Massive AI expenses will start coming back to roost soon. Analysts say tech workers should be worried.

The vibes should have been good. 

Meta’s fourth-quarter numbers had just trounced expectations. Revenue hit an all-time high. The stock price was ripping after-hours.

But as longtime tech analyst Laura Martin listened to CEO Mark Zuckerberg’s post-earnings conference call late last month, she heard what she later described as “a veiled threat.” 

It came as Zuckerberg touted efficiency gains of 30% for Meta’s software engineers thanks to new “agentic” AI coding tools. Productivity jumped as much as 80% for some “power users,” he said.   

“We expect this growth to accelerate through the next half,” he said. 

Martin, who covers Meta for Needham & Co., has been interpreting CEO-speak since starting as an analyst for Credit Suisse First Boston in the late 1990s. She had a pretty clear idea of the message Zuck was sending.

“If you’re only getting 20% or 30% more productivity as an engineer,” she said, “you’re going to lose your job.”

Giant tech companies at the heart of the AI boom have never been more profitable, powerful, or highly valued.

But for anything non-AI, Wall Street expects increasingly tight budgets at hyperscalers like Meta, Microsoft, Alphabet, and Oracle, as these companies start to digest massive costs related to their huge spending binge on AI data centers. Morgan Stanley estimates that group could recognize a whopping $680 billion of depreciation costs over the next four years. And that doesn’t even include Amazon, one of the most prolific capex spenders.

As those depreciation costs rise in coming years, hyperscalers’ profitability will hinge in part on how effectively they control non-AI operational expenses. That can be through lower labor costs, higher productivity, or spending less time and money on non-AI endeavors. Analysts say such adjustments are already starting to happen. 

A little over two weeks ago, Amazon announced 16,000 new layoffs, on top of the 14,000 job cuts it had already announced in October. Less than a month ago, Meta laid off about 1,500 workers from its Reality Labs division.

“Companies are trading opex for capex,” said Brian Pitz, an analyst covering Alphabet, Amazon, and Meta, among others, for BMO Capital. “They’re kind of shifting away from some of these lower return threshold projects in favor of AI.” 

Companies could also cut back elsewhere, including on dividends and stock buybacks, a point recently made well by The Information.

Before the AI boom, tech companies praised their lack of large-scale capital investment as a key attribute of their “asset-light” business models. The low-cost structure consisted largely of employees, intellectual property, and computing power, resulting in big profit margins and growing piles of cash that companies could use to buy back stock and boost returns.

“For two decades-ish, they talked about how they’re asset light and they’re efficient and they can buy back stock and they don’t need debt,” said Todd Castagno, a Morgan Stanley managing director specializing in accounting research. “And all of that has changed within the past 18 months.”

Castagno argued in a recent report that Wall Street seems to be hitching its near-term hopes for these companies to a large decline in the importance of non-AI expenses, which may not materialize.

Bank of America semiconductor analyst Vivek Arya, who tracks hyperscaler capex because of its importance to chip sales, noted that one of the biggest costs for these companies has long been employees, largely represented in the form of stock-based compensation expense. 

“They would hire a lot of people, and those people would consume a ton of stock compensation expense,” he said, adding that now “they are not hiring people with the same intensity. They would rather buy more GPUs. And guess what? Those GPUs don’t need as much stock-compensation expense.” 

In other words, Arya thinks that the potential returns to hyperscalers over time in the AI era could be larger than expected, as companies continue to effectively replace people with high-powered AI technology.

That seems to be the story that the companies themselves are telling. As part of their recent flurry of earnings reports, tech executives have stressed how they expect rising productivity to help offset the impact of depreciation expenses. 

On its post-earnings conference call last week, Alphabet said roughly 50% of code at the company is now written by AI agents, and only then reviewed by flesh-and-blood programmers. A day later, Amazon executives highlighted plans to expand on the 1 million robots it already has working at its retail distribution operation.

And on its call late last month, Microsoft executives used the phrase “efficiency gains” five separate times — the most on record in conference call transcripts kept by FactSet — all in the context of offsetting rising expenses and preserving profit margins.

So far, Wall Street appears to be buying it, with estimates seeming to show a large reduction in the impact non-AI investment costs on profits over the next few years.

To be sure, tech companies have gone through periods of belt-tightening before. That’s clear from the recent layoffs we mentioned earlier.

But their current focus on boosting efficiency is a part of the somewhat ironic situation tech workers face during what’s arguably the biggest technology boom of modern times.

During previous manias, like the dot-com blitz of the late 1990s or the proliferation of privately funded unicorns during the 2010s, tech companies became top destinations for educated workers looking for pleasant, high-paid jobs with stock options. 

But in those episodes, tech disrupted long-established non-tech industries — consider the impact of e-commerce on the shopping mall. This time, some of the most significant productivity improvements from AI seem set to disrupt tech workers themselves, as companies increasingly rely on AI coding tools such as Cursor and Claude Code.

A January research report from Goldman Sachs found that while the impact of AI on the overall job market remains limited, employment headwinds were visible in some specific occupations “and especially tech.” 

In a recent note, analysts at Evercore ISI spotlighted how productivity — defined in terms of sales per employee — increased sharply between 2023 and 2025, rising 22% at Amazon, 29% at Alphabet and Microsoft, and 48% at Meta. 

All else equal, such a burst in productivity might be expected to filter through to the bottom line, boosting profit.

But all else isn’t equal.

Now that productivity is likely to be rivaled by a parallel surge of delayed costs, in the form of the large, difficult-to-predict depreciation expenses these companies will report in coming years.

(Companies typically reflect the cost of such investments in the form of regularized depreciation that gets recognized each quarter, reducing earnings over time.) 

And since those AI investments have been massive, the depreciation costs will be, too. Is it possible that jumps in productivity can offset costs of this scale?

Morgan Stanley’s Castagno is skeptical that Wall Street’s consensus view of plunging non-AI expenses is correct. 

“In our view, it’s kind of starting to price in a free lunch,” he said. “And generally, free lunches don’t exist.”