By Jonathan Weil
The stock market is salivating over big increases in corporate earnings, but there's a catch. Much of the growth is due to a lag in when the costs of huge AI investments hit the books.
Analysts expect S&P 500 companies' earnings growth to top 20% for a second consecutive quarter in the current period, in part because of soaring profits by semiconductor makers and other AI-infrastructure companies. This has lent crucial support to the booming stock market, helping keep valuation ratios in check.
These equipment suppliers are benefiting from surging capital spending by Meta Platforms, Microsoft, Google parent Alphabet and other hyperscalers that are building the data centers necessary to power artificial-intelligence technology.
This combination has worked wonders for corporate earnings broadly. When Nvidia sells its chips, for instance, it books the revenue and earnings quickly. But its end customers treat the purchases as capital assets, meaning they defer the upfront costs and recognize them only gradually as depreciation on their income statements.
So, while the outlays immediately hurt the customers' free cash flow, they get to spread out the acquisition costs over many years on their income statements. In addition, they might not have to start recognizing depreciation expenses until long after the purchases if, for instance, the equipment is going into a new facility still under construction and isn't being put to use immediately.
Todd Castagno, an accounting analyst at Morgan Stanley, calls the current period "a golden window where everybody looks good." Revenues and margins look strong across the AI ecosystem at both the hardware suppliers and the big-spending buyers.
There is nothing unique here about the accounting treatment. This is the normal way assets get booked and then gradually written down. What stands out is the sheer size of the capex boom.
A big wave of depreciation expense is coming, and the size of the hit to hyperscalers' income statements is far from clear. Estimates of hyperscalers' future capital expenditures keep climbing sharply, and analysts' estimates of future deprecation expenses diverge widely.
Consider analyst estimates for Meta and their standard deviation, which measures how widely dispersed they are from the average. The standard deviation for their 2028 revenue forecasts is tightly bound at just 4% of the average estimate, while the standard deviation for their 2028 depreciation-and-amortization estimates blows out to 24% of the average, according to data compiled by Visible Alpha.
The wide divergence in D&A estimates shows a big blind spot that leaves future margins vulnerable.
Depreciation costs are difficult to model for a host of reasons. Most hyperscalers shifted to capital-intensive from asset-light business models only recently, so historical data isn't all that helpful. Transparency is limited, because the companies don't say how much depreciation is allocated to each of the various expense lines on the income statement.
Companies have wide discretion to shorten or lengthen their fixed assets' useful lives, which changes their yearly depreciation numbers. And a lot of the data-center build-out is being financed off-balance-sheet, which adds to the complexity. Many analysts don't even attempt to isolate the specific depreciation of AI equipment, and instead lump it into a broad D&A estimate.
S&P 500 companies reported about $1.3 trillion of capex for calendar 2025, according to data compiled by S&P Global Market Intelligence. That included $412 billion for just five large hyperscalers: Alphabet, Amazon.com, Meta, Microsoft and Oracle.
Estimates for those companies' capex in calendar 2026 total about $760 billion, according to Visible Alpha, compared with estimated D&A expenses over the same period of about $211 billion. And, as Zion Research Group founder David Zion says, "analyst depreciation estimates for each of these companies are all over the place." He says that the "consensus D&A estimates could be systematically understated."
The lag effect means the spending shows up in free cash flow long before it hits earnings. For 2026, the combined free cash flow at the five hyperscalers is expected to drop 91% to about $16 billion, while net income is projected to rise 25% to $506 billion. Free cash flow this year is expected to be negative at Amazon and Oracle and only slightly positive at Meta.
Current forecasts have the five hyperscalers' earnings compounding at about a 20% annual rate through 2029, with free cash flow rebounding sharply to $185 billion in 2028 and then soaring to $387 billion in 2029. The consensus view for now is that capex growth will taper off after next year while revenue keeps surging, allowing free cash flow to rebound in a V-shaped recovery. But nobody really knows how this will play out.
The stakes are high for investors trying to make sense of the numbers, as well as for passive index investors. The forward price-earnings ratio for the S&P 500 using estimates for the next 12 months is about 22 times earnings, which is above historical averages even before depreciation expenses ramp up.
A lot of today's earnings are coming from spending that won't appear on the hyperscalers' income statements for years. Everyday investors have a great deal riding on whether the AI titans can find the revenue someday to justify those costs.
Write to Jonathan Weil at jonathan.weil@wsj.com
(END) Dow Jones Newswires
June 18, 2026 05:30 ET (09:30 GMT)
Copyright (c) 2026 Dow Jones & Company, Inc.
Comments