Here’s our text for the day on the LLM/AI “boom”, which is now the center of the US economy and to a great degree the center of political power as well.
Four years into the AI boom, the companies selling it are cutting prices. Microsoft, Amazon, Workday and Figma are offering discounts and free access to hold on to customers drifting toward Anthropic and OpenAI, and the two labs cut prices on their newest models by as much as half.
The discounting has a plain explanation: EY says only one in 10 companies can show AI’s return on its income statement. The cost of building AI kept rising. Anthropic is negotiating a single data center lease that would require at least $40 billion, and Blackstone concedes no one has mapped out who will buy all the debt.
The takeaway: AI is getting cheaper to buy just as it becomes more expensive to build. How that gap closes, whether through vendor margins, credit markets or Anthropic’s coming IPO, will shape the boom’s next phase.
This comes in an email from The Information, the high dollar subscription publication covering Silicon Valley. As far as I can tell, this text is only in the email. It introduces a suite of articles expanding on the topics discussed. So I can’t link it.
In any case, falling prices for a product that is becoming more expensive to create is a formula which expresses lack of demand. Even my generally innumerate, non-economics-expert brain can understand that. This is the potential disconnect that has loomed over the AI boom from the start. It’s clear that LLMs can do big things and that there is demand and economic utility for those things. The question is whether there is enough demand, enough productivity gains for the companies and consumers who consume the product to justify the capital expenditures which are driving the LLM/AI boom and – this isn’t an exaggeration – sustaining most of the US economy.
That seems highly questionable.
It’s not a purely binary question. Most of us over 45 know that there was a major bubble and bust at the beginning of the internet era. But that didn’t mean the internet was a fad or a failure. Lots of start-ups went under and there was a big stock market slump. But the internet and tech powered ahead and were genuinely transformative for the whole US economy. The railroad boom was similar and even more chaotic and bumpy. The US economy is still massively fueled by railway infrastructure built in the final decades of the 19th century. (The bulk of other transport today runs on highway infrastructure build-outs in the 1950s and 1960s.) The physical rails and ties and sleepers have almost all been replaced over time. But the infrastructure is in the rights of way, the grading, the curvatures and embankments, the depots and cities built around them. For the US economy the railroads were a huge success. But there were massive boom and bust cycles and most of the concerns that built the railroads went bankrupt and were bought out by others who inherited the gains. Massive private sector infrastructure build outs almost always involve booms and busts and bubbles, even when they’re successful. We simply don’t know if AI will follow that pattern. What is important to remember, as we discussed last week, is that AI is fundamentally unproven in economic terms.
One more point to consider.
We are mostly thinking of data centers, which are really computing centers, as infrastructure for the LLM/AI economy. (They’re not big hard drives.) But there’s a lot of data to suggest that the lifespan of the chips that are the muscle of the data centers have a pretty short lifespan. That is both in the physical sense of when they stop working but also in the functional sense of when they might become obsolete. This is a complex question involving a lot of factors which are not only beyond my understanding but, I think, to a significant extent unknown. So see this not as declarations of fact but pointing to serious possibilities but also unknowns. In any case, if those lifespans are significantly short then this build out isn’t so much infrastructure – as in one time expenditures which yield longterm productive value on which economies are built – and more like fuel. And if that is the case the economics change a lot. And not in a good way.