The tech industry boarded the AI “crazy train” this summer, and it shows no signs of slowing down.
The latest fuel for this boom comes from Wall Street. Some of these projects are now being financed by elaborate borrowing methods and unusual, circular deals.
In these moments, I lean on Dakin Campbell, a BI reporter who’s covered Wall Street for almost two decades. He wrote a great story about this AI financing frenzy, so I asked him to weigh in here:
Alistair Barr: Structured credit is showing up in AI infrastructure financing. Does that worry you?
Dakin Campbell: The easy answer is we have seen this movie before. Structured credit isn’t fundamentally dangerous. But it does distribute risk throughout the system in a way that makes it harder to see and track and understand. And yes, that does worry me. It makes the job harder for investors, regulators, journalists, and others who act as a natural counterbalance against excess.
Do founders such as Mark Zuckerberg and Sam Altman care about investors’ return on investment, or just winning the AI race?
At some level, I do believe that Zuckerberg and Altman and others believe that there is money to be made. They think it will become a profitable, a very profitable business, at some point in the future. I do think their egos are involved in the belief that they could be the ones to usher in AGI and become the legends of history. These are men who read science fiction books when they were kids. I don’t think we can overlook that aspect.
Related stories
Business Insider tells the innovative stories you want to know
Business Insider tells the innovative stories you want to know
Are railroads a fair analogy for the AI buildout — big losses at first, lasting assets later?
Railroad tracks and locomotives are long-lived assets, whereas that is not true of GPUs.
Tech blogger Paul Kedrosky, who I quoted in my piece, says about 60% of the cost of data centers is the GPUs. You can argue whether their depreciable life is three years or six years, but whatever that is, it is considerably shorter than railroad assets. There is of course the shell of the data center and the cooling and electrical infrastructure for the building but if Kedrosky is right, that means less than half of the spending is going into an asset that could reasonably be considered long dated infrastructure, like the railroads.
The analogy to fiber overbuilding in the first dot-com boom is also instructive. Fiber networks last longer than GPUs.
Can inference demand sustain this AI infrastructure boom, or will it depend on real-world AI products?
Inference is the process of getting AI models to deliver answers for users, which one would argue is the same goal as real-world end products.
At some point, the industry will need to figure out how to design products with repeatable outcomes, based on AI that they can sell to corporations and consumers. You’re already seeing people arguing that we should stop reaching for AGI or superintelligence and instead focus on using AI, today, to solve real-world problems. And you hear from smart AI researchers who say we are still several advancements from achieving AGI.
So when I think about it like that, yes, at some point, it feels like the markets or investors or public perception will force these companies and these CEOs to focus less on AGI and more on solving real-world problems.
Have you personally seen any real value from generative AI?
Yes, I have. I have friends who love Grammarly for helping to improve and copy edit their writing. I like asking other models research questions, which I find helpful for ideation and my own thinking.
But when I ask it to solve problems for me in a repeatable fashion, or to look at a pile of documents and come up with an answer that strictly adheres to the information in the documents, it fails pretty miserably.
So I do see the potential, and I am not sitting here calling this all nonsense, but when I talk to people about AI, the answer I get back again and again is that they would like AI to be able to help them solve problems easily and repeatably, without having to know the exact words to use in the prompt. It doesn’t feel like we’re there yet.
Sign up for BI’s Tech Memo newsletter here. Reach out to me via email at abarr@businessinsider.com.
