What a difference a year makes in AI accounting
BREAKINGVIEWS-What a difference a year makes in AI accounting The author is a Reuters Breakingviews columnist. The opinions expressed are his own. Refiles to add topic code.
By Jeffrey Goldfarb
NEW YORK, Oct 29 (Reuters Breakingviews) - A simple ledger entry threatens the vast AI kingdom. Spending on the chatbot boom is heading into the trillions of dollars, with servers set to devour electricity now measured in gigawatts. Yet a minor change to one small figure buried inside the books of technology giants has enough power to wobble the whole edifice. It’s a question of just how long data-center kit, including all-important specialty chips, will keep contributing cash flow.
The four companies building and operating the most sprawling computational networks used to train, deploy and run large-language models have independently reached a broad consensus. After tinkering with their accounting treatment, Alphabet GOOGL.O, Amazon.com AMZN.O, Meta Platforms META.O and Microsoft MSFT.O for the most part now spread the cost of graphics processing units and other equipment over five to six years. Alphabet, for example, switched to six years from four for its servers in 2023. A year later, Meta changed its server and networking equipment time-line to between five and 5.5 years from a range of between four and five years.
These forecasts are of significant consequence. The longer the span, the lower the annual depreciation charge that hits the bottom line. Small tweaks imply big movements. An interactive Breakingviews calculator suggests that trimming a single year from the estimated “useful life” of these AI assets would erase some $36 billion of combined net income next year from the Big Tech foursome, theoretically leading to nearly $800 billion of lost market capitalization at their current valuation multiples.
Click here to view calculator in browser.
The lack of detailed disclosure in this area invites questions about whether the accounting choices match up with reality. It also will attract renewed and greater scrutiny as concerns mount about how long it will take to generate returns on all the spending. Meta, for one, just started burning cash in its most recent quarter after generating more than $150 billion in free cash flow over the previous three-and-a-half years. The social-media empire led by Mark Zuckerberg said it was holding $134 billion of servers and other related gear at the end of September and had reduced their value on the books by more than $14 billion over the first nine months of the year.
Meanwhile, the cost to insure against Oracle ORCL.N failing to repay its debt is surging in the credit default swaps market. Cybersecurity vulnerabilities have rattled investors and prompted companies to restrict autonomous AI agents from accessing their systems. The once-mooted $2 trillion valuation for Anthropic, the lab behind robot coder Claude, in its initial public offering is now meeting investor skepticism, Reuters has reported.
Among these debates, the accounting question has proved a particular flash-point. In 2025, Amazon shortened the projected useful life for a subset of its servers and networking equipment to five years from six, acknowledging the growing pace of technological development and consequently quicker lapse into obsolescence. The decision reversed two earlier extensions, from three years to four in 2020 and then from four years to between five and six in 2024. None of its major peers have followed suit, but Microsoft now discloses an unhelpfully broad range of two to six years.
Accounting standards are as much art as they are science, especially when it comes to the complicated assumptions around hundreds of thousands of Nvidia NVDA.O GPUs at the heart of AI development. At the heart of it all is the presumed impact of technological evolution. For a while, the chipmaker followed a steady cadence, introducing new generations of its silicon architecture every two years. Then, in June 2024, boss Jensen Huang announced a new “one-year rhythm,” starting with the company’s Rubin platform.
There’s little history to go on to determine how long such technology will last. Cutting-edge GPUs are most important in training advanced models, described by the industry as inhabiting the frontier. Second- or third-best designs, however, often provide sufficient computing power for fine-tuning existing models or calculating responses to user queries. An important consideration is whether older silicon, by dint of being less energy-efficient, is simply uneconomical to keep running.
To gauge the effect of any changes to these assumptions, the Breakingviews calculator uses three steps. First is an estimate of just how much server kit, mostly GPUs, is on the books. This starts with the most recent change to “useful life” for each company, dividing the reported impact to operating income that accompanied the revisions by the gap between the two depreciation rates to derive the implied historical asset pool.
These figures are then increased in line with the subsequent growth in AI capital expenditure at Alphabet, Amazon, Meta and Microsoft, giving a projection of how much server equipment there will be in 2027. After resetting the useful life schedules, the hit or benefit to depreciation is factored into 2027 earnings per share, based on analyst forecasts compiled by Visible Alpha before Alphabet, Amazon and Microsoft reported their latest quarterly results this week. If useful life shrinks, so does income because of the additional depreciation cost, thus reducing a company's tax liability. This, too, is incorporated into the calculations.
The final step multiplies any EPS change by each company’s respective share price multiple of estimated 2027 earnings to determine the resulting gain or loss in market value. Any extension in useful life would provide an uplift in net income. It’s unlikely to happen, but recent debates from Silicon Valley to Washington, Brussels and Beijing over slowing down the pace of developing frontier models over safety concerns could mean infrastructure gets left in place for longer.
A more plausible outcome is that the pace of technological improvement accelerates, old kit becomes a great burden and depreciation schedules shorten, leading to less profit. The accounting may seem like an afterthought, but it crystallizes many of the biggest concerns clouding the boom.
Follow Jeffrey Goldfarb on X and Linkedin.
(Editing by Jonathan Guilford; Production by Pranav Kiran)
((For previous columns by the author, Reuters customers can click on GOLDFARB/jeffrey.goldfarb@thomsonreuters.com))