The largest American technology companies are producing more cash than at any point in their history and returning less of it to shareholders than at any point in fifteen years. Reuters counted the first nine months of the year and found buybacks down about 83%. For every dollar of stock repurchased a year ago, roughly seventeen cents was repurchased this year.
The cash did not vanish. Microsoft, Amazon, Alphabet and Meta all raised capital spending in the same stretch, and their data centre and power contracts now run past 2027. The buyback was the mechanism that made that spending easier to own, and it has been switched off.
Here is the part that matters for anyone holding the shares. The multiple on these companies was built during a decade in which three things happened at once: revenue grew, margins expanded and the share count fell. The third was mechanical, cost nothing to produce, and added a low single-digit percentage to per-share earnings every year without a single operating improvement. Consensus models for next year still carry a version of that assumption. If buybacks stay near zero, the assumption is wrong, and the operating growth needed to reach the same earnings per share is higher than the models show.
The comparable that makes the point is 2022. The index fell about 19% that year and companies in it still set a record for buybacks, because cash flow was strong and capital spending was modest. That pattern held for a decade: capital intensity fell, cash rose, and the difference went back to shareholders. The present cycle inverts all three at once. Cash is higher, capital intensity is rising faster than at any point since the buildout of mobile networks, and the return of capital has stopped.
The fifteen-year low is worth putting in context, because it is a comparison with a period that includes 2020, when repurchases collapsed during the pandemic and recovered within two quarters. The current decline has run longer, and it coincides with record free cash flow rather than a shortage of it. That combination has no close parallel in the series, which is why the market has treated it as a valuation event rather than a cash-flow event, and why the operating numbers have not yet caught up with it.
The bull case is not foolish. A buyback executed at a premium multiple is a poor use of money, and the same dollar spent on capacity that customers have already contracted can earn more than the shares it retires. The spending is also, in principle, reversible: accelerators can be depreciated over five years, buildings can be leased rather than owned, and a company that pauses repurchases for two years has not renounced them. On this reading the pause is correct allocation, and the market is confusing a decision to invest with a refusal to return cash.
The bear case rests on the distance between a decision and a commitment. A power purchase agreement signed for fifteen years is not reversible. Neither is a multi-year reservation of advanced packaging capacity, nor a lease on a shell that must then be fitted out at a cost nobody has published. Capital spending that began as discretionary has been converted into contractual obligation, and the accounting follows with a lag of two to four years. Depreciation arrives after the cash has gone, in exactly the period when the returns are supposed to appear. Strip out the buyback and dilution from stock-based compensation lands directly in the share count at the same moment. Fund the gap with debt and the interest line rises while the equity that was being retired is now being issued.
The valuation question is where the two cases meet. These stocks still trade at premiums to the broad index, and those premiums were set when the business model was asset-light, licence-free and able to convert nearly all of its cash into either growth or repurchases. If the artificial intelligence buildout makes them permanently capital-intensive, the right comparison is no longer the software sector. It is the infrastructure sector, where heavy spending earns a lower multiple because the assets are long-lived and the pricing power is contractual. Nobody has yet repriced the first group as the second.
There is a further asymmetry that the sell side has mostly skipped. Buybacks support earnings per share hardest in the periods when depreciation is climbing fastest, which is precisely when a capital-intensive business needs the support. Cancelling them removes the offset at the moment the cost arrives, so reported numbers will look worse than underlying cash generation for several quarters. The market’s reaction to that gap will test patience rather than arithmetic, and patience is the input that has been in shortest supply in every previous capex cycle.
A note on measurement. Free cash flow, the figure the companies emphasise, subtracts capital spending from operating cash flow and leaves the buyback out entirely. A shareholder who wants to know what is left for them has to subtract the repurchase too. On that definition the four companies returned the smallest share of their cash in fifteen years while telling investors that free cash flow had never been stronger. Both statements are true. Only one of them describes what a holder receives.
The test is falsifiable and the data are public. Two figures, disclosed every quarter, both due within four quarters: net share count and depreciation per share. If the share count is flat or lower a year from now while capital spending stays at current levels, the buildout is being paid for by the business and the sceptics are wrong. If the share count rises and depreciation per share grows faster than revenue per share, the bill was handed to shareholders, and the premium multiple is still discounting an arrangement that ended this year.
