What AI's spending boom is really buying

The big picture: Of course everyone in the AI race — Microsoft, Google, their giant rivals and their startup challengers — views AI as the industry's next big platform and believes it will keep getting more useful and lucrative.

• But if or when the current investing frenzy subsides, or even if there's a big AI bust, a company that built a giant data center for AI still has a giant data center.

That's why no one in Silicon Valley is terribly worried about the risk of overbuilding.

• Everyone above a certain age in the industry remembers what happened to overenthusiastic investments in fiber capacity during the late '90s dotcom bubble.
• The timing was off, and plenty of investors got burned. But before long we needed all that bandwidth and more.

Between the lines: It's not that hard to repurpose an AI server farm as a computing center dedicated to some other kind of work.

• Remember that the Nvidia chips powering today's massive AI model training projects were, until the advent of ChatGPT, known as GPUs — graphics processing units.

Originally, GPUs were for manipulating images, and that made them valuable for gaming rigs as well as professional workstations.

• In the 2010s, the crypto industry discovered that the same devices were the perfect tool for bitcoin mining and other blockchain-related computing needs.
• Then researchers discovered that these processors are also perfect for training generative AI — and that turned top-of-the-line GPUs into the most prized chips on the planet.

Today, the companies investing billions in new data centers say it's just the price tag of training new AI frontier models on the path to AGI or "superintelligence."

• They also say they're going to keep needing more data center capacity to run those models ("inference") as they rebuild the foundations of business backends and consumer online services around AI.
• Finally, they say these investments strike a blow for the U.S. in a geopolitical struggle with China for control of global AI.
• That argument has gained traction in Washington even though no one has been able to specify how you know who's won this conflict or what the winner gets.

What they're saying: In 2011, browser pioneer and venture capital veteran Marc Andreessen declared that "software is eating the world." But more recently he's taken to touting an agenda of physical infrastructure development with the slogan "It's time to build."

Yes, but: Another way of looking at this capital spending is that the tech industry's leaders have more money than they can productively deploy.

• As of the end of the most recent quarter, tech's trillion-dollar giants — Microsoft, Google, Apple, Amazon and Meta — had nearly $400 billion cash on hand. And more keeps pouring in.
• There's only so much money you can spend on product development and researchers. (Meta's Mark Zuckerberg has lately tested the outer limits of the market's tolerance in that realm.)

These companies hate the idea of giving profits back to their shareholders directly in the form of dividends — though Apple and Microsoft have been doing so regularly for some time now.

• That's because paying dividends sends a signal to investors that the firms' growth days are over and their stock prices should take a more muted trajectory.
• The tech giants are less hostile to giving the money to shareholders indirectly through stock buybacks, but there are limits to how much money you can funnel via that route.

The bottom line: On the financial landscape as seen from the Big Tech boardroom, a data center project looks like just another useful way to park some billions.

• If AI keeps booming, the companies will have capacity to keep up.
• If it goes bust, they have computing assets they can retool to support whatever comes after AI.

Our thought bubble: Data center projects can spike construction jobs but they're lousy at creating long-term job growth. As the well-known investor Nancy Tengler wrote Friday, "We believe companies are investing in technology instead of human capital."

• If only AI infrastructure spending were as effective at goosing employment as it is in training models, the U.S. economy might really benefit.