Earnings Season 2025: AI Capex Eats the Income Statement

The Year the Spending Became the Story

For two decades the argument for owning big tech was light capital. Software margins modest factories money pouring out the back door. Looking back at the 2025 earnings seasons that identity cracked visibly and repeatedly. Microsoft Alphabet Amazon Meta and Oracle spent the year raising their capital spending forecasts almost every quarter up from the roughly $477 billion the sector spent between2022 and 2024 to commitments of 660 to 690 billion dollars for 2026 alone. Goldman Sachs counted more than a trillion dollars expected between 2025 and 2027 when the year closed. Each of those dollars buys something physical: data centers chips electrical infrastructure the machinery that our real estate coverage examines from the owner's side. In 2025 the income statements began to reflect it andQuarter by quarter the market reaction function changed

Beat and Drop

The pattern that defined the year was easy to describe and strange to observe. A giant reports revenue and earnings above consensus raises its capex guidance again and the stock falls anyway. Investors had stopped grading the quarter. They were grading the spending plan stapled to it. A demand-driven push for cloud and AI confirmed the revenue thesis of course but another $10 or $20 billion added to the annual capex budget further pushed back the target date.payout and increased the size of the bet. By late 2025 an increase in capital budget guidance was as likely to knock a stock down on earnings day as it was to bust. The market had no doubt about the demand for AI. It was about applying a discount rate to a promise that kept growing. Anyone who has lived through previous capital spending supercycles telecommunications in 1999 and shale drilling in 2014 recognized the reflex. Construction enthusiasmfirst anxiety about the bill then in that order every time

Two Statements, One Confusing Number

This is where I think there really was most of the confusion in the 2025 earnings calls and it's worth being precise about it because the entire debate over capital spending revolves around a distinction that is constantly blurred in the headlines. A company maintains some financial statements and two of them disagree regarding intentional spending. The income statement measures profitability. Revenue minus expenses equals net income and is based on matching an expense to the period in which it helped generateincome.The cash flow statement measures something different: real dollars flowing in and out of the bank account and it doesn't care about matching anything to anything. Capital expenditures the money spent building data centers and purchasing servers appear immediately and in full on the cash flow statement in the investing activities section the quarter the check clears. It doesn't show up at all on the income statement for that quarter. Not a dollar. That gap is neither a trick nor a loophole. That's thegoal of accrual accounting and it's exactly why a company can report healthy growing net income in the same quarter that its bank balance is hit by construction spending. The two statements answer different questions. Confusing them together makes a hyperscaler's earnings report look like nonsense. Keep them separate and 2025 will make a lot more sense

Why a Dollar of Capex Doesn't Hit Earnings on Day One

So if the cash runs out the day the data center is built where do the expenses go? On the balance sheet first as an asset. A server rack or data center is not treated as an expense when it is purchased because accounting rules require it to be capitalized meaning it is recorded as property plant and equipment an asset that is expected to help generate income for years not just this quarter. From there it loads the income statement gradually a little each period until depreciationDepreciation is simply the mechanism that spreads the cost of an asset over the years it is used rather than dumping the entire cost in the year it was purchased. A company that spends $10 billion on servers this year does not report $10 billion of expenses this year. It reports a fraction determined by how long the company's accountants believe those servers will remain useful a number called the asset value. useful lifeThat single assumption is doing a tremendous amount of quiet work on every hyperscaler's income statement and 2025 was the year analysts started reading the footnote where it lives instead of omitting it

The Useful Life Assumption Is Doing More Work Than It Looks

Useful life is an estimate not a law of physics which is precisely why it is so important. If you extend the assumed useful life of an asset the same total cost is spread over more years meaning that less of it hits the income statement in a single year which means that reported short-term profits increase even if not an extra dollar of cash has been earned. Shorten the assumed useful life and the opposite happens. More expenses per year lower short-term profits sametotal cost over the life of the asset in any sense. This is not a debate about whether the spending was prudent. It is a mechanical fact about how depreciation schedules work and that is why a change hidden in a 10-K footnote extending servers from five years to six say can move next year's earnings per share by a significant amount without a single customer doing anything different. The bears on 2025 argued that AI chips age quickly and that within a few yearsCutting-edge technology becomes obsolete in a year and a half so depreciation schedules based on a five- or six-year lifespan underestimate the true economic cost of building. The Bulls countered that older chips continue to earn their keep in inference workloads for years after they stop training frontier models so a longer lifespan was simply honest. Both sides were arguing about the same contribution to the same formula. I want to show you exactly whatthat input makes the output with numbers you can check yourself

A Worked Example: The Same Server Rack, Two Depreciation Schedules

Suppose purely for illustration and not as an actual figure that a company spends $10 billion $10 billion building a single data center campus in the first year and every dollar of that money is capitalized as equipment with an assumed useful life. I'll ignore taxes here to keep the arithmetic clean. They would scale everything by the same factor and wouldn't change the point

low straight line depreciation the simplest and most common method is to divide the cost of the asset equally over its useful life. First assume a useful life of five years. The annual depreciation expense is $10 billion divided by 5 or $2 billion per year every year for five years. By the fifth year the $10 billion has been completely spent and the asset is on the books with zero value. You get nothingmore of future earnings because there is nothing left to depreciate

Now change one assumption. Instead assume a useful life of six years. Nothing more about the changes in expense. The annual depreciation expense is $10 billion divided by 6 or about $1,666.7 million per year for six years

Line them up side by side and the effect is exactly what the lifespan debate was about. In each of years one through five the six-year plan spends $1,666.7 million versus $2 billion for the five-year plan a difference of $333.3 million a year.year to reported revenue in each of the first five years with exactly the same money spent exactly the same servers exactly the same business. Added up over five years the six-year assumption has kept $1,666.7 million of expenses off the income statement that the five-year assumption would have carried at that time

That money doesn't disappear. It appears in the sixth year. The five-year schedule will have fully depreciated its asset by then and doesn't record any additional depreciation expense a small tailwind to that year's earnings. The six-year schedule continues to collect $1,666.7 million in the sixth year because there is one more year left to run. The total lifetime depreciation is identical either way $10 billion because it has tobe. That is the actual amount of cash that was spent. What changes the useful life assumption is not how much it is spent. It is when. A longer assumed life delays the recognition of expenses and favors the years in between. That is the entire mechanism behind why the footnotes on 2025 depreciation became a valuation argument instead of a rounding detail

Case Study: Waste Management and the Trucks That Refused to Wear Out

If you want to see how much weight a lifespan assumption can have when no one checks it look at Waste Management in the 1990s. This isn't a story about AI or data centers it predates all of that by decades but it is the clearest real-world illustration I know of the exact mechanism above that was pushed beyond the point of honesty

Waste Management ran a fleet of garbage trucks and containers ordinary equipment that paid for itself nothing exotic. During the 1990s the company extended the assumed useful life of those trucks and containers well beyond what the equipment actually lasted in the field and inflated the assumed useful life. salvage value of that equipment that is the amount for which the company assumed it could still sell a worn-out truck for at the end of its useful life. Both movements push in the same direction. A longer useful life spreads the cost over more years and a higher assumed salvage value reduces the amount that must be depreciated since only the cost is depreciated minus what is expected to be recovered in the end. Together they allowed the company to report depreciation expenses significantly lower than the trucks actuallythey justified year after year which favored pre-tax profits by a wide margin without realizing an extra dollar of real profit

The SEC eventually filed a case that found the company had understated depreciation expenses and overstated profits by something on the order of $1.7 billion in the mid-1990s one of the largest earnings restatements of that era and ensnared top executives along with the company's external auditor. What strikes me about the case is how boring the mechanism actually was. No one hid income or invented a fake customer. They just kept pushing an assumptionThat's precisely the lever in my example above. Waste Management is that same lever pulled dishonestly and for years and it's why regulators and analysts alike treat a sudden unexplained extension of supposed useful life as a real warning sign rather than a radical change

The Free Cash Flow Gap Nobody's Income Statement Shows You

If you put together the mechanisms of the last sections the central financial fact of 2025 follows almost automatically. free cash flow roughly operating cash flow minus capital expenditures began to diverge from net income at some of these companies by a wide margin. Net income was being hurt by depreciation schedules that had not caught up with the actual pace of spending since much of the capital spending was so recent that it had barely begun to depreciate. Meanwhile real cash was leaving the building in full immediately to build the next data center without any delay in depreciation forcompanies reported record accounting profits in quarters where cash generation stagnated or fell and analysts increasingly cited free cash flow per share rather than earnings per share specifically to eliminate that illusion

I think this is the most useful habit that has been taught to investors in this sector in 2025. Net income tells you what accountants think happened after applying a lifespan assumption that is at best an educated guess. Free cash flow tells you what really happened to the bank account without the need for assumptions. No number is more real than the other they are measuring different things on purpose but when they start moving in opposite directions over several quartersconsecutively that gap tells you something about the timing of expense recognition that the headline earnings per share number is actively hiding from you

Where This Breaks: When Extending the Life Is Honest, Not a Trick

I've made life extension seem like a lever that companies use to achieve flatter profits and sometimes it is exactly that. I want to honestly defend the other side because the model breaks in a specific and common way and it's not always in bad faith

Lifetime is supposed to be an estimate of how long an asset will continue to produce value and sometimes the honest estimate actually changes. If a company originally assumed five years for a server because that's how long chips historically remained competitive at training the largest models and then discovers through real operating experience that those same chips continue to get solid performances running inference workloads for a sixth year after they age out of frontier training use extending the assumed life to six years is nota trick. It's accounting catching up with a fact about hardware. Cloud providers actually use older hardware for cheaper less demanding workloads well beyond their cutting-edge years and if that behavior is real and sustained a longer lifespan describes the business more accurately than the original more conservative assumption

The honest version and the aggressive version look identical from the outside. They both appear as a footnote change and an increase in short-term profits. The only way to tell them apart is to compare the claim with something external to the accounting itself. Does the company's own capital spending pattern suggest that it really believes the hardware lasts longer? Is it buying less replacement equipment than the shorter assumption would imply? Does the new assumption match what its peers with similar hardware are doing or the management changeIs it suspiciously scheduled for a quarter that needed earnings help? I don't think there's a clear rule here. I think it's a judgment call and honestly I find it really hard to make with confidence from outside the company since I can't see the actual failure rates or usage patterns of a data center's chip inventory. That's the real limit of everything in this piece. The worked example tells you exactly what a lifespan assumption does mechanically

How I Actually Read a Depreciation Footnote

If I was given a hyperscaler 10-K and told to spend twenty minutes specifically on the capex story this is the order I would actually work in as it is different from how I would read a normal industrial company's presentation

I'd skip the headline earnings per share number first on purpose and go straight to the property plant and equipment footnote to find the assumed useful lives for servers networking equipment and buildings noting whether those figures changed from the previous year. A change isn't automatically bad according to the section above but it's always worth noting and reading the reason given. Then I'd compare the capital expenditures on the cash flow statement to the depreciation expenses on the income statement. When the expensecapital is way ahead of depreciation which was true almost everywhere in 2025 that tells me the income statement hasn't caught up to actual spending yet and today's earnings per share measures a smaller business in terms of assets than actually exists right now. The gap between those two lines is a rough indicator of how much depreciation spending will still occur in future years regardless of what management decides to spend next. Only after thatI would look at free cash flow per share and compare its trend with earnings per share because a widening gap between them in either direction is usually the most honest chart in the presentation

My own opinion and this is clearly an opinion is that useful life assumptions deserve roughly the scrutiny that revenue recognition received after the accounting scandals of the early 2000s and they mostly aren't getting it yet because depreciation reads as boring in a way that a fake sale never is. I don't think that means every extension is Waste Management in a nicer suit. I think it means the burden should be on the company to explain the change inplain language and that a footnote that quietly pushes an assumption without a clear operational reason attached deserves more attention from an investor than the headline of the quarter

Winners of the Bill

None of the previous accounting mechanisms indicate who actually gets paid first in a capital spending supercycle and the 2025 earnings seasons answered that question clearly even as the lifespan debate remained unresolved. Nvidia remained the toll collector throughout the development with data center revenues dwarfing the size of all addressable markets from previous cycles.surplus to sell they reported delays and power prices they hadn't seen in decades. Memory networking and optics vendors followed the same wave. The simplest way to look at it: the market didn't go sour on AI in 2025 it changed the price of who grabs the cash first and the vendors got paid today in cash while the platforms are still waiting for a return that has to clear their own depreciation schedules before it even hits the profit line investors are watching

Whether that eventual return justifies the expense is a real question and is the subject of its own extended argument elsewhere on this site including the comparison that analysts continued to look for in 2025 between this build and previous infrastructure cycles. I deliberately don't raise that comparison here. What I wanted to isolate in this piece is more limited.Profits reported in hundreds of millions of dollars a year without a single customer or dollar of demand changing at all

The Bottom Line

The 2025 earnings seasons were dominated not by whether Big Tech could beat estimates - they mostly did - but by the accounting mechanics associated with the spending behind those results. Capital expenditures hit the cash flow statement completely and immediately.better off without additional cash actually earned a mechanism that the worked example above shows moves hundreds of millions of dollars a year and a mechanism that Waste Management shows can become a total fraud when no one controls it. Free cash flow diverged from net income because depreciation had not caught up with actual spending and that gap became the most honest number to look at. The model breaks down when a useful life extension is actually justified by actual operating experience rather than the need to hit a number andDifferentiating those two cases from outside the company is really difficult. My habit is to read the useful life footnote before the earnings per share line because by the time a headline number hits a research report one assumption has already decided how flattering it should look