Gary Marcus：生成式AI泡沫正在破裂

Since August of 2023, I have been warning that, financially, generative AI might turn out to be a dud.

As I outlined in the very next article after that was that the problem with generative AI was that everyone was following the same path, using very similar techniques, training on very similar datasets. I warned that as a consequence OpenAI (which was then flying high) would not be able to keep its lead; that nobody would have much of a technical moat, and that there would be price wars, and the profits would be hard to eke out. Three years later, prices are lower, competition is higher, and profits remain elusive.

Training is ridiculously expensive and getting more expensive. The Claude 5.5 model family alone reportedly cost $10 billion to develop. And inference is expensive too. That leaves very little room for profits, especially in such a competitive market.

I also warned that LLMs lacked reliability; a raft of studies showing a lack of ROI seems to bear that out. There is an argument that things are improving, but even if they do they may not improve fast enough to justify the trillions being invested.

Meanwhile, nobody will be surprised if the massive Microsoft outage stems from code that was written by generative AI. And the systems remain poor at following instructions. The Klarna effect, in which companies rehire humans after making extravagant claims around AI, is yet another sign that GenAI is still not as good as it was cracked up to be. (Some future of AI may be better, of course, but we are not there yet).

All that said, I didn’t get everything right. In particular hypey valuations persisted a lot longer than I expected. I am really surprised that the dance lasted as long as it did.

But it does look now like we may finally be near the end.

The Huang Affair and the Blue Wave have led to some pretty dramatic changes in fortunes. Rockets like Nvidia and Coreweave and Oracle which used to gain every month, are all losing steam; Anthropic was hoping for a $2T valuation at IPO and will now count themselves lucky if they get $1.5T.

Nvidia is down 29% over the last six months, Oracle 35%, CoreWeave 54%, and Anthropic’s perceived valuation is down 25% in just a matter of weeks.

People may not want to admit it, but those numbers sure look like a bubble popping to me.

And that downward slide is continuing today.

§

In a wild moment last night, I was live on CNBC Fast Money, warning that the whole thing might fall apart, in light of Nvidia’s troubles, because so many companies were reliant on Nvidia’s financing. Suddenly my conversation with the host, Melissa Lee, was interrupted by a breaking news bulletin: a neocloud company called N-Scale was in deep doo-doo, having overcommitted.

We may soon see a lot more moments like that.

§

Psychologically I always said that the real issue around timing the deflation wasn’t the economics, which obviously never made sense, but the psychology of it all. When would the world, like Wile E. Coyote, realize that (economically speaking) generative AI was over a cliff? I couldn’t really say.

My sense, which could once again be wrong, is that the market has finally picked up on the irrationality of generative AI’s economics. And of course the change in regime at Nvidia may have massive repercussions.

So much was resting on Jensen playing YOLO (you only live once) with truly massive bets on the companies that Nvidia was effectively subsidizing. But that may be coming to an end.

In an article yesterday, that’s what The Wall Street Journal seemed to be hinting:

If I read the essay correctly, the interim CEO is less likely to be inclined to make such big bets. She might, for example, refuse to prop up Nscale, and let nature take its course (whereas Jensen might have stepped in).

If Nvidia stops acting like everybody’s banker, a bunch of neocloud companies are in trouble. Anthropic’s IPO could be jeopardized, too.

“Planned.” Will they still do so? Or will Colette Kress cut back the allocation, as she tries to right the ship?

It’s already starting to get bad:

§

The grim fact is that companies like OpenAI, Anthropic, Nscale, and CoreWeave have never been demonstrably profitable (aside from a single tokenmaxxing-mad quarter in which Anthropic received heavy subsidies from Musk). And it’s never been clear that they can stand on their own two feet.

Now they may have to, and that may not be pretty.

Subscribe now