Benefiting from the booming AI industry, companies are racing against time to prove that their huge investment in GPU chips is paying off. However, an obscure issue will make this effort even more difficult.

Barclays analysts pointed out in a recent report that the depreciation related to large-scale AI chip investment is the 'not-so-hidden' cost of AI, and few investors calculate this in their valuation analysis of these companies. #AI trend#

Depreciation is an accounting method that allows companies to spread the cost of capital investments over their expected useful lives. This means that when a high-market-cap technology company purchases GPU chips worth billions of dollars, it does not immediately record them as expenses but as capital expenditures.

This can bring huge profits in the short term, as capital expenditures are not immediately recorded on the company's income statement, but as depreciation expenses during the asset's useful life.

The potential problem is that the useful life of AI GPU chips may be much shorter than many people expect, especially when AI chips go through an accelerating innovation cycle, resulting in higher-than-expected depreciation expenses and ultimately dragging down profits.

The depreciation cost related to GPU chips will be very high, so Barclays lowered its profit expectations for cloud providers Alphabet, Amazon, and Meta Platforms next year by 10%.

Ross Sandler, a Barclays internet analyst, said: 'Depreciating AI computing assets is the biggest expense for these leading companies. We believe that as we look towards 2025, this risk may become more apparent, so we have been paying attention to it early on.'

Due to the large technology companies spending hundreds of billions of dollars to purchase expensive GPU chips from companies such as Nvidia, depreciation costs will increase significantly in the coming years, especially as Nvidia switches to launching a new product every year.

Ted Mortonson, managing director and technology strategist at Baird, told Business Insider: 'Because Nvidia's design cycle is so tight, its main product release interval is about one year, so all these products have different characteristics, functions, and power configurations.'

Mortonson said: 'This is a resistance', adding that this resistance is enough to affect valuations and cause AI stocks to fall in the next year.

Barclays estimates that Wall Street consensus underestimates depreciation costs over the next two years.

For example, the bank expects Alphabet's depreciation costs in 2026 to reach $28 billion, 24% higher than the current widely expected $22.6 billion.

For Meta Platforms, the difference between Barclays' expected depreciation and Wall Street's expectations is even greater, at $30.8 billion and $21 billion, respectively, meaning potential costs are 47% higher than expected in 2026.

Sandler of Barclays said: 'We believe that with this modeling error, the stock prices of GOOGL, META, and AMZN are 5% to 25% higher than the general expectation.'

He added: 'Although we do not think that the valuations are too high compared to the historic bubble era of 2021, the prosperity of AI makes it clearer whether large technology companies need to expand in multiple ways, so under this background, depreciation (and valuation) dislocation may be subject to strict examination.'

The chief financial officers of large technology companies use an accounting method to extend the useful life of server assets from five years to six years or longer, as this can spread costs over a longer period of time and reduce the impact on earnings.

However, due to the fast pace of new GPU chip releases by Nvidia, there are limitations even in this method.

'We believe that after the six-year plan, no large enterprise will extend the useful life of servers, because the GPU cycle is increasing rapidly. As a result, large enterprises may have to bear higher depreciation costs in the future, unlike the past few years, when they adjusted the useful life.' explained Sandler.

For Mortonson, all of this comes down to the return on investment in AI capital.

'Wall Street has a big problem. They're now spending over $200 billion, and capital expenditures have increased by over 50%. Where is the investment capital return?' Mortonson asked. 'We're still early in this phase, and with all the accounting work, it all comes down to the investment capital return, which I think you won't see until 2025 or 2026.'

Mortonson added,"I don't think there is a conclusion at present. I think accountants must master it and there must be greater transparency between extending the life of networks, storage, servers and GPUs. That is the bottom line."