Earnings Season 2025: AI Capex Eats the Income Statement
Looking back at 2025, the biggest companies on earth kept beating estimates and kept getting punished for the spending attached. The market spent the year learning to price a capex supercycle running through income statements built for capital lightness.
The Year the Spending Became the Story
For two decades, the magic of big tech was capital lightness, software margins, modest factories, cash gushing out. Looking back at 2025's earnings seasons, that identity visibly cracked. The largest cloud and AI companies, Microsoft, Alphabet, Amazon, Meta, and Oracle, spent the year raising capital expenditure guidance almost every quarter, on a path from the roughly 477 billion dollars the sector spent across 2022 through 2024 toward commitments of 660 to 690 billion for 2026 alone, with Goldman Sachs tallying more than a trillion dollars planned for 2025 through 2027. Every one of those dollars buys data centers, chips, and power infrastructure, the physical machinery of AI our real estate coverage examines from the landlord's side. In 2025 the income statements started showing it, and earnings season by earnings season, the market's reaction function changed.
Beat and Drop
The pattern that defined the year: a giant reports revenue and earnings above consensus, raises its capex guidance again, and the stock falls anyway. In hindsight the logic was straightforward, investors stopped grading the quarter and started grading the spending plan attached to it. A beat driven by cloud and AI demand confirmed the revenue thesis, but another 10 or 20 billion dollars added to the annual capex budget pushed the payoff date further out and raised the stakes on the bet, and by late 2025, guidance hikes to capital budgets were about as likely to knock a stock down on earnings day as to lift it. The market was not doubting AI demand, it was applying a discount rate to a promise that kept growing. Investors who lived through prior capex supercycles, telecom in 1999, shale in 2014, recognized the reflex, enthusiasm for the buildout, followed by anxiety about the bill, in that order, every time.
The Accounting Mechanics That Took Over the Debate
Capex does not hit earnings when the cash leaves. It lands on the balance sheet as an asset and then charges the income statement gradually as depreciation over the asset's useful life. That lag created 2025's two defining wedges. First, free cash flow diverged from net income, companies reported record accounting profits while cash generation flattened or fell under the weight of data center construction, and analysts increasingly quoted free cash flow per share instead of earnings per share to strip the illusion. Second, depreciation became a forecast battleground, tens of billions of dollars of servers were beginning to depreciate, and small assumptions about useful life moved future earnings by whole percentage points. Bears argued that AI chips age quickly, state of the art one year and obsolete within a few, so depreciation schedules understated the true cost. Bulls answered that older chips keep serving inference workloads for years. The debate was unresolved in 2025 and remains the single biggest accounting question in big tech.
Capex is a promise the income statement pays for later. In a capex supercycle, earnings per share overstates health on the way up, and depreciation, the dullest line in accounting, quietly becomes the number the whole valuation argues about.
Winners of the Bill
Every capex boom pays someone, and 2025's earnings seasons sorted the sellers of shovels from the buyers. Nvidia remained the toll collector on the entire buildout, with data center revenue that dwarfed prior cycles' entire markets. The electrical and industrial complex, turbine makers, transformer manufacturers, power utilities with capacity to sell, reported the kind of backlogs and pricing power they had not seen in decades, and their stocks behaved accordingly, a rotation our macro coverage picked up from the power side. Memory, networking, and optics rode the same wave. The clean way to see 2025: the market did not sour on AI, it repriced who captures the cash first, and the answer, in hindsight, was the suppliers, paid today in cash, over the platforms, paid tomorrow in expected returns on nearly a trillion dollars of concrete and silicon.
What It Rhymes With, and What Is Different
The comparison every strategist reached for was telecom's fiber buildout of the late 1990s, enormous infrastructure spending justified by real but early demand, followed by a bust when capacity outran it. The parallel is imperfect in an important way, the 2025 spenders fund the buildout from operating cash flow of monopoly grade businesses rather than from debt, so the failure mode is wasted shareholder money and multiple compression rather than bankruptcy cascade. The parallel is uncomfortable in a different way, returns on this capital are not yet proven at scale, and the spending is concentrated in a handful of hands making correlated bets, the same concentration problem our piece on 2024's index math describes. In hindsight, 2025 was the year the market moved from asking whether AI is real, it is, to asking the harder investor question, what is the return on 700 billion dollars a year, and nobody's answer fit in an earnings call.
The Bottom Line
2025's earnings seasons were dominated not by whether big tech could beat estimates, it mostly did, but by the capital spending attached, roughly doubling from 2024 levels toward 660 to 690 billion dollars committed for 2026. Free cash flow split from earnings, depreciation assumptions became valuation arguments, beat and drop became a reaction pattern, and the suppliers of the buildout captured the cash while the platforms carried the promise. The demand was never the question by year end. The return on the trillion dollar bet still is.