AI Infrastructure Debt Repricing and Hyperscaler Capex Signals a New Financing Regime

The bond market has decided that AI infrastructure is investment-grade. The question worth putting to the board is not whether it is right, but what the firms that financed first now own that the firms that waited cannot replicate. Oracle's planned $40 billion FY2027 raise and the $121 billion that five hyperscalers placed into the U.S. corporate bond market in 2025 [1] are not just a financing story. They are the moment a buildout becomes a barrier to entry.

The repricing has already happened. The consequences have not.

Five hyperscalers, Amazon, Alphabet, Meta, Microsoft, Oracle, sold roughly $121 billion of U.S. corporate bonds in 2025. The 2020-2024 average was around $28 billion a year. That is a more than fourfold step-up in a single year, with Meta's October 2025 deal of $30 billion ranking as the largest non-merger high-grade bond sale on record, and Amazon's March 2026 sale of $37 billion across 11 tranches drawing orders close to four times the amount on offer. [2] Barclays projects total U.S. investment-grade issuance reaches $2.46 trillion in 2026, up 11.8% year-on-year, with hyperscalers the marginal driver. [3]

What matters for the next twelve months is not the volume but the duration. Alphabet's February 2026 global offering included a 100-year "century" bond. [4] Hyperscalers are matching long-dated paper to multi-decade asset lives, locking in cost of capital at 10-year Treasury yields above 4.5% and 30-year above 5%. [5] If rates fall materially, the first movers have made an expensive mistake. If rates stay here or grind higher, the first movers have locked competitors out of the buildout for a generation. That is the bet the board needs to debate, and most boards are not debating it.

Oracle as test case

Oracle's FY2026 capex landed at roughly $55.7 billion. FY2027 guidance reaches $95 billion, with up to $25 billion of that expected to be repaid by customers through take-or-pay contracts. [6] Clay Magouyrk told the market that Q1 FY27 delivery is approaching one gigawatt, "nearly the same capacity as we've delivered in the previous four quarters combined." [7]

The stock fell roughly 8.9% in after-hours on the disclosure. [8] Equity holders are pricing the take-or-pay receivable at a discount, which means they doubt the counterparty. Oracle's major data-centre customers include OpenAI and Meta, and the WSJ has reported that banks struggled to syndicate large loans tied to Oracle-leased sites. [9] The bond market is buying Oracle paper at investment-grade pricing while the loan market is choking on its underlying customer concentration. Both cannot be right.

The TS Lombard observation tightens this further: half of the combined order backlog of Microsoft, Oracle, Google, and Amazon comes from OpenAI and Anthropic. [10] Neither generates the cash to honour those commitments today. The investment-grade rating on hyperscaler paper is, in part, a derivative bet on two private companies whose revenue trajectory remains the central speculative question of this cycle.

The equity-financing pivot

For fifteen years, the marginal buyer of large-cap tech equity was the company itself. Buybacks supported multiples. That pillar is being dismantled. Big tech is reportedly preparing dilutive equity offerings to fund AI capex, with Alphabet's $80 billion plan and Meta potentially following, alongside the debt wave. [11] (Both remain reported plans, not filings.)

If the capital structure of the five most heavily index-weighted equities in the world shifts from net-repurchaser to net-issuer, the passive flow that has supported the S&P 500's concentration trade weakens at the same moment the credit market is digesting record supply from the same names. That is one cyclical exposure stacked on another, and most asset-allocation models do not capture the correlation. Broadcom's $444 billion two-day market-cap loss on a tepid outlook is a small preview of what violent repricing looks like when AI infrastructure forecasts slip. [12]

The downstream beneficiaries tell the same story from the other side. Ciena's CEO told the Q2 call that since March, "the largest hyperscalers have increased their 2026 capital expenditures with indications of continued expansion into 2027 and beyond." [13] Dell'Oro projects data-centre interconnect optical spending grew 40% in 2025 and grows another 40% in 2026. [14] The supply chain is now dependent on hyperscaler debt-funded capex continuing through at least 2027. If the bond market loses its appetite, the cascading revenue gap runs from Ciena and Coherent through to the smaller optical and power names re-rated on the AI tailwind.

The counter-case

The strongest opposing view is that none of this is a regime shift. It is liquidity. Stacie Ware at Fidelity, co-managing the Total Bond Fund, told subscribers in the June 3, 2026 midyear outlook that investors "aren't well compensated to own corporate securities right now," and Fidelity is steering core bond portfolios away from hyperscaler AI debt entirely. [15] The cost of insuring hyperscaler debt through credit default swaps has climbed since mid-2025, even as spreads on the cash bonds remain thin, meaning the marginal investor is paying for protection rather than trusting the rating. [16] DoubleLine and Oaktree are reported to be positioning in instruments that perform if the AI boom turns into a credit bust, with a DoubleLine portfolio manager warning that bond valuations will reach frothy levels "in the coming months or years." [17]

The counter-case is correct that spreads are thin and the underlying credit story is partly circular: hyperscalers book rising valuations on their OpenAI and Anthropic equity stakes as income, which flatters the same earnings figures that support their debt ratings. [18] But it underestimates one thing. Even if today's spreads are wrong, the bonds have already been issued and the proceeds already spent on land, power contracts, accelerator hardware, and long-lead-time switchgear. A 2027 repricing of hyperscaler credit does not unwind the assets that the 2025-26 issuance has already locked in. The first-mover advantage is physical, not financial. The credit market may revolt; the gigawatts will not move.

The honest position is narrower than either bulls or bears suggest. The financing regime may prove fragile. The infrastructure footprint it has already paid for will not.

What to watch

1. Rating agency action by end of Q1 2027. Moody's, S&P or Fitch placing any of the top five hyperscalers on negative outlook citing AI-related leverage would mark the first formal break in the investment-grade story. Absence of any such action through Q1 is itself a signal that the cash-flow models are holding.

2. Oracle FY2027 customer prepayment realisation. Oracle has guided up to $25 billion of FY2027 capex to be repaid by customers under take-or-pay terms. If actual prepayment receipts at H1 FY2027 reporting come in materially below the run-rate implied by that guidance, the customer-concentration thesis is being validated and the bond market should reprice the issuer.

3. Bank syndication outcomes on the next Oracle-linked or hyperscaler-linked data-centre loan package. A failed or materially repriced syndication, particularly one reported in WSJ or Bloomberg before March 2027, would confirm that loan markets and bond markets have diverged on the same credit. That divergence cannot persist; whichever market moves second will move violently.