The AI CapEx Arms Race

Why Cloud Providers Are Spending Hundreds of Billions in AI CapEx.

Just a few weeks ago Microsoft published their earnings, beating expectations on both the top and bottom lines. Total revenue was up 16%, yet the stock fell.

The reason for this was because Microsoft Azure, their cloud division, fell short of expectations with “only” quarterly growth 29% compared to expectations of around 31%. It was the first time since 2022 that Microsoft has failed to meet or beat expectations for the cloud segment.

Beyond the fact that Wall Street has seemingly become anesthetized to extremely strong, double-digit quarter-over-quarter growth that would make most other Fortune 500 companies blush, the real news behind Microsoft’s earnings was deep within Microsoft’s cash flow statement.

The company’s capital expenditure (CapEx) soared by 79% from a year earlier to reach an astonishing $14 billion in the last quarter alone. Over the past year, Microsoft has poured $44.4 billion in CapEx, with almost all of it going to AI. To put this number in perspective, that’s roughly the amount Elon Musk overpaid paid for Twitter, or if you prefer a military metaphor, almost 700 F-16 jets.

However, ludicrous investments in AI are not unique to Microsoft. In fact, Amazon, Meta, Microsoft and Google have all been pumping billions of dollars into AI.

The 4 tech giants alone will spend a combined $200 billion this year on AI-related CapEx, mostly data centers, chips, cooling systems and building materials. To put that number in perspective, $200 billion is twice what the USA spends on Homeland Security, twice the yearly budget of Ireland, or 4x the yearly budget of the French army.

Whatever way you spin it, it’s an enormous amount of money.

The question is why. Why are these cloud providers plowing hundreds of billions into AI CapEx and should they be ?

The Inevitability of AI Investment: A Game Theory Perspective

Consider for a moment that you are at the helm of one of two major cloud providers in a world where AI is set to revolutionize the industry. However, put the case that in order to succeed in AI and win market share/increase revenues, you needed to invest a lot of money into AI.

Do you invest?

Of course you do.

Think it through: if you invest and your competitor abstains, you gain market share, revenues soar, and your stock goes up. As the CEO, your equity in the company means you get to pick out a new penthouse with a view of Central Park.

However, if you abstain and your competitor goes all in on AI, you lose market share, your stock tanks and you’re suddenly calling your parents to see if your old room is still available.

On the flip side, if both you and your competitor invest, sure, you might drive up costs, but hey, Wall Street loves a good AI story so your stock might still rise. However, if neither of you invests, well, nothing changes… No bonus for you.

If you map out the situation with game theory, you find that you land in a situation where both players are doing the best they can (i.e. earning the highest profit) given what all the other firms in the market are doing. No player can gain an advantage by changing their own strategy, thus creating a Nash Equilibrium. The catch? This particular equilibrium could lead to companies overbuilding and overspending in a category that perhaps doesn’t warrant such levels of investment.

Here’s what that looks like in a game theory matrix, with the Nash Equilibrium circled in red:

Now, if you’re thinking this is just a fancy way to say, “Everyone’s investing in AI,” you’re not wrong. But here’s the twist: this isn’t just theory. It’s the reality of today’s cloud oligopoly.

Amazon, Microsoft, and Google—together holding 67% of the market—are locked in this high-stakes Nash Equilibrium. Changing course would mean surrendering market share and profits to the other. So, as per the theory, it seems the only option is to keep throwing billions at AI, even if it means overshooting the mark.

Beware of Over-investing

Back in the 90s, telecom companies thought they’d struck gold with the internet. Eveyrone realized the tremendous potential of the internet and telecom companies invested significant amounts to lay the infrastructure for the internet. The companies believed individuals would be willing to pay high prices to go online, and as a result, CapEx investments surged by 350% and they laid 600km of cable.

The only problem was that sure, people wanted the internet, but they wanted it for free.

Fast forward to today, and the tech giants today seem to be repeating history, betting big on AI under the assumption that the willingness to pay exists (sounds like the narrative during the crypto boom: “the use case hasn’t been found yet but we’re building the infrastructure”).

Yet, according to The Information, OpenAI’s recent struggles is a stark reminder that not everyone’s willing to pay for AI. The company lost $540 million last year, on just $30 million in revenue. Not exactly a goldmine.

There’s also the risk is that AI models could become a commodity. As open-source models become “good enough” for a large majority of use cases, it will be increasingly difficult for AI providers to offer differentiated models with a strong value-add versus open-source models. The more market share the open-source models win, the more difficult it will become for the big-name models to justify their costs.

And let’s not forget the law of diminishing returns. Sure, investing in computing power enables the tech giants to train bigger models. But nobody should assume that models are proportional to the amount of money thrown at them — or their size. Case in point, MistralAI is succeeding because their AI models are smaller, less costly to run, and more accurate. Plus, the size of the models are inherently limited by the amount of data available. Most of the general-purpose AI models available today have already been trained on the appropriate, human-generated data, and AI-generated data just doesn’t cut it — research shows that AI models collapse when trained on recursively generated (artificial) data.

In the end, the cloud giants find themselves locked in a high-stakes game where the rules demand relentless investment in AI CapEx. Yet, as history has taught us, the allure of cutting-edge technology often leads to overspending. The risks of overbuilding and overspending loom large, and the echoes of the 1990s telecom bubble serve as a cautionary tale. As AI evolves and the arms race intensifies, the question remains: Will these investments yield the transformative returns they promise, or will the top cloud providers find themselves stuck in a web of diminishing returns? Only time will tell, but in this game, the stakes have never been higher.