The hidden menace behind Big Tech’s AI arms race: Meta, Amazon, and others are spending billions on hardware that’s worthless in 3 years, says CEO of Research Affiliates

There’s a wild paradox in the middle of the biggest story in tech right now. The GPUs and other essential hardware that the hyperscalers are spending on so lavishly to pack into their data centers, it turns out, go obsolete in a hurry. That’s the view detailed in an excellent new report from Research Affiliates, a firm that oversees around $200 billion in investment strategies for its RAFI index funds and ETFs. Author Chris Brightman—he’s RA’s CEO—contends that the AI arms race has effectively created a new industrial era. In this transformed ecosystem, companies aren’t “investing” in the traditional sense. Rather, they are churning equipment at such an incredibly rapid tempo to generate sales that it’s changing the very definition of capital expenditures.

Recommended Video

“They’re more like supermarkets than traditional tech or industrial enterprises, but their turnover isn’t in the likes of grocery items. It’s the stuff that generate their large language models, vector search, and other products,” Brightman told me in a phone interview. “They’re in an arms race where they need to replace their hardware very rapidly, in other words, restock their shelves in a hurry.” The problem, Brightman asserts, is that hyperscalers are taking losses on the large language models, vector databases, and other products they’re selling to companies and consumers, so the more hardware they buy, the more money they lose. “Right now, each is using AI to maintain crucial dominance in their field, and that makes sense,” Brightman observes. But, he adds, the immense spending needed to maintain those “moats” and keep rivals at bay could generate puny returns going forward, and harm their overall profitability.

In the article, Brightman spotlights the historic surge in AI capital expenditures (capex) that has mushroomed from $250 billion in 2024 to $650 billion this year by Bloomberg’s estimate, equal to 2% of GDP. That industry’s historic appetite for capital spawned the view that AI is becoming the new steel or railroads. But as Brightman points out, the equipment and infrastructure that supported those businesses is far different from the gear that drives AI. “Steel mills and rail tracks depreciated over 40 to 45 years,” he writes. He then contrasts those multi-decade useful lives to the scenario in AI. Hyperscalers such as Microsoft, Amazon, Alphabet, and Meta are depreciating their GPUs and other hardware over roughly five or six years on their income statements. Although those spans appear short, he says, their real “lives” are much shorter.

In an economic sense, assets become fully depreciated, or turn obsolete, when the revenues they generate no longer cover their cost of acquisition (reflected in yearly depreciation), operating expense, and cost of capital. According to Brightman, the industry numbers show that AI hardware loses its value over about three years. As proof, he cites data on the profitability of Nvidia’s industry-standard H100 GPUs. In their second year, an H100 spawned $36,000 in annual profit for a 137% return on investment. But by year four, the product was losing over $4,400 for a negative ROI of 34%, and the results sank fast from there. Writes Brightman: “The economic life of AI hardware is [a lot] shorter than its accounting life.”

It’s not that the equipment wears out. Physically, it can actually run a lot longer. The reason AI hardware loses potency so fast: Nvidia, AMD, and the other producers are crafting fresh offerings that each year provide enormous increases in computing power per watt deployed. Since the hyperscalers face tough energy constraints, they are constantly seeking gobs of new “compute” using dollops of extra electricity. Normally, if typical manufacturers were adding capital at the pace the hyperscalers are setting in AI, they would already have built a gigantic base of equipment and infrastructure they could deploy for years, without the need to keep buying more. Not so in this brave new business. AI equipment is evolving so fast that each year, the hyperscalers need to replace an immense part of their capital base just to maintain the same capacity for forging AI wonders. “Most of their spending isn’t growth capex, it’s ‘maintenance’ capex,” says Brightman. Nevertheless, the overall numbers are so huge that although only about one-third goes to expansion, that’s still good enough to hugely grow the volume of products and services they can deliver each year.

The hyperscalers are using AI, and taking big losses, chiefly to protect their turf

On our phone calls, Brightman nailed the conundrum for the giants of AI. “As they ramp the compute, they lose more and more money,” he says. “But they have plenty of rationale to do so for now.” All of the Big Four aim to provide the best AI features to enhance their signature offerings, and recognize that they will lose their leadership in those staples if the AI component isn’t top-notch. Amazon makes most of its money providing comput