I had a front-row seat to that era while at Mail.com/EasyLink (later acquired by OpenText). Watching WorldCom implode, IBM scramble, and the telecom infrastructure unravel taught me something I still believe today: technology revolutions are real, but valuations often outrun business fundamentals. The hard part isn’t predicting the technology. It’s predicting which business models will actually survive once the hype fades.
Very interesting. What caused the steep decline in manufacturing productivity in the 1970s? Did the OPEC oil embargo play a role? It just made me wonder what other price shocks and global uncertainties today might end up shaping the current AI bubble as it plays out…
Galloway is right that AI is entering a new phase—but I think he’s mistaking maturation for collapse.
The strongest part of his argument isn’t “AI is another Pets.com.” It’s that enterprises are shifting from “deploy AI everywhere” to “prove ROI.” Uber blowing through budgets, Meta and Salesforce emphasizing proven use cases, and tighter governance are real signals.
But that’s optimization, not necessarily demand destruction.
The missing link is evidence that this translates into lower infrastructure demand. To make the 1999 analogy work, you’d expect to see slowing frontier-model revenue, hyperscalers cutting GPU orders, HBM inventories building, packaging utilization falling, and sustained weakness in AI infrastructure spending. We’re not there.
In fact, hyperscalers are still committing record capital while the real bottlenecks remain HBM, advanced packaging, power, and transmission—not excess compute. Unlike telecom in 2001, supply is still constrained.
Galloway is right to ask whether lower token prices automatically translate into higher infrastructure demand. So far, the evidence largely says yes: falling inference costs have been accompanied by rising usage, stronger AI revenues, and record hyperscaler capex. Jevons isn’t a mathematical law, but the burden of proof now sits with those arguing that enterprise cost controls will overwhelm the powerful demand created by cheaper intelligence.
The more interesting debate isn’t whether AI is a bubble. It’s which layer captures the economics. Frontier labs? Hyperscalers? Semiconductor suppliers? Enterprise software? Or end users?
My base case remains that private AI valuations compress, enterprises become more disciplined, and capital shifts toward measurable ROI. That is very different from concluding AI infrastructure demand collapses. The analogy to 1999 skips a critical difference: today’s largest AI investors aren’t Pets.com—they’re some of the most cash-generative companies in history, operating in an environment where key components remain supply constrained.
The evidence increasingly points to a transition from indiscriminate AI spending toward ROI-driven deployment—not yet to a telecom-style unwind. If that changes, the signals to watch aren’t anecdotes about budget discipline. They’re slowing frontier-model revenue, hyperscaler capex cuts, weakening HBM demand, easing packaging constraints, and sustained excess GPU supply. Those are the dominoes that would validate Galloway’s thesis. Until then, the stronger evidence supports normalization, not collapse. Hat trick AI
I, too, have seen this movie before. I was a top-ranked internet analyst in the 1990's at a firm called Alex. Brown (then Deutsche Banc). I was the analyst who took Amazon, eBay, and other companies public, and I covered many more. I passed on the pets.com IPO, along with 50 others. I even wrote and taught a case study about it before the crash! 67% of the companies we backed are either still in business today or were successfully acquired by other companies. ALL of these companies' IPOs would now be considered VC A or B rounds, definitely not IPOs.
In the end, the dot-com bubble was just a blip, a healthy flushing out of the marginal players. Those who were in it for the long game not only survived, they became massive winners. The likes of Bezos were never building companies for the 1999 market; they were building companies for the 2009-and-beyond market they knew was coming. Staying power was the key, and the cost of capital was higher then than it is today on a relative basis.
There are some notable differences in today's AI economy (I'm not saying "it's different this time", but there are differences): (1) most of the funding for the LLMs is coming from private, not public markets, with the exception of the big players who already generate massive cash flows. This gives companies time to figure things out without the capricious nature of public markets driving their strategies; (2) the ease of adoption, and thus diffusion across markets, is much easier and faster than the Web was in 1999; (3) the war chests among the leaders are huge, and so are their backers.
AND, yes, I agree, none of the LLM players have put out a convincing financial story, especially when the circular financing is taken into account. This is the fly in their ointment (though I don't think they see it). The social hazard arguments are mute; there are enough people becoming reliant on AI at work and in personal endeavors that it will continue to grow... planet be damned! Which is why I am finding it really hard to use AI in my own work!!!
Warren Buffett was right when he said: "trees don't grow to the sky"!
Having lived through the dot-com bubble and bust, I can definitely see the parallels. What we know is that the greatest value companies are built on the ashes of the “visionaries.” Wake me up when we cross the chasm, I’ll be more interested then.
Oh, I definitely think it’s a bubble, but I think we’re still on the early side of this bubble inflation. There’s an enormous about of money committed to building out the infrastructure, so that money will be flowing to the chip companies (and related businesses) for at least a couple of years.
I have a question about this FT quote below, I read the linked article too but didnt find an answer. How do we know productivity is truly exploding at OpenAI and Anthropic?
“The fact that incumbent software and knowledge work companies are finding only modest productivity gains by incorporating AI into existing workflows and organizational structures, while usage, revenue and productivity explode at Anthropic and OpenAI...”
The strongest parallel may be less 1999 valuations than who captures the productivity surplus after the shakeout. If models commoditize and enterprise buyers move from token volume to completed-task economics, value can migrate simultaneously to cheaper open-weight models, orchestration layers, and end users—while some infrastructure vendors still benefit from higher aggregate demand. That makes the key split not “AI works versus AI fails,” but gross demand versus supplier pricing power. What evidence would convince you the capex cycle is correcting rather than merely reallocating toward lower-cost inference and new geographies?
I would like to hear more about Scott’s dot.com years in the 90s / early 00s and the large amount of money changing hands in those days pre cell phones, pre gmail.
I think those are some of Scott’s best examples of storytelling.
Read years ago in a scholarly economics treatment (can't recall where now) that the U.S. has something like an average 15 year boom and crash cycle over its history.
I had a front-row seat to that era while at Mail.com/EasyLink (later acquired by OpenText). Watching WorldCom implode, IBM scramble, and the telecom infrastructure unravel taught me something I still believe today: technology revolutions are real, but valuations often outrun business fundamentals. The hard part isn’t predicting the technology. It’s predicting which business models will actually survive once the hype fades.
Very interesting. What caused the steep decline in manufacturing productivity in the 1970s? Did the OPEC oil embargo play a role? It just made me wonder what other price shocks and global uncertainties today might end up shaping the current AI bubble as it plays out…
Cycles...
Its my favorite word.
Counterpoint: Normalization, not collapse.
Galloway is right that AI is entering a new phase—but I think he’s mistaking maturation for collapse.
The strongest part of his argument isn’t “AI is another Pets.com.” It’s that enterprises are shifting from “deploy AI everywhere” to “prove ROI.” Uber blowing through budgets, Meta and Salesforce emphasizing proven use cases, and tighter governance are real signals.
But that’s optimization, not necessarily demand destruction.
The missing link is evidence that this translates into lower infrastructure demand. To make the 1999 analogy work, you’d expect to see slowing frontier-model revenue, hyperscalers cutting GPU orders, HBM inventories building, packaging utilization falling, and sustained weakness in AI infrastructure spending. We’re not there.
In fact, hyperscalers are still committing record capital while the real bottlenecks remain HBM, advanced packaging, power, and transmission—not excess compute. Unlike telecom in 2001, supply is still constrained.
Galloway is right to ask whether lower token prices automatically translate into higher infrastructure demand. So far, the evidence largely says yes: falling inference costs have been accompanied by rising usage, stronger AI revenues, and record hyperscaler capex. Jevons isn’t a mathematical law, but the burden of proof now sits with those arguing that enterprise cost controls will overwhelm the powerful demand created by cheaper intelligence.
The more interesting debate isn’t whether AI is a bubble. It’s which layer captures the economics. Frontier labs? Hyperscalers? Semiconductor suppliers? Enterprise software? Or end users?
My base case remains that private AI valuations compress, enterprises become more disciplined, and capital shifts toward measurable ROI. That is very different from concluding AI infrastructure demand collapses. The analogy to 1999 skips a critical difference: today’s largest AI investors aren’t Pets.com—they’re some of the most cash-generative companies in history, operating in an environment where key components remain supply constrained.
The evidence increasingly points to a transition from indiscriminate AI spending toward ROI-driven deployment—not yet to a telecom-style unwind. If that changes, the signals to watch aren’t anecdotes about budget discipline. They’re slowing frontier-model revenue, hyperscaler capex cuts, weakening HBM demand, easing packaging constraints, and sustained excess GPU supply. Those are the dominoes that would validate Galloway’s thesis. Until then, the stronger evidence supports normalization, not collapse. Hat trick AI
I, too, have seen this movie before. I was a top-ranked internet analyst in the 1990's at a firm called Alex. Brown (then Deutsche Banc). I was the analyst who took Amazon, eBay, and other companies public, and I covered many more. I passed on the pets.com IPO, along with 50 others. I even wrote and taught a case study about it before the crash! 67% of the companies we backed are either still in business today or were successfully acquired by other companies. ALL of these companies' IPOs would now be considered VC A or B rounds, definitely not IPOs.
In the end, the dot-com bubble was just a blip, a healthy flushing out of the marginal players. Those who were in it for the long game not only survived, they became massive winners. The likes of Bezos were never building companies for the 1999 market; they were building companies for the 2009-and-beyond market they knew was coming. Staying power was the key, and the cost of capital was higher then than it is today on a relative basis.
There are some notable differences in today's AI economy (I'm not saying "it's different this time", but there are differences): (1) most of the funding for the LLMs is coming from private, not public markets, with the exception of the big players who already generate massive cash flows. This gives companies time to figure things out without the capricious nature of public markets driving their strategies; (2) the ease of adoption, and thus diffusion across markets, is much easier and faster than the Web was in 1999; (3) the war chests among the leaders are huge, and so are their backers.
AND, yes, I agree, none of the LLM players have put out a convincing financial story, especially when the circular financing is taken into account. This is the fly in their ointment (though I don't think they see it). The social hazard arguments are mute; there are enough people becoming reliant on AI at work and in personal endeavors that it will continue to grow... planet be damned! Which is why I am finding it really hard to use AI in my own work!!!
Warren Buffett was right when he said: "trees don't grow to the sky"!
What about the effect LLM data centres have on the environment, surely that should be a concern? Look outside!!! The world is burning.....
I agree with everything except your contention that our universal hatred of Zionist child killers is a poor byproduct of partisan agreement.
Having lived through the dot-com bubble and bust, I can definitely see the parallels. What we know is that the greatest value companies are built on the ashes of the “visionaries.” Wake me up when we cross the chasm, I’ll be more interested then.
Oh, I definitely think it’s a bubble, but I think we’re still on the early side of this bubble inflation. There’s an enormous about of money committed to building out the infrastructure, so that money will be flowing to the chip companies (and related businesses) for at least a couple of years.
I have a question about this FT quote below, I read the linked article too but didnt find an answer. How do we know productivity is truly exploding at OpenAI and Anthropic?
“The fact that incumbent software and knowledge work companies are finding only modest productivity gains by incorporating AI into existing workflows and organizational structures, while usage, revenue and productivity explode at Anthropic and OpenAI...”
The strongest parallel may be less 1999 valuations than who captures the productivity surplus after the shakeout. If models commoditize and enterprise buyers move from token volume to completed-task economics, value can migrate simultaneously to cheaper open-weight models, orchestration layers, and end users—while some infrastructure vendors still benefit from higher aggregate demand. That makes the key split not “AI works versus AI fails,” but gross demand versus supplier pricing power. What evidence would convince you the capex cycle is correcting rather than merely reallocating toward lower-cost inference and new geographies?
⭕️
I would like to hear more about Scott’s dot.com years in the 90s / early 00s and the large amount of money changing hands in those days pre cell phones, pre gmail.
I think those are some of Scott’s best examples of storytelling.
Read years ago in a scholarly economics treatment (can't recall where now) that the U.S. has something like an average 15 year boom and crash cycle over its history.
Fascinating! One feels it in the air, the eery calm before the storm...