Hyperscaler Capex Passes Free Cash Flow as Debt Issuance Triples: Five companies that once funded the largest capital program in corporate history out of pocket now spend more than they earn, and the gap is being closed in the bond market.
The arithmetic flipped quietly. Microsoft, Alphabet, Amazon, Meta and Oracle grew operating cash flow 29% year over year to a record $542.9bn in the first nine months of 2026 — and still ran a $14bn free-cash-flow deficit, because capital spending on data centers and chips doubled to $556.9bn over the same stretch. A year earlier, the identical group posted a $143.5bn surplus.
That is a $157.5bn swing in nine months. It did not come from a collapse in earnings. It came from capex absorbing 94% of operating cash flow in 2026-27, against roughly 40% in 2023.
Capex absorbed 94% of operating cash flow in 2026-27, against roughly 40% in 2023 — a $157.5bn swing in nine months.
The tell is not the capex line. It is the buyback line.
Share repurchases across the group fell 83% to $13.6bn. Alphabet and Meta bought back no shares at all in 2026, after deploying $66.5bn between them the prior year. Total payouts dropped to 9.3% of operating cash flow from 27.4%.
Those are not the numbers of a company that thinks it is spending opportunistically. They are the numbers of a company that has repriced its own equity against a bond coupon.
The quarterly prints show where the squeeze sits. Alphabet reportedly ran negative free cash flow of $9.1bn, its second consecutive quarter in the red, with capex at $51.4bn for the quarter and guidance up to $205bn for the year. Amazon's trailing-twelve-month free cash flow was reportedly negative $42.4bn, down from positive $14.8bn a year earlier, against roughly $220bn of planned 2026 capex. Meta reportedly ran a $2.1bn quarterly deficit while guiding to as much as $145bn.
Microsoft was reportedly the lone exception — $18.6bn of positive quarterly free cash flow alongside $45.2bn of capex — and even that came with a cut, its calendar 2026 forecast reportedly trimmed to about $170bn from $175bn because third-party capacity deliveries slipped.
Last week the finding was that the buildout had shifted from cash flow to credit. This week the shift has a size and a shape. Full-year 2026 spending across the group lands between $803bn and $828bn, with the four largest alone reportedly guiding to more than $740bn. And the financing has scaled with it: incremental annual debt rose from 9% of capex in FY24 to 32% by mid-2026, while first-half issuance of nearly $200bn already exceeded the $136bn the five borrowed in all of 2025. Year to date, hyperscalers have issued $219bn of investment-grade bonds, including $62bn equivalent in euros, sterling, Canadian dollars, Swiss francs and yen.
This is not distress. It is a deliberate maturity transformation, and the tenor choice is the evidence.
Across 2025 and 2026 supply, hyperscalers placed 41% of issuance beyond 15 years, against 16% for the ICE BofA US Corporate Index. In the over-26-year bucket their weighting is 31.2% versus 13.0% for the broad market — nearly two and a half times.
Alphabet's £1bn century bond drew £9.5bn of orders at 120bp over gilts, the first such deal from a technology company since Motorola in 1997. Amazon has extended dollar paper to 2076. Meta's and Oracle's longest runs to 2066. Alphabet also printed a debut A$5.5bn deal with a 20-year tranche and the largest-ever corporate bond in Swiss francs.
The duration matches the asset. A data center is a 30-to-50-year cash-flow stream, so funding it at 30 years is coherent rather than reckless. Microsoft is the outlier here, with a materially shorter issuance profile and — so far — no terming out of its 2026-27 capex in the bond market. The largest single source of long-end supply in this cycle has not yet arrived.
What has arrived is spread widening, and the issuers have shrugged at it. Hyperscaler spreads have widened about 30bp year to date, against 2bp for the broader USD investment-grade index. The arithmetic still clears comfortably: a 20-year average return on invested capital of 17.75% against funding at 5.5% to 7%. When the 30-year Treasury yield hit its highest level since 2007 in August, hyperscalers absorbed a comparable move in their own funding costs and changed nothing.
Demand has held, with one crack. The 30-year tranches were covered 4.6x for Alphabet, 4.5x for Oracle and 3.6x for Meta, all ahead of the 2.98x average oversubscription across the broader US investment-grade primary market this year. Amazon's July deal, by contrast, caught the market off guard and drew the smallest book of the four, at 2.5x.
That single 2.5x is the whole point. Four issuers are absorbing duration capacity that would otherwise have gone into governments, and each $5-6bn of 30-year paper competes directly with the 30-year Treasury for the same pension and insurance bid. The 30-year now yields 5.25%, roughly 40bp above where it started the year. Hyperscaler supply is adding to long-end pressure rather than causing it, but it is no longer a rounding error in that market.
The consequence is not a chip shortage and not a permitting fight. It is that the binding constraint on data-center timelines has moved to a desk where a credit committee decides whether to buy paper maturing in 2076 from an issuer that just stopped buying its own stock.
The exposure is not evenly distributed. The equipment chain gets paid on order; it is not waiting on the bond market or a regulator. Capacity operators are the opposite — their revenue begins at energization, and their balance sheets cannot idle while a book builds. Those are the names where rate sensitivity is real rather than theoretical.
Reported leverage also understates it, which is where the vendor-credit work picks up: disclosed data-center lease obligations run around $1.4trn, roughly $1.1trn of which is not yet recognized on balance sheet because it is booked only when the leases commence.
The test ahead is narrow and specific. Not whether the spending was justified — the ROIC arithmetic says it clears. The test is whether free cash flow recovers after 2027, as capex moderates and the energized capacity finally starts billing. Every bond sold to 2076 is a bet on that recovery arriving inside the window the coupon assumes.