Tokenmaxxing versus Valuemaxxing
Notes from a room full of AI leaders who can't agree on where the spending curve goes next. Also: Anthropic nerfed Fable, Altman's ominous IPO message to OpenAI employees, and more.
In addition to my Monday evening of programming with Yahoo, I’m excited to announce that I’ll be participating in a panel titled “Attention, AI and the Future of Marketing” with Joe Marchese of Human Ventures and Bryson Gordon, Microsoft’s CVP of global marketing, at the OpenAI Villa during Cannes Lions. Let me know if you want to attend.
If you believe the recent press cycle, corporate America has decided that tokenmaxxing, or burning through AI tokens as a proxy for productivity, has run its course. After spending Tuesday afternoon with leaders from Databricks, Cognition, Pinecone, LangChain, and others at Nebius’ first Inflection forum in San Francisco, I can tell you there’s no such consensus among the people actually building this stuff.
During his opening remarks, Nebius CRO Marc Boroditsky made the pitch that the industry needs to move from tokenmaxxing to “valuemaxxing”: making tokens count rather than counting tokens. The leaders from other AI companies that Nebius put onstage didn’t exactly agree.
The spending split
During one of the early panels, The Information’s Amir Efrati set the stage by comparing this moment to the early public cloud era, when AWS customers were saying, “Wait a minute, I just spent $20 million more than I expected to this year.” Back then, he noted, $20 million was a lot of money.
From there, the panel split:
