Record earnings, shrinking free cash flow, and the accounting gap between them
S&P 500 companies are expected to post year-over-year earnings growth of roughly 38% for the second quarter, and the full-year 2026 consensus has been revised up from about 14% in February to north of 23%. Analysts normally trim estimates at this point in the calendar. This year they are doing the opposite.
Part of the explanation is less exciting than the headline. A large share of what the artificial intelligence buildout has cost so far has not reached the income statement yet.
Exhibit 1. Capital spending is closing on cash generation.
Microsoft, Amazon, Alphabet, Meta and Oracle have grown capital spending at roughly 70% a year since GPT-4 shipped, against operating cash flow growth of about 23%. Those two trends converge in the third quarter of this year.
Exhibit 2. Capex as a share of operating cash flow.
In 2023, the group spent roughly 41 cents of capital for every dollar of operating cash flow. Last year, it was 68 cents. On trend, it reaches the mid-90s this year. PIMCO independently arrived at approximately 94% for 2026 and 2027, providing useful corroboration. Above a dollar, the buildout is no longer self-funded.
In practice, external financing is already playing a larger role. Incremental annual debt has risen from 9% of capital spending in fiscal 2024 to 32% over the twelve months through mid-2026. Alphabet, a company that has generated cash at scale for two decades, announced an $84.75 billion equity capital raise in June to fund AI infrastructure and compute, reported as the largest in US corporate history. Roughly $35 billion was immediately priced, Berkshire Hathaway took $10 billion in a private placement, and a $40 billion at-the-market program was left to run from the third quarter. Moody’s calculates that lease commitments not yet on hyperscaler balance sheets already equal 113% of their adjusted debt. By early May the five accounted for more than 15% of all US investment grade issuance for the year, a concentration the Bank of England thought worth flagging in its July Financial Stability Report.
None of this is a solvency question. All five are profitable and increasingly so. It is a timing question.
Exhibit 3. The bill arrives later than the spending.
Cash leaves at once. Cost returns gradually. Exhibit 3 sets cumulative capital spending since 2022 against the portion that would have been recognized as depreciation expense by now under a straight-line schedule. On a five-year assumed asset life, roughly $1.1 trillion of that spending will remain capitalized on balance sheets at year end rather than having passed through earnings.
Measured over 2026 alone rather than cumulatively, that gap between capital spending and the depreciation recognized on it runs north of $400 billion on our numbers. An RIA Advisors analysis circulating this month puts the same annual figure at $549 billion. Ours is lower by design. Applying a single five-year life to all capital spending depreciates the longer-lived buildings and power infrastructure inside it faster than reality does, making our estimate conservative.
The sensitivity that matters is asset life. Moving the assumed useful life of compute from five or six years down to three or four raises depreciation charges materially, and compute is now roughly 60% of hyperscaler capital spending against 43% in 2022. Current market evidence is consistent with the longer schedules, since contracted leasing rates and secondary prices for chips launched three to six years ago are holding up well. That evidence is a fact about compute scarcity today rather than a fact about 2029.
Exhibit 4. One transition, five different stages.
The aggregate also conceals a good deal of dispersion. Oracle passed the point where capital spending exceeds operating cash flow some time ago. Amazon reached it around the middle of this year. Based on current trends, Alphabet is not projected there until the first quarter of 2027, Meta until the third quarter of that year, and Microsoft not until late 2028. Anyone holding the group through a broad index holds five companies at five different stages of the same transition, which is worth knowing before treating them as a single position.
For allocators, the useful reframe is that this is a question about how durable today’s reported earnings are rather than a question about whether artificial intelligence works. If current earnings are flattered by costs that will be recognized in later periods, those earnings may prove less durable than the prevailing multiple assumes. That is not an argument for being absent. It may be useful to consider position sizing and diversification among assets that do not depend on the same spending cycle or accounting assumptions.
SOURCES AND METHODOLOGY
Quarterly capital spending and operating cash flow for Microsoft, Amazon, Alphabet, Meta and Oracle are from Epoch AI, “Hyperscaler capex is on trend to outpace their cash inflows by the end of 2026,” June 16, 2026, used under CC BY 4.0. Epoch parses those figures directly from SEC EDGAR 10-Q and 10-K XBRL tags. Company-level crossover dates in Exhibit 4 are Epoch’s exponential trend extrapolations fitted from the second quarter of 2023 through the first quarter of 2026, and are extrapolations of existing trends rather than forecasts.
Exhibit 2 and Exhibit 3 are Xtollo Investment Partners calculations from the Epoch series. Exhibit 3 applies straight-line depreciation to capital spending from the quarter incurred and excludes the runoff of asset vintages placed in service before 2022. It applies a single assumed life to all capital spending, including longer lived assets such as buildings and power infrastructure, which recognizes their cost faster than reality and therefore understates rather than overstates the gap shown. It illustrates the timing effect and is not a forecast of reported depreciation. Financing statistics are from FactSet Insight, Moody’s via secondary reporting, the Bank of England July 2026 Financial Stability Report, and Alphabet’s Form 8-K of June 2, 2026 and Form 10-Q for the quarter ended June 30, 2026. Earnings growth expectations are from FactSet.
IMPORTANT DISCLOSURES
This material is provided for informational and educational purposes only and does not constitute investment, legal, tax or accounting advice, nor an offer or solicitation to buy or sell any security. Company names appear solely to illustrate the accounting and cash flow dynamics discussed and are not recommendations to buy, sell or hold any security. Xtollo Investment Partners and its affiliates, employees and clients may hold positions in the securities reference.
Statements regarding future periods reflect historical trend analysis and modeled assumptions, including assumed asset useful lives, and are subject to significant uncertainty. Actual results will differ materially from those discussed. Trend-based estimates are sensitive to the fit window, model assumptions, seasonality in operating cash flow, changes in capital spending, finance conditions, and other factors, and should not be viewed as predictions of company results.
Dates and crossover points shown in Exhibit 4 are model-based estimates derived from historical trends and current assumptions. Actual timing and outcomes may differ materially. Third-party data is believed reliable but has not been independently verified. Past performance is not indicative of future results.
Xtollo Investment Partners, Austin, Texas. Document XT20260730.
