Today the Wall Street journal published an article (paywall) asking if the AI-boom data center mania might be inflating a bubble along the lines of the “internet’s infrastructure build-out in the late 1990s”. Well, gee, ya think?
I lived through that bubble. Right when I was graduating college, the Internet (I still use the capital “I”) was a new thing and the dotcom boom was taking off. My second job in IT was at a computer consulting company in San Francisco, the epicenter of the dotcom mania. The consulting company specialized in advertising agencies. While Madison Avenue in New York may be most famous for advertising, SF has a healthyr ad business too.
I saw the bubble firsthand. Venture Capitalists (VCs) were throwing cash at these new dotcoms, many of them sketchy. The dotcoms would then pump their money into advertising. In turn, the ad agencies expanded rapidly. Since they needed infrastructure for their expansion, they hired us. We were a general IT shop, but we came to specialize in moving ad agencies from small offices to larger digs.
The entire thing had a Ponzi-scheme sort of feel. Who advertised on dotcom web sites? Other dotcoms, of course. By 1999 I was feeling uneasy about the dotconomy, as I used to call it (TM), and I kept what money I had in cold hard cash, and avoided the markets. I advised others to do the same. Eventually I got a job running the network at the local newspaper. I figured (correctly) that it would be a good place to weather the coming storm, as I figured (incorrectly) that the daily paper would be around forever.
We all know what happened after that. One of my friends went from nerd to nerd chic. He was profiled in Details magazine in a $400 designer hoodie with his collection of vintage pin-ball machines in his costly Seattle flat. His millions of dollars (on paper) were in stock options, which are taxed when you exercise them. He exercised them when the stock for his pre-Google search company was at $125/share. He got a tax bill for the income at $125/share. He wasn’t able to pay the tax bill when the stock went to pennies/share. If only he had sold.
I went to a bunch of dot com auctions when the crash happened. These were a blast. Back then, starting up a company required a significant capital outlay. Dotcoms had to buy servers and network equipment and then rack and stack the gear in colo facilities. There was no AWS at the time. When they went under, they’d open their offices up. There you could pick up Cat 6500s, Lightstream ATM switches, servers, and Herman Miller Aeron chairs for a bargain. A lot of folks built CCIE labs out of doctom auctions.
The other problem some of these companies had was exactly how much to build. Webvan was the classic example. They had a good idea (online grocery delivery), but how many warehouses to build, and where? They operated on an “if-you-build-it-they-will-come” model. Only they built it and no one came. (Apologies to the 1980’s baseball movie reference for my non-American readers.)
I expected something similar with the smartphone/app craze, but it never happened. I saw the same crazy overvaluations. However, the difference in my non-finance-guy opinion was the lack of capital expenditures to create these companies. If you needed server power for your app, instead of buying a pile of HP servers you would just host your app in Amazon. Scaling up and down became simple. And, with remote work, you don’t even need the Aeron chairs.
AI is different. Behind the magic of LLMs is computing power, and lots of it. AI requires massive capital outlays, beyond what dotcoms needed, and it requires massive building projects done way in advance of their use. It requires speculation as to demand. Like Webvan, the data center builders are assuming if they build the infrastructure, it’ll get used.
The problem is, the utility of AI is so unknown. I’ve found it does a great job as a search engine, because I can feed it very specific queries with nuance that gets lost in basic Google queries. Its facility for language is nothing short of amazing. But beyond that, nobody really knows if anyone will make any money from it.
Aside from disturbing episodes where AI coaches people into suicide or advocates for genocide, Harvard business review has called into question the quality of AI’s output as a productivity tool. Calling AI-generated output “workslop”, Harvard says “few are seeing it create real value”. Employees “use it to create content that is actually unhelpful, incomplete, or missing crucial context about the project at hand.” This generates more work downstream by “requiring the receiver to interpret, correct, or redo the work.” (For the record, AI generates zero percent of the content on this blog, except when directly quoted.)
For us network engineers, AI can produce some flashy demos. You can speak into your microphone and say “go configure loopback zero on router R1” and, by golly, it happens. How useful is this, really? I can type “int lo0” pretty fast. I’m not convinced it’s as awesome as promoters seem to think.
Technology, of course, improves. Today’s limitations might be conquered tomorrow. As a technologist (I hate that word) I know that there is danger in being too skeptical of new technology. However, technology also plateaus, and reaches a point where improvements become incremental instead of drastic. My gut (which could be wrong) tells me we are at that point with AI.
If so, and if the business case doesn’t work out, then we may be at a dotcom level of bubble. But hey, if you’re working on your NVIDIA certification, maybe you’ll be able to get some cheap gear.