Join Puck to listen to this article
The HumanX conference is only 2 years old, but it’s already become the quintessential A.I. conference for operators, investors, and executives to discuss the industry’s breakneck momentum—and its risks. As expected, the Moscone Center in San Francisco was pulsing with energy: gaudy booths, dancing robots, fog machines, etcetera. But in my dozens of conversations with industry insiders, I detected an undertone of anxiety that wasn’t quite so palpable last year. As a number of A.I. executives pointed out, the conference floor was buzzing with salespeople and investors. Largely absent, however, were potential customers.
In theory, enterprise procurement specialists are the most important attendees for startups looking to generate new business—and the mismatch between buyers and sellers points to deeper challenges for the industry. Some executives told me they worried about a tightening macroeconomic environment and increasingly cost-conscious customers. Others spoke with trepidation about how the market has become exceedingly competitive, especially as the dynamics of institutional lock-in take hold. Still others fretted about whether the frenzied venture capital environment may have obscured a lack of demand for their products.
These weren’t exactly idle concerns. One executive told me that if genuine enterprise adoption lags the hype, then the A.I. industry is on track to experience one of the most painful bubble deflations in recent history. But no one really knows what’ll happen next. As HumanX C.E.O. Stefan Weitz told me in the days leading up to the conference, “Certainty, at this point, is not a thing in this space. The velocity of change is so high. Who knows what’s going to happen tomorrow?”
Claude Mythology and P&L Magic
Bubble agita aside, there’s no question we’re in the midst of a transformative investment cycle: Enterprise A.I. spend has been ratcheting up since before 2024 and growing rapidly still. Just last week, Anthropic reported a surge in run-rate revenue, from $9 billion at the end of last year to $30 billion—an expansion that coincided with the reigning popularity of Claude Code, as well as the nice P.R. boost the company got from its clash with the Pentagon. But compute is expensive, and not all boats are rising at the same pace.
As Boomi C.E.O. Steve Lucas told me, all that money has to come from somewhere. “Our spend went up 10x from last year to this year, just on Anthropic,” he said. “Every C.E.O. I talk to—software, non-software—it’s the same.” (Indeed, several other executives also told me that their Anthropic bills had multiplied in recent months. Good work, Dario.) One reason costs have surged, Lucas pointed out, is that Anthropic has moved away from its unlimited-token pricing model and now charges enterprises twice: one subscription fee to gain access to the model, and another fee based on usage, billed at A.P.I. rates. As GitLab C.E.O. Bill Staples pointed out last week, “The $200 max plan can easily turn into $2,000/user/month (or more!).”
At some point, Lucas posited, this is all going to “blow up in [Anthropic’s] face.” Enterprises, he said, don’t know exactly what they’re burning tokens on, and aren’t able to distinguish R.O.I. on a per-token or even per-task basis. So even as enterprise spending continues to reorient around A.I.—often at the cost of other, mature categories, like cybersecurity—Lucas is convinced the music will eventually stop. “The chickens will come home to roost on this cost thing,” he said. “Unless OpenAI and Anthropic figure out how to make their services profoundly less expensive, they’re going to force everyone to use these open-weight models and intelligently route prompts to the lowest-cost model for the best outcome.”
Florian Douetteau, the C.E.O. of Dataiku, agreed that enterprises are getting more cost-conscious about A.I. and will eventually shift from running numerous experiments to focusing only on those that positively impact their bottom line. Arun Chandra, the C.O.O. of NiCE, said that this shift is already underway: Sooner or later, he told me, model developers will create a stronger bias for cashflow, which means they’ll pass those costs onto their enterprise customers. “So the R.O.I. has to be pretty clear, and the paths to R.O.I. have to be pretty clear,” Chandra said. “I think that’s the forcing function that is here and now. Those companies and partners that can show there’s a path are going to be the most successful. There’s nothing artificial about your P&L.”
Yahoo-ification
Silicon Valley is famous for proclaiming that things will never be the same, as PagerDuty C.E.O. Jennifer Tejada reminded me. Likewise, she said, the industry narrative surrounding A.I. is often oversimplified. She has large enterprise customers, for instance, that are still in the midst of a long-awaited transition to digital—only “20 percent of the way through their cloud transformations.” And with A.I., she said, “there’s still a big question” around what the high-value use cases are. Meanwhile, widespread A.I. tool adoption requires enterprises to both understand the risks and be able to mitigate them. “These kinds of transformations take a long time,” she said. “They move as fast as society can move, as fast as humans can move.”
This has put the industry in a bit of a bind. The frontier labs—even with their constant influx of venture capital and hyperscaler cash and credits—need adoption and enterprise lock-in to keep the lights on and the investor momentum rolling. But if adoption proceeds at the pace of past innovations and takes a decade or so, that crucial proof-of-R.O.I. likely won’t arrive in time for some companies. “I wouldn’t want to be Yahoo right now,” Lucas said, offering a pithy summation of the dilemma facing his peers. “One of them is Yahoo.”