- Companies are facing rising depreciation costs from their massive chip investments, Barclays says.
- Barclays warns these costs will significantly impact earnings estimates for top tech firms.
Companies getting a boost from the booming AI trade are in a race against the clock to prove that their massive investments in GPU chips are paying off, but there's a little-talked-about issue that will make that endeavor even harder.
Depreciation related to massive AI chip investments is the "not-so-hidden" cost of AI that few investors are factoring into their valuation analysis of these companies, analysts at Barclays said in a recent note.
Depreciation is an accounting method that allows companies to spread out the cost of a capital investment over its useful lifetime. That means that when a mega-cap tech company buys billions of dollars worth of GPU chips, it doesn't immediately record that as an expense, but rather as a capital expenditure.
That can lead to big profits upfront, as the capital outlays don't hit a company's profit and loss statement immediately but are rather recorded as a depreciation expense over the asset's useful lifetime.
The lurking problem is that the useful lifetime of AI GPU chips can be a lot shorter than many expect, especially as AI chips go through an ever-accelerating innovation cycle, leading to higher-than-expected depreciation expenses that ultimately drag down profits.
The depreciation costs related to GPU chips will be so big that Barclays is trimming its earnings estimates of cloud hyperscalers Alphabet, Amazon, and Meta Platforms by as much as 10% heading into next year.
"Depreciation of AI compute assets is the biggest expense for these leading companies," Barclays internet analyst Ross Sandler said. "We think this is a risk that may rear its ugly head as we start looking ahead into 2025, so we are flagging it early."
With mega-cap tech companies spending hundreds of billions of dollars on pricey GPU chips from the likes of Nvidia, massive depreciation costs will add up over the next few years, especially as Nvidia shifts to a new product launch cadence of one per year.
"Because Nvidia has this very aggressive design cycle of roughly a year between major releases, all of those products have different skews and functionality and power profiles," Baird managing director and tech strategist Ted Mortonson told Business Insider.
"It is a headwind," Morton said, adding that it is big enough to impact valuations and send AI stocks lower over the next year.
Barclays estimates that Wall Street consensus is underestimating just how big the depreciation costs will be over the next two years.
For example, the bank expects Alphabet to record $28 billion in depreciation costs in 2026, which is 24% more than current consensus estimates of $22.6 billion.
For Meta Platforms, the mismatch between Barclays' depreciation estimate and Wall Street's is even further askew, at $30.8 billion versus $21.0 billion, respectively, representing potential costs being 47% higher than expected in 2026.
"GOOGL, META, and AMZN shares are between 5% and 25% more expensive than the consensus estimates perceive given this mis-modeling, in our view," Barclays' Sandler said.
He added: "While we don't think valuations are stretched vs. a historic bubble-y era like 2021, the AI boom has shined a brighter light on whether multiple expansion for big tech is warranted, so in light of this backdrop the depreciation (and hence valuation) disconnects are likely to be scrutinized."
One accounting method mega-cap tech CFOs are using is extending the useful life of their server assets from five years to six years or more, as that would spread out the costs over a longer time period and dampen the hit to earnings.
But even that has its limits because of how quickly Nvidia is releasing new GPU chips.
"We don't see any mega cap extending useful life of servers after this 6-year schedule, as GPUs cycle times are increasing rapidly. The result of this is mega caps are likely to have to absorb the higher cost of depreciation expense going forward, unlike the last few years when there useful life tweaks were happening," Sandler explained.
And for Mortonson, it all comes back to the return on invested AI capital.
"Wall Street has a big question. They are now spending over $200 billion and their CAPEX is over 50% up. Where is the return on invested capital?" Mortonson asked. "We're so early in this, that combined with all the accounting, it all wraps up to return on invested capital, and I don't think you see a return on invested capital till sometime in 2025 or 2026."
Mortonson added: "I think the jury is still out. I think the accountants got to take a hold of it, and there's got to be much more transparency between extending useful life of networking, storage, and servers versus GPUs. That's the bottom line."