When I saw this note from Torsten Slok, the chief economist for Apollo Capital, I knew I had my topic for the week. The point is that the big money in AI is far removed from the end product. The chipmakers are making money hand over fist, the energy providers are doing okay, the hyperscalers have less to show, and the AI companies are losing bucks big time.
This matters, because at the end of the day, if the AI companies are not making money, the whole thing breaks down. To use a common analogy, suppose that steel companies are making huge bucks producing steel for rails, and construction companies are making money laying the rail, but the companies that run the railroads are all going broke. That doesn’t look like a story of long-term prosperity. In the great minds think alike category, Ed Zitron jumped on the same point in his excellent newsletter.
Anyhow, I take a somewhat different tack than Ed and focus on the Chinese competition. I realize that even if there was no competition from China, it is unlikely that AI would ever have the massive payoffs the hyperscalers are banking on, but the existence of that competition makes the story considerably less likely. And developments in the last couple of weeks seem to make the case for American AI even weaker.
As I have frequently noted here in the past, Chinese AI costs far less per input or output token than U.S. AI. For the cutting-edge models, the Chinese AI sells for one-fifth or even one-tenth of the price of U.S. AI. One response I have seen is that, even though the Chinese AI costs less per token, it can still end up being more costly because the systems are less efficient and require more tokens per task.
I am not sure that the measure of cost per task is a sufficiently standardized metric to allow it to be compared in a meaningful way, but insofar as it can be, it looks like the U.S. advantage has gone away. According to the Korean electronics industry publication, The Elec, the leading Chinese AI model is now cheaper on cost per task than the leading U.S. model, and performance gaps continue to narrow.
In the same vein, both Google and DeepSeek released new flash models last week. The DeepSeek model scored better on several benchmarks. And it sells for less than one-tenth the price.
If that makes the picture look bleak for U.S. AI producers, don’t worry, it will likely get worse. Alibaba reports having developed a modular design that will allow it to build data centers in 100 days. Compare to 12-18 months in the United States. This should mean lower costs and greater capacity for Chinese AI producers. That means the flood of low-cost, high-quality Chinese AI is likely to get even larger in the months ahead. Given its huge cost advantage, it’s not surprising that Chinese AI models are gaining ground rapidly at the expense of U.S. models. I’ve noted before that Chinese AI seems to be winning out by large margins in most regions of the developing world. However, it also seems to be gaining ground in Europe. There are political considerations that could make European companies reluctant to rely on Chinese AI; however, given the erratic behavior of Donald Trump, it’s not clear that going with U.S. provides greater security.
And it looks like Chinese AI is continuing to gain ground in the U.S. market. It seems that Apple is looking to Chinese AI as a cheaper alternative to the Silicon Valley producers. Apple by itself is potentially a huge market, but perhaps more importantly, it is a company that has been at the cutting-edge of innovative technology for more than a quarter century. Its decision to go with Chinese AI is sending a serious message.
And remember, the question for those expecting really big bucks for the AI makers is not just whether Anthropic, OpenAI, and the rest can hang onto a large share of the market. It’s whether they can do so while selling at prices that give them the huge profits the stock market is banking on.
As mortgage issuers sold ever more dubious mortgages to further inflate the housing bubble, the wizards of Wall Street assured us that their financial magic would make it all work. This attitude was best conveyed by former Treasury Secretary Larry Summers at an academic conference in 2005, where he dubbed a critic of the growing house of cards a “financial Luddite.” Somehow, Summers thought innovative finance would make the millions of underwater mortgages issued to people with weak employment prospects and no reserve assets all work out fine.
We might be getting the same story with the AI bubble. Getting back to Torsten Slok’s point about the chipmakers making big bucks, while AI producers are making big losses, it seems Nvidia is looking to address the problem. It has just arranged $500 billion in financing from major banks for the hyperscalers that buy its chips. Fans of markets everywhere are asking the obvious question: if there is so much money to be made in building the data centers, why does Nvidia have to arrange the financing?
The details are not clear at this point, like whether Nvidia will in any way be on the hook for the financing, but there is a suggestion that it could involve securitization with tranches carrying different levels of risk, sort of like mortgage-backed securities or collateralized debt obligations. It could be lots of fun!
*Dean Baker is a senior economist at the Center for Economic and Policy Research and the author of the 2016 ** book *Rigged: How Globalization and the Rules of the Modern Economy Were Structured to Make the Rich Richer