Meta's $20 Billion Bond Is the AI Buildout's Honest Price Tag — and Its Own Project Debt Trades at 92 Cents

Meta is funding one year of AI capital spending with every instrument the market offers. On November 2 it began marketing at least $20 billion of investment-grade bonds — dollar tranches of five to 30 years, plus a first euro-denominated issue running from just over four years to just over 38 years. Six months earlier it had signed an underwriting agreement for $25.0 billion of senior notes, and a week before this launch it priced $25.0 billion of Class A stock and $12.5 billion of mandatory convertible preferred. The equity market is no longer the marginal buyer of Meta, and the bond market has taken over the job.

The bond and the equity now price AI capex at almost exactly the same place. On October 26, Meta sold three-year mandatory convertible preferred at a 6.50% dividend on its $1,000 liquidation preference — $3.25 a year on each $50.00 depositary share — while it pays 6.450% on bonds due in 2066. The equity-linked paper prices no cheaper than forty-year straight debt, which is the first clean market read on what an AI buildout costs now that the credit market rather than the equity market is writing the cheque.

Third-quarter results, reported October 21, were the moment the buildout stopped being self-funding. Revenue grew 26% to $64.60 billion. Capital spending was $34.91 billion and operating cash flow was $32.79 billion, so free cash flow was negative $2.12 billion against positive $10.63 billion a year earlier. Operating income fell 37% to $12.99 billion after a $9.89 billion legal charge, and diluted earnings per share were $3.83. Meta spent more on servers and data centers in the quarter than its operations generated.

The ratio did the damage. Capital spending consumed 59% of operating cash flow in the first quarter of 2026, 95% in the second and 103% in the third; across all of 2025 it never left the band between 54% and 65%. Guidance for 2026 puts capital spending at $135–145 billion, on a basis that includes principal payments on finance leases, against $69.7 billion of property and equipment purchases in all of 2025. A $2.12 billion quarterly shortfall becomes a funding problem when the spending plan is $140 billion a year.

The obligations are not remote. Of the $454.52 billion of non-cancelable commitments, $121.4 billion falls due in 2027 and $287.3 billion after that, largely for third-party cloud capacity, servers and network equipment, energy service agreements for data centers, and Reality Labs hardware. The $355.73 billion of leases not yet commenced — data centers, energy infrastructure, colocation and network equipment, with terms running up to 30 years — begin landing from the rest of 2026 through 2036, and a contract that cannot be cancelled is a debt in everything but its accounting.

S&P moved Meta's outlook from stable to negative on October 30, 2026, citing elevated capital spending and negative free cash flow, and said its own adjusted leverage measure could rise above 1.0x — the threshold at which it would downgrade. Moody's still has the company at Aa3 with a stable outlook, and told investors in its most recent rating action that pro forma gross adjusted leverage, including lease liabilities, should land around 0.6–0.7x at the end of 2026, and one agency sees headroom while the other sees a line being crossed.

Meta's own project debt is the sharpest evidence available. In April and May 2026 the bonds backing the Hyperion campus in Richland Parish, Louisiana were among the best paper in the AI buildout. By mid-October they had slipped from above par in early July to just under 92 cents on the dollar. CoreWeave's bonds fell below 80 cents at yields near 15%, and QTS's Project Excalibur dropped below 88 cents: the market is not repricing Meta's balance sheet, it is repricing Meta's contracted rent.

The cost side is visible in the Texas campus. BlackRock raised more than $12 billion of bonds for Meta's El Paso project in July 2026, pricing roughly 40 basis points wider than the comparable Hyperion deal at an interest rate around 7.5%. Oracle's roughly $18 billion of Project Jupiter loans in New Mexico moved into stressed territory after power permitting delays, and every one of these instruments is a claim on a hyperscaler's promise to pay rather than on the hyperscaler's balance sheet.

Meta's own filing explains why project bonds can fall while the corporate credit does not. Meta holds a 20% membership interest in each of the two ventures — the Louisiana project entered in October 2025 and the Texas project in July 2026 — and neither is consolidated; Meta is not the primary beneficiary of either. Its equity-method carrying value in the two was $4.73 billion at September 30, 2026, against $18.48 billion of lease commitments and $3.70 billion of future funding commitments to them, so Meta moved roughly $41 billion of data centers off its balance sheet while bondholders took on the asset risk.

Goldman Sachs estimates that roughly 33% of hyperscaler capital expenditure was debt-financed in 2026, against 27% in 2025, and expects the share to peak at 35% in 2027. Hyperscaler bond issuance went from $108 billion in 2025 to $194 billion in the first eight months of 2026, with about $250 billion expected for the full year, and AI-related issuance is now roughly 18% of US investment-grade supply, up from 7% in 2025 and 1% in 2024. It accounts for 40% of all investment-grade issuance with a maturity of 15 years or more.

That is the transmission chain, and it runs through structured credit before it reaches the corporate bond. Barclays Research puts outstanding data-centre securitisation at $61 billion year-to-date in 2026, up from $4 billion in 2020, and estimates that $180 billion by the end of 2028 would make it roughly a fifth of the market. Data-centre securitisation alone ran at $21.9 billion year-to-date as of November 2, 7.0% ahead of 2025, with sell-side desks projecting $30–40 billion a year of gross supply through 2027. Spreads on data-centre ABS have been flattening or widening even as broader structured credit has held, so the tenant-credit question gets asked first in the ABS and private-credit market, and that is where the marginal price-setter for AI capex now sits.

The buildout chain has been re-sorted by credit rather than by equity. The supplier of the compute is the strongest balance sheet in it, the tenant is the weakest, and the intermediaries that package the tenant's promise into tradable paper are the ones now earning the spread: value has moved from the chip maker to the arranger, and risk has moved from the chip maker to the bondholder.

Meta's November bond is not really a financing event. The company has $82.45 billion of cash and marketable securities and quarterly revenue of $64.60 billion growing at 26%. It does not need the money. The bond is a price-discovery event: by asking the investment-grade market to price forty years of AI buildout risk in public, Meta converts a private and unobservable problem into a quoted one, and the quote is the number every project lender, private-credit fund and structured-finance investor will mark their own exposure against, so the marginal price-setter for the AI buildout has moved from shareholders to bondholders and the first price is already on the tape.