Investors Question AI Accounting Which Extends Useful Life, Buries Costs

In what is turning out to be yet another significant challenge to the veracity of AI companies' financial projections, a growing number of venture and other investors are closely examining the way both startups and established enterprises are accounting for AI costs. The issue is that they appear to be conflating both the extended lives of data centers and the potentially short-lived expectations for AI chips and related costs. The result falsely extends the life of one while shortening the life of the other, thereby exaggerating profits. 

This latest set of questions contributes to the broader undermining of the industries claims about AI's extraordinary financial and operational impact, raising yet another reason for investor prudence. JL

Mark Maurer reports in the Wall Street Journal:

Tech companies are  spending hundreds of billions of dollars on data centers, chips and networking. (But) the cost of chips that have to be replaced in a few years can be lumped with buildings that stand for decades. This has investors seeking more details about tech spending on AI - and some data-center projects could face delays. Concerns over spending are leading investors to watch for signals that infrastructure projects will take longer than companies projected. “The construction-in-progress account is this big hole where hyperscalers can bury a lot of their costs.” Companies count AI chips because they buy GPUs during a data center’s construction. Public companies made 79 changes to estimates of the useful lives of assets, many of them increases in the lives of servers and networking equipment. Shareholders in tech companies are struggling to gauge the risk AI investments.

The massive AI build-out comes with a transparency problem.

Tech companies often provide the cost of AI data centers and chips associated with a long-term construction project. The catch: They generally don’t break out the costs for each, nor are they required to do so, despite the vastly different time periods in which facilities and chips depreciate.

That means the cost of chips that may have to be replaced in a few years or less can be lumped together with buildings that can stand for decades. This has some investors seeking more details about tech giants’ surging capital spending on AI infrastructure.

“The construction-in-progress account is this big hole where hyperscalers can bury a lot of their costs,” said Gaurav Kumar, an accounting professor at the University of Arkansas at Little Rock.

Large tech companies are collectively spending hundreds of billions of dollars on data centers, chips and networking. Spending on data-center construction is expected to exceed office-building construction for the first time as soon as next year. Some data-center projects could face delays due to issues such as power-supply limitations and local politics. Concerns over postponed spending are leading investors to watch for any signals that infrastructure projects will take longer than companies have projected.

 

For companies that share their construction-in-progress spending, they rarely go any further in breaking it down. The lack of detail can be significant when the figure includes both data centers that could have a depreciable life of between 20 and 40 years and AI chips that could become obsolete in less than three years.

“The disclosure is not evolving quickly enough to keep up with real-life demand for information about AI investment,” said Olga Usvyatsky, an accounting consultant.

The debate over the longevity of AI chips comes as many tech companies in recent years have said they expect their servers and network equipment to last longer without replacing or discarding them. Replacing the equipment less frequently helps preserve cash flows. It also reduces depreciation expense and increases reported profit, sometimes by hundreds of millions of dollars.

Companies include their AI spending in total figures on capital expenditures and property, plant and equipment that they are required to provide in their financial statements. They must report capex on assets not yet in service and include them in their total PPE capex, but aren’t specifically required to break these expenditures out separately. Accounting rules also generally require companies to disclose balances of major classes of depreciable assets.

The construction-in-progress account, which is immune to depreciation, holds all the costs for a fixed asset until it is ready for use. Companies then move the individual costs to building, machinery or another specific PPE category.

 

Google parent company Alphabet recorded $50.6 billion in assets not yet in service in 2024, up 44% from a year earlier. Amazon booked $46.4 billion in 2024, up 62% from the previous year. Meta Platforms, whose favorable treatment of a data center off the balance sheet has attracted scrutiny, said its construction in progress totaled $26.8 billion, up 10%. The account represents a chunk of net PPE for these companies, at 30%, 18% and 22%, respectively, filings show. Amazon declined to comment, and Alphabet and Meta didn’t respond to requests for comment.

 

Microsoft, in contrast, doesn’t give the number. The Redmond, Wash.-based company disclosed in its latest annual report that it committed $32.1 billion primarily related to the construction and improvement of data centers. In some cases, companies’ commitments may not equate to actual spending. Microsoft declined to comment.

Construction in progress collectively totaled about $214.5 billion this year for public companies with at least a $2 billion market capitalization that referenced AI infrastructure in their filings, according to investment research software provider Hudson Labs.

Companies tend to count AI chips in the construction-in-progress amount because they often buy graphic processing units, or GPUs, during a data center’s construction.

Public companies made 79 changes to their estimates of the useful lives of assets, many of them increases in the lives of servers and networking equipment, in 2024, the highest level since at least 2020, according to research firm Ideagen.

Shareholders in the tech companies are struggling to adequately gauge the risk of the AI investments. That’s in part because companies disclose very little about their experience using chips, said Ravi Gomatam, co-founder of Zion Research Group, an accounting and tax-research firm.

“It would be hugely helpful having the company talk more about what kind of assumptions go into coming up with the life determination,” Gomatam said.

Investors want to see a breakdown by project and components, which could give them a better idea of how much of the total is at risk of becoming obsolete, said Jack Ciesielski, owner of investment research firm R.G. Associates.

“Any frustration about construction-in-progress would be the lack of delineation between electronics and hard physical stuff like buildings,” Ciesielski said.

Any such changes in U.S. accounting rules would be a while away. The Financial Accounting Standards Board, which sets accounting rules for U.S. companies and nonprofits, has no projects on the standard-setting agenda to explore requiring a breakdown of construction in progress, Chair Rich Jones said. The organization will do additional outreach on the general topic of improved disclosures as part of a continuing agenda review, he said.

“We’re not hearing that people that are reading the financials aren’t getting the information they need, but we’re always receptive to understanding that,” Jones said, referring to capitalized property, plant and equipment.

Once a project is complete, companies’ disclosures on the value of the assets usually become clearer, but they still generally don’t provide a line item on how long they expect chips to be profitably used, said Bobby Carnes, an associate professor of clinical accounting at the University of Southern California.

“It’s a black box in the beginning with the construction-in-progress bucket and then once the project is complete, it’s a black box in the back end,” Carnes said.