Could 2025 Represent a Near-Term Peak in AI Capex?

Forbes

23-04-2025

As John Templeton once said, believing that 'this time is different' can be one of the costliest ideas in investing. So why should we think the AI boom is any different?
Futuriom research shows that five of the top hyperscalers (Alibaba, Alphabet, Amazon, Meta, and Microsoft) are set to spend about $350 billion on capital spending to build out AI capabilities in 2025. This is, of course, an exciting development. But our research shows that this capex spending is well above a long-term mean as a percentage of revenue, a trend that could pull back very soon.
Technology investment surges are exhilarating and fun. Just as the period of 1996-2000 gave us fun new things like web browsers, Pets.com, and Amazon.com, over-enthusiastic investors got over their skis and the capital boom eventually slowed.
Why should this time be different? I believe the AI investment boom is real and here to stay for at least the next decade, but just as the initial Internet boom was met with a roller coaster of pullbacks and surges, this first stage of AI capital spending may be at a dangerous point and due for a pullback.
The Internet industry reset from 2000-2002, but then later gave birth to next-generation plays like Google and Facebook in its next surge (the Google IPO wasn't until 2004!). I think we are ready for the AI reset, as hyperscalers and technology investors reassess how much they are spending on AI infrastructure. The global trade crisis may be a catalyst.
Enterprises are also in an early stage of evaluating their investments in AI and need time to digest the results. The chaotic and unpredictable macro environment, with see-sawing policies on tariffs and global trade, won't help the situation.
In the year 2000, I was working at Light Reading, a dot-com news and analysis firm that follows the telecom industry. In other words, we were in a bubble following a bubble.
At Light Reading, one of the first integrated digital B2B news and research firms, we launched a research service in 2000 to evaluate what the telecom boom would mean for our readers, clients, and us. What we found was interesting. By analyzing capital spending trends in the industry—specifically capex as percentage of revenue—we found that the capex investment far exceeded prior cycles and was due for a pullback.
We announced these research much to the chagrin of the industry, which attacked us. But we were right. And our readers were thankful for that.
And as Mark Twain said, history rhymes. My research firm Futuriom's analysis service, Cloud Tracker Pro, recently conducted a similar analysis of hyperscaler capex trends, using the capex/revenue ratio as an anchor, and found similar results to telecom in 2000: Capex ratios are well above historical means and likely to revert.
The major hyperscalers we studied have boosted capital spending 50% from just two years ago and are set to target an astounding $350 billion to build AI and cloud datacenters and facilities in 2025.
Historically, the top five hyperscalers spent an average of 12.5% of their revenue on capital spending in the years prior to 2025 (2021-2024). This represents the period of time spanning the COVID buildout (2021), as well as a capex slowdown in 2023.
With hyperscaler capex approaching 22% on average—well above the historical mean of 12.5%—it's unlikely to stay that high. And if it pulls back toward the historical mean, it could drop by 20% or more.
Futuriom shows that hyperscaler capex of 22% is well above the long-term mean of 12.5%
What's concerning is that the surging ratio of capex-to-revenue is accompanied by slowing cloud revenue growth and a growing global trade war. The capex might be sustainable, for example, if revenue growth accelerated. But it's not accelerating. The rate of growth has slowed.
The parallels with the telecom industry aren't exact. The telecom industry was in a worse place in 2000 than the hyperscalers are now. The telecom boom was financed primarily with debt, while most of the hyperscalers finance their capex with profits. And many telecom companies were marginally profitable before they loaded up their balance sheets with debt. The largest hyperscalers are cash-printing machines.
But still, we believe the capex trend is not sustainable over the next couple of years, as additional factors weigh on budgets going forward: 1) Cloud revenue growth is slowing 2) Earnings are likely to be hit by accelerated depreciation of AI chips and 3) Tariffs, tariffs!
Tariffs and the global trade war may just be a catalyst to more cautious thinking. As demonstrated by NVIDIA's recent announcement of potential export fees and the trade war's impact on costs, there are many risks growing in the global supply chain. The sentiment is shifting.
This sentiment shift and uncertainty about global trade, along with capex spending ratios that are well above the historical norms, means that investors should be cautious and expect a capex pullback.
So buckle up for some more to-and-fro in the cloud industry! As I said, I remain unduly optimistic about advances in AI in the long term, but it's still early. And this early surge in spending and investment may soon be tempered by a global slowdown.
Like most industry research and analyst firms, Futuriom provides paid research and marketing services to technology companies. These services include subscription research, custom research, and report sponsorships. Of the companies mentioned in this article, Futuriom has had a paid relationship with NVIDIA in the past 12 months. Individuals from some of the companies mentioned may subscribe to our premium research service, Cloud Tracker Pro.