The second-quarter earnings season is in full swing. So far, the results are landing in line or better than the upwardly revised Wall Street estimates. That’s the opposite of how this usually works. Analysts normally trim their forecasts as a year wears on. In 2026, they’ve done the opposite. The is on track to grow earnings north of 20% for a second straight quarter. The earnings are real. However, a meaningful slice of them is also an accident of accounting timing. That timing, the AI capex depreciation risk, hasn’t hit the income statement yet. But it is about to turn from a tailwind into a headwind.
The Golden Window
Currently, the entire earnings growth story is concentrated in the semiconductor and AI-infrastructure names. The accounting underneath it is where the catch hides. Here is what I mean. When Nvidia (NASDAQ:) sells a chip, it books the revenue and the profit almost immediately. The hyperscaler buying that chip does the opposite. It records the purchase as a capital asset and spreads the cost across years through depreciation. So the seller’s earnings jump now, while the buyer’s costs arrive later, in slow motion.
Here’s what makes this run unusual. Analysts normally walk their forecasts down as the year unfolds. Over the past five years, consensus has trimmed full-year estimates by about 2% on average at this point on the calendar. In 2026, they’ve gone the other way. The full-year S&P 500 earnings growth estimate has climbed from roughly 14% in February to north of 23% now, a swing of nearly nine percentage points in the wrong direction for anyone expecting the usual fade. With Q2 results landing through late July, that bar keeps moving higher.
That upward march is the golden window in motion. Every beat this quarter lifts the bar for the next one, and the more confidently the Street marks earnings higher, the more those forward numbers lean on costs that haven’t shown up yet. Make no mistake: this is the same setup I flagged in “Earnings Estimate Revisions Are Very Optimistic.” The AI capex depreciation missing from today’s numbers is exactly what those rising estimates are quietly assuming away.
Todd Castagno at Morgan Stanley calls this “a golden window where everybody looks good.” He’s right. Revenues and margins look strong among chipmakers and the companies buying the chips at the same time, which is exactly the kind of broad, simultaneous strength that convinces investors a cycle is durable rather than borrowed from the future. Make no mistake, there’s nothing improper here. This is how companies book capital assets. What’s different this cycle is the sheer scale of the spending, and the eventual AI capex depreciation is being overlooked.
Where The Bill Actually Lands
Here’s the problem with the everything-is-fine read. The spending is enormous, and it shows up in cash long before it shows up in earnings. The five biggest hyperscalers, Alphabet (NASDAQ:), Amazon (NASDAQ:), Meta (NASDAQ:), Microsoft (NASDAQ:), and Oracle (NYSE:), spent about $412 billion on capex in 2025. For 2026, the estimates run to roughly $760 billion. Yet the AI capex depreciation and amortization that those companies expect to recognize against all that spending in 2026 is only about $211 billion.
Read those two numbers again. They’re spending $760 billion and expensing $211 billion. The other $549 billion sits on the balance sheet, waiting. It becomes an earnings cost later, once the equipment goes into service and the AI capital depreciation clock starts. A good chunk of it isn’t even running yet, because the data centers housing it are still under construction.
The cleanest way to see the gap is the cash. For 2026, combined free cash flow at those five companies is projected to fall 91% to about $16 billion, while net income is projected to rise 25% to roughly $506 billion. A business can report half a trillion dollars of profit and throw off almost no cash in the same year. That’s not fraud. In this case, that’s the AI capex depreciation timing at work.
The Number Nobody Can Model
However, here is where it gets interesting. If depreciation is the future cost of today’s earnings, you’d want analysts to have a tight handle on it. They don’t. Look at the consensus estimates for Meta in 2028. The standard deviation of the revenue forecasts is just 4% of the average. The standard deviation on the depreciation-and-amortization forecasts blows out to 24%, six times wider. Translation: analysts broadly agree on what Meta will sell. They have almost no agreement on what it will cost to run the machines that produce it.
Why so uncertain? A few reasons. Most of these firms only shifted from asset-light to capital-heavy models in the past few years, so there’s little history to model against. Companies also have wide latitude to lengthen or shorten the useful lives they assign to equipment, and that single assumption swings the annual depreciation number significantly. On top of that, a growing share of the buildout is financed off-balance-sheet. As David Zion of Zion Research Group puts it, consensus depreciation estimates “could be systematically understated.”
That table is the entire bull case in five rows. The market isn’t paying for the $16 billion. It’s paying for the snapback to $387 billion. And the snapback is an assumption, not a result.
“You’re paying 22 times earnings today for profits whose single biggest future expense the analysts modeling them can’t agree on within a quarter of a trillion dollars. That’s the catch.”
“But The Revenue Will Come”
Let me steel-man the optimists, because they aren’t wrong about everything. The consensus view holds that capex growth tapers after 2026 while revenue keeps climbing, so free cash flow rebounds in a clean “V.” The same forecasts that show $16 billion in free cash flow this year also show it recovering to $185 billion in 2028 and $387 billion in 2029, with earnings compounding at around 20% a year through the end of the decade. If that plays out, today’s multiple looks reasonable in hindsight, and the depreciation wave gets buried under a bigger revenue wave.
Put real numbers on the bet. Consensus has the five hyperscalers’ capex climbing from $412 billion in 2025 to roughly $760 billion this year, then to about $820 billion in 2027 and $930 billion in 2028. Watch the growth rate, not the level. Spending jumps 84% into 2026, then the annual increase collapses to single digits. That deceleration is the entire argument. If capex growth stalls while net income keeps compounding near 20% a year, free cash flow snaps back on its own, because the cash stops rushing out the door faster than it comes in. The chart below is the bull case drawn to scale.
It’s a coherent story. It also leans on a capex taper that the same companies have run straight through at nearly every guide. Each time the Street pencils in a slowdown, the next quarter’s guidance lands higher. The 2026 consensus alone climbed from about $600 billion last November to $760 billion by February. So the model that gets you back to $387 billion of free cash flow assumes spending discipline from an industry that hasn’t shown any.
Maybe. But notice everything that case requires. It needs capex to slow on schedule, revenue to accelerate on schedule, and depreciation that everyone admits they can’t model to behave itself along the way. has aged well for a reason. When all the experts and forecasts agree, something else usually happens. The V-shaped recovery isn’t a forecast. It’s an assumption wearing a forecast’s clothes.
The issue is NOT whether AI is real. It is. The issue is whether the price already paid assumes a clean landing that the people closest to the numbers can’t promise.
What This Means For Your Portfolio
So what do you do with it? Start with the multiple. The S&P 500 trades around 22 times forward earnings, above its historical average, and that’s before the depreciation wave ramps. If the forward earnings inside that ratio are flattered by deferred costs, then the real multiple on fully loaded earnings is higher than the sticker says. You’re paying more than it looks.
I made a related point last month in Earnings Estimate Revisions Are Very Optimistic. Strip AI infrastructure out of the index, and the other 470-odd companies have seen their 2026 earnings estimates revised lower over the prior 17 months. This is the same warning from a different angle. The index’s earnings engine is concentrated in a handful of names, and a chunk of those names’ reported profits carries a deferred bill that the consensus is probably underpricing. Concentration risk and earnings-quality risk are now stacking on top of each other. I walked through the valuation side of this in as well.
In the Equity 60/40 Growth model, we’ve held the AI infrastructure sleeve at model weight rather than letting price appreciation push it higher. That’s not a call that AI ends. It’s a recognition that when a position’s earnings quality gets murkier, and its valuation gets richer at the same moment, you don’t add to it. You trim the names that have done the most work, you hedge the largest exposures while protection is still cheap, and you keep dry powder for the first real disappointment. Howard Marks has spent a career making the same point. The riskiest moment is usually the one that feels the safest.
The AI capex depreciation wave is coming. That part isn’t in dispute. The only open questions are how big it is, when exactly it lands, and whether the revenue arrives in time to absorb it. Right now, the market is answering all three with optimism and pricing the answer as though it were already known. When the first hyperscaler guides depreciation higher than the Street modeled, and one of them will, the golden window closes fast. Better to position for that before the tape forces the issue.
