AI bubble may have more in common with tulips than railroads

One of Australia's leading experts in AI reckons all the signs are there for a bubble, but said the eventual pop is unlikely to look much like the fallout that followed recent events including the telecom and dotcom collapses, or even the railroad frenzies of the 19th century. Instead, the coming pain might look more like the aftermath of the 17th century Dutch flower fiasco for investors - while yet leaving lasting benefits for humanity.

Professor Toby Walsh, in his calm and disarming manner, said a common theme in the technology space is playing out again as he applies his thoughts to where AI is going in the short term and long terms. Notably, "We overestimate the impact of a technology in the short term, and we underestimate its impact in the long term, and that is incredibly true of this moment in time," said Walsh. "The speculation and amount of investment going into artificial intelligence is appropriate for the longer term, but it's not appropriate for the short term. There will be some correction," he cautioned.

His observation is backed by relatively recent events including the telecoms and dotcom boom and bust cycles around the year 2000. But the echoes of past frenzies might include the Railway Mania in the 1840s and the subsequent Railroad Boom and Bust in the USA towards the end of the 19th century. And no discussion of bubbles is complete without reference to the 17th-century market craze where contract prices for tulip bulbs soared to extreme heights before collapsing in February 1637.

These and other historic economic waves were generally driven by various combinations of overspeculation, heavy debt financing, and rapid technological excitement triggering major financial panics. FOMO, too, played its part.

"The thing that's focusing a lot of the public's attention is all the investment going into AI data centres," noted Walsh. "The estimate is that there's going to be something like $7 trillion being invested before 2030. To see a return on that investment is hard to understand, given that it's not like the investments that were made in previous technologies."

Instead, this particular bubble may have more in common with that most notorious example of human folly and FOMO greed. When the wipeout came, little of enduring value remained in the tulip market, beyond aesthetics.

Walsh went on to explain that the $2 trillion invested in fiber optics in the late 90s saw the book value of US telecoms companies falling by roughly the same amount. But all was not lost, because the physical infrastructure eventually proved essential. "Most of that fiber was dark for the first 10 years. Today, it is being used to screen Netflix and all the other stuff that we do, so it proved to be a reasonable, long-term investment."

AI, he said, is going to have a wider impact, but probably not one that justifies the pace and extend of the buildout. "It will take longer to happen than the investment market is predicting; people change more slowly than technology," - as demonstrated starkly by the dotcom bust, where the impact of the internet was ultimately overestimated short term, and under in the long.

Right now, he sees warning signs in AI spending. "The only comparable point in history is the great railway buildout. It's worth looking at history, where lots of investors lost everything but a few winners came out of it very well. Of course, we got the railways which allowed the development of the U.S. economy and the marketplace that happened there. The same, I'm sure, will be true of spending on data centres, with one proviso. Railways can last 100 years. That's not true of data centres."

Or, indeed, of tulips. When the perceived value of tulip bulbs evaporated, so did their prices.

"So they're unlike, and that's a good third of the cost of a data centre is the GPUs, and that third of your asset is going to have to be replaced in five or 10 years, at a significant depreciation cost."

Competition, always a powerful and welcome force driving down prices for customers, doesn't work in the favour of those making large investments with questionable paths to returns, added Walsh. Essentially, he said that as companies quickly face the reality of tokenomics, that most powerful of invisible forces is driving discernment and a shift to 'right sized' LLMs. Good enough, especially at a price, is better than perfect.

"Companies are waking up to the idea that you can use a less capable model to answer mundane queries…and that there's a very competitive marketplace out there [with, for example] open weight Chinese and other models that run on your own hardware."

One only need examine the previously prohibitive costs of most technologies and their downward trajectory over time to clearly see Walsh's point. "Technical advances mean we can do the same stuff with cheaper and smaller models, and there's an open model trend of giving away the models so people can run them themselves. So, this is good for the planet [meaning people]. But not very good if you're trying to justify a trillion dollar valuation for an upcoming IPO."

Walsh believes, with the backing of a good deal of precedent, that AI is here to stay on the one hand, and will become a commodity. "You're going to get cheaper and cheaper, more and more competitive. The models are going to become a commodity, a general-purpose technology. And like other general-purpose technologies before it, like electricity, like the internet, like the steam engine. These were general-purpose technologies that were pervasive in our society; no other GPT was a monopoly...and the price will always be going down."

That's certainly unfortunate for those shelling out billions and trillions to get the AI ball rolling. But where the AI rollout differs from tulips and still has something in common with railroad and fibre networks, is an enduring benefit even if conceptual rather than physical.