on CNBC last week. “These people are livid,” he said. “They’re like, ‘I am paying for tokens that create no value… and they’re going to get my IP.’”

Forward-deployed engineers, dispatched by AI labs to help customers implement their models, will come back to the mother ship with a deep understanding of how banks, consulting firms, retailers, manufacturers, and consultants operate. That customer support, to a suspicious eye, looks a lot like reconnaissance.

Anthropic has already launched products for law firms and design firms; the primary losers there are software companies that hawk those services, like Harvey (law) and Figma (design). But what’s to stop the labs, after ingesting enough data, from launching an operating business to compete directly with law firms and design firms?

At the same time, open-source models — many Chinese, but also a growing roster of American alternatives like Reflection — are rapidly closing the capability gap at a fraction of the cost. They can be downloaded and run locally without sending data back to the labs.

Companies have to use AI labs’ products “to stay in the race, but doing so requires sending all of their IP, and that is a very uncomfortable thing to do with somebody who might be trying to replace you,” Will Wilson, CEO of code-debugging startup Antithesis, told me. (Wilson said he’s “not scared of Anthropic” and is loving Fable 5.)

To be clear, the AI labs say they will not do this; here is OpenAI’s terms of service for enterprise customers, and here is Claude’s. But they trained on the internet’s content without asking permission and have a credibility problem.

“Claude Code didn’t kill Replit and Cursor,” Anthropic’s Mohiuddin said last month, referring to AI-native coding apps that have flourished in the world Anthropic built. “We’re not trying to be kingmakers and we’re not trying to be market replacers. What we’re trying to do is elevate the floor of what’s possible.”

The real question isn’t whether AI labs can move downstream and compete directly with their customers. It’s whether they can make more money by helping companies like Sullivan & Cromwell draft legal contracts, Gensler draw up architectural plans, and Pfizer discover new drugs — or by simply doing all that themselves.