“We had a lot of new ideas and really cool stuff that [Meta] should implement. But they were just going for things that were essentially safe and proved. When you do this, you fall behind. A lot of people have left, a lot of people who haven’t yet left will leave. I’m sure there’s a lot of people at Meta, including perhaps Alex [Wang], who would like me to not tell the world that L.L.M.s basically are a dead end when it comes to superintelligence. But I’m not gonna change my mind because some dude thinks I’m wrong. I’m not wrong.” —Yann LeCun, the industry titan and former chief scientist at Meta, opening up about some of the circumstances surrounding his recent exit… while pitching his own venture, a company called Advanced Machine Intelligence.

And now for the main event…

For as long as the A.I. industry has existed in its current hyperscale mania, I’ve been reading nearly identical surveys—whether they’re time-stamped from 2023, 2024, or 2025—that tend to reveal the same conclusion: While the C-suite class is euphoric about the economic promise of the technology, the tech isn’t quite delivering on the hype. In fact, it’s not even close.

In just the last couple of months, there was an M.I.T. report uncovering that 95 percent of enterprise A.I. pilot programs fail; an ISG report stating that R.O.I. remains “uneven” for enterprises experimenting with A.I.; and a McKinsey report from the end of the year basically saying the same thing. “The gold is even more elusive than perhaps was initially thought,” Dr. Seena Rejal, the veteran tech executive and chief commercial officer at the British A.I. company NetMind, told me.

Yes, we’re still at the dawn of a tectonic platform shift, and no one ever expected that hundreds of billions in investment would pay off overnight. But, as the calendar year turns, this burgeoning industry is under pressure to produce practical value that’s at least partly commensurate with its astonishing promise. As former Meta chief scientist Yann LeCun recently noted, and Sam Altman undeniably knows, L.L.M.s are cool and all, but they aren’t necessarily a transformative business opportunity by themselves. Indeed, the true enterprise value of A.I. has hardly been unlocked, or even really fully conceptualized. “No one ever promised this would be easy. I think ChatGPT and others probably made it appear that way, and therefore they gave us false hope that it's as easy as just talking to your phone. But the reality was much more complicated than that,” Rejal said. “But humans are humans. We do this every time. We never learn. We go through these hype cycles.”

Rejal, a former venture capitalist who also founded Shapes AI, described to me the yawning chasm between technological advancements and genuine business applications—in other words, the gap between the technology and making it work. “We’re probably oversaturated in terms of really exciting developments on the technical A.I. front that we don’t have much use for just yet,” he said. “The gap between research and business needs to close in the following year to make this whole cycle sustainable, and that readjustment will take place. It’s just a natural leveling out of things.”

Given the extraordinary investment in the space, the technical advancements will likely keep coming, however incremental they may be. Meanwhile, Rejal expects, or perhaps hopes, that significantly positive business outcomes will become apparent this year. Practical and commercial success, he said, is becoming increasingly urgent and even existential for many companies. “People are speculating, and they’re investing, and they’re doubling down, because they think there’s a gold rush,” Rejal said. “But unless that gold actually gets dug up somewhere, this whole thing might end up with lots of spades having been sold and no gold found. If that gap isn’t filled or closed quickly enough, the equation won’t make sense anymore.”

And if that happens, corporate disillusionment with the industry could diminish investment outside of a protected class of first movers and send plenty of A.I. startups to the cleaners. “I think it desperately needs this kind of all-encompassing infiltration or penetration of legacy markets and legacy sectors,” Rejal told me when we spoke in December. “I think that can happen. Next year is when things get real.”

Michael Bachman, the head of research at Boomi, an enterprise A.I. integration company, took things a step further when we spoke recently. He explained that markets are “forcing value on to why A.I. is even being used in the first place. We’re starting to see this now.” Clients, he said, are increasingly interested not only in cost savings but in actual dollar generation associated with the application of the technology. Thus far, alas, that kind of investment return has been largely elusive.

At this point, there have been two major A.I. winters. The first occurred in the late ’70s and the second in the ’90s, and both periods resulted from times of peak investment that failed to meet expectations. Bachman told me that he’s lately been thinking a lot about when the next A.I. winter will strike. Over the next year or two, he told me, he thinks we’ll experience something more akin to a “frost”—essentially a “more condensed winter.” At some point soon, he posited, the industry will suffer the results of a dearth of chips and power, something that he expects will lead to a market pullback complete with all the normal signs of a correction: the death of startups, and plenty of consolidation.

As M&A activity picks up, the valuations of the major A.I. startups (and of many minor players as well) will likely continue to soar, even while profit remains elusive. Meanwhile, dreams of hyperscale are already running up against energy constraints and growing community resistance to data center construction. Indeed, Lin Qiao, the C.E.O. of Fireworks AI, expects adoption to “outpace infrastructure build-outs next year. Efficiency won't just be about cost savings anymore,” she told me. “Under capacity constraints, it will become the opportunity cost of growth.” She added that the race to secure users will be won by the companies that can scale despite constraints in chips and power. It’s an impression shared by Natalie Hwang, the founding managing partner at Apeira Capital, who told me that the biggest constraints in A.I. no longer have to do with capability but rather power and infrastructure. “The most important development for 2026 is that A.I. economics are finally colliding with physics,” she told me. “This shift is pushing innovation beyond model size and toward efficiency, architecture, and deployment economics. That transition will define which companies are able to create durable value in the next cycle.”

If 2024 was the year of the chatbot, and 2025 the year of the agent, then 2026 is already shaping up to be the year when technical advancements and iterations begin to matter a hell of a lot less than making the tech we have today work reliably, safely, and at scale. Three long years into the A.I. boom, we’re approaching a self-evident inflection point. “Any advanced technology can be indistinguishable from magic, and I think the marketers that have been selling [A.I.] as magic… those chickens are starting to come home to roost right now,” Bachman said.

For what it’s worth, though, he isn’t all that bummed about a looming correction. In many ways, the oscillations in the A.I. cycle mimic the dialectic between the caution of the scientists who helped advance this technology from the blackwater and the investors and capitalist maximalists who plan to convert it into billions of dollars in value. A frost “is going to create good research,” Bachman added. “And I think it’s going to give us a chance to recalibrate, as a society, what we want out of these systems. Because once we start understanding that these things are here to stay, and they’re probably going to be governing our lives, or used to govern our lives… we’re probably going to need to have some dramatic policy changes all across the world. It’s going to give us a chance to pause.”

That’s all for today. I’ll see you on Thursday.

Ian

Need help? Review our FAQ page or contact us for assistance. For brand partnerships, email ads@puck.news.

You received this email because you signed up to receive emails from Puck, or as part of your Puck account associated with {{customer.email}}. To stop receiving this newsletter and/or manage all your email preferences, click here.

Puck is published by Heat Media LLC. 107 Greenwich St., New York, NY 10006