This week we talk about tech bubbles, building moats, and infrastructure investment.

We also discuss capital expenditure, data centers, and employee compensation.

Recommended Book: The Art of Gathering by Priya Parker

Transcript

Many technology booms have early periods in which innovators have a first-mover advantage, and a lot of what happens in their industry is informed by the decisions those innovators make.

After that—depending on the technology, but this is common enough to be considered a trend—after that there tends to be a period of build-out and consolidation amongst the people and business entities that survived that initial, innovation-focused throw-down.

In the context of personal computers, this moment saw computer-makers like Microsoft and Apple scramble to pivot from figuring out what an operating system should look like and whether or not to use mice to navigate user interfaces, to a period in which they were rushing to scale-up the manufacture of now-essential, but previously comparably rare components: suitable screens for their monitors, chips that could power their increasingly graphical machines, and the magnetic materials necessary to produce floppy disks and spindle-based hard drives.

There’s an initial period in which new ideas and approaches provide these entities with a moat that protects them against competition, in other words, but then the game they’re playing changes, the rules are more fully understood and to some degree locked into place and agreed upon, and instead of competing for the biggest, most brazen new ideas, they lock onto one set of ideas that seemed to be the best of what’s available at that moment and build on those, iterating them at a regular cadence, but focusing especially on scaling them.

So at this second stage, they’re investing in the ability to out-produce their competition in some way, so they can eventually bypass that competition and (they hope) safely increase their prices and make a profit, as opposed to just larger and larger revenues with equal or greater expenses, continuing to be reliant on investor injections of capital, rather than generating their own surplus returns.

By many analysts’ and insiders’ estimates, we’ve just entered that second stage in the generative AI industry. That’s the sort of AI that generates text and images and code and such, and it’s increasingly becoming a sort of commodity, rather than a new, hot things that few companies can offer the market.

What I’d like to talk about today are the increasingly massive financial figures associated with this industry’s shift to that second stage of development, and why some of those insiders and analysts are voicing fresh concerns that this could all lead to a bubble, and possibly an historically large one.

—

There are many ways we could measure the growth of the AI industry over the years.

The US market size, for instance, which is a measure of the value of AI-oriented companies based on how much shares of their company cost or would cost on the open market, has ballooned from just over $100 billion in 2022 to an estimated $174 billion in 2025. That figure is expected to grow at a not quite 20% compound annual growth rate through 2034, which, if accurate, would put this market, in the US alone, at more than $850 billion.

Another metric we might use is that of capital expenditure, or capex, in this corner of the tech industry, which refers to the amount of money AI companies are using to buy, upgrade, or maintain their long-term assets, like new computer chips or the data centers they fill with those chips.

The seven most valuable US tech companies—Meta, Alphabet, Microsoft, Amazon, Apple, NVIDIA, and Broadcom (that last spot formerly held by Tesla, which was dropped from this designation in late-2024)—just those seven companies have spent $102.5 billion on capex this last financial quarter (and most of that was from just four of them, Meta, Alphabet, Microsoft, and Amazon, the remainder only spending something like $6.7 billion).

That’s a staggering amount of money, and due to a recent drop in consumer demand—the money individual US citizens spend on things like food and clothes and smartphones and cars and all the other things people buy—AI-related capex, spending by these massive US tech companies, has added more to GDP growth than consumer spending for the past two quarters.

All the things all the people in the US bought over the past two quarters did not cost as much, in aggregate, as what these companies spent during the same period, on new and existing assets. That’s pretty wild.

And it’s the consequence, partly, of the shift in these companies’ focus from providing goods and services that relied heavily on people—salary and stock compensation, basically, which is not a capex expense, because its spent on employees, not stuff—to spending heavily on all that infrastructure that they believe will be required to help them compete with those other companies that are also frantically investing in the same.

Whomever can built the biggest, baddest, most reliable and powerful data centers, and can get the AI-optimized chips to fill them, will have an advantage over their opponents in the new, developing tech world paradigm, it’s thought, so they’re pumping gobs of resources into exactly those sorts of assets, hoping to get ahead, build an insurmountable advantage, and put their competition out of business—or failing that, to establish themselves as the AI Coca-Cola, versus their opposition’s AI Mr. Pip.

Similar dynamics are playing out elsewhere, especially in China, where the market could reach a value approximating today’s US AI market in 3-5 years, and several times that, up to $1.4 trillion, by 2030—though like all of these figures, it depends on how we choose to measure these sorts of things, including what counts as an AI company, and in China, several of their major AI players are heavily involved in automation, robotics, which itself is expected to be a $5 trillion industry in that country by 2050.

Europe’s market is comparably smaller, as is its overall tech industry, but the AI market is now just shy of 15% of its total tech sector, up from 12% in 2022, and AI startups are attracting about a quarter of all VC funding in the bloc right now—so they’re starting from a less spendy start, but like pretty much everywhere the necessary knowledge and manufacturing base exists at the moment, the European AI market is growing a lot faster than anyone would have expected even just five years ago.

And there are real-deal innovations coming out of this tech; these investments are flooding into AI companies because these technologies, this version of them, the generative AI stuff, has completely rewired the programming world, AI bots and agents helping coders achieve a lot more, faster, and non-coders make things they wouldn’t have been able to build lacking these tools, imperfect as many of those tools are, under the hood.

We’re also seeing an explosion of other sorts of generated content, and the injection of these tools that make such content into Hollywood studios and consulting firms and government agencies, and everything in between, is causing equal parts panic and excitement, depending on whether you’re one of the people who feels like they might be laid off soon, replaced by software, or if you’re someone who profits from all those layoffs, and the payments from the companies that hope to save money by conducting them, replacing their comparably expensive employees with cheaper AI tools.

Things have gotten so wild that Meta’s CEO Mark Zuckerberg has started offering compensation packages ranging from $200 million to more than a billion dollars to top AI talent. Meta’s AI spending is already massive, and could hit $72 billion this year, but the company has said it could hit $100 billion in 2026, while Microsoft’s leadership suggested their 2025 spending of $30 billion could balloon to $120 billion in 2026.

OpenAI recently offered their employees large bonuses, in the hundreds of thousands to millions of dollars range, to counter those sorts of overtures from the likes of Meta, but there’s a lot of money flying around from all direction right now, much of it aimed at more AI infrastructure, or the relatively few people on the planet who understand this tech well enough to make a competitive difference in this industry.

That’s…a lot of money. There’s just so much spending happening, so many resources sloshing around in this one space right now, and all this investment is predicated on the idea that AI will change everything, we’re stepping into a new paradigm, and those who control the AI, will basically own the next game. So they’re all trying to set things up so they win the next game, or at least have the best hand possible when it arrives.

There have been increasingly loud arguments, made by long-time generative AI critics, but also, more recently, ardent AI boosters, that we might be running up against a wall of what these things can do for us; this version of the AI concept, at least.

And these arguments got louder with OpenAI’s release of their long-teased GPT-5 model, which some expected to be true AGI, human-grade, flexible, omni-capable intelligence, while others thought it might be a mono-focused superintelligence of some kind: the perfect coder, the perfect image generator, something like that.

What users got was not that. It seems to be better at some things, still not great at others.

This was an incredibly expensive model to produce—the training costs alone are estimated to be something like a half-billion dollars, and that’s just a portion of the total costs of creating this sort of model—and