Charted: Hyperscalers Will Fund a Third of AI Capex With Debt by 2026

The build-out now has a balance-sheet story

For three years the AI infrastructure debate lived almost entirely in equity and product narratives: GPU shortages, campus megawatts, token demand, and ever-larger capex dollar totals. That framing is incomplete. Once hyperscaler capital expenditure climbs toward $750 billion in 2026 against operating cash flow that only barely clears the same neighborhood — and after buybacks and dividends are paid — the residual must be financed. Credit markets, primary equity, leases, and structured vehicles are no longer footnotes. They are co-pricing the cycle.

Goldman Sachs credit strategists led by Amanda Lynam put a clean number on the shift: investment-grade bond issuance from the five large hyperscalers (Amazon, Alphabet, Meta, Microsoft, Oracle) totaled about $108 billion in 2025 — roughly 26% of capex. Through early July 2026, four of those names alone had already printed about $194 billion, and Goldman’s full-year path points to roughly $250 billion of IG issuance (~33% of capex), rising toward $400 billion and a ~35% debt-funded share in 2027. FactSet’s July 2026 note rhymes with the same geometry: incremental annual debt jumped from about 9% of capex in FY24 to roughly 32% on a trailing twelve-month basis by mid-2026.

The interactive dashboard above is a financing map — not another spend total. Toggle 2025 / 2026 / 2027, switch the IG credit / funding mix / equity-ETF lens, and read issuance, debt share, deal sizes, new-issue spreads, and public-market flows side by side.

What the 2025–27 path actually says

Two readings matter. First, absolute supply into IG credit is exploding: a single mid-year 2026 print already exceeds the entire 2025 calendar total for the Amazon–Alphabet–Meta–Oracle subset. Second, the share of capex funded with debt is climbing into the low-to-mid thirties — still not a leveraged-buyout story, but a regime change from the cash-funded cloud decade. Pair that with our capex intensity research: Meta and Oracle already reinvest mid-thirties percentages of revenue into PP&E. Debt does not lower intensity; it defers the cash-flow reckoning.

Goldman also notes that 2026 operating cash flow for the cohort (~$778B) can still sit near or slightly above aggregate capex (~$750B) — until shareholder returns, working capital, and cash M&A eat the buffer. That is why “they generate enough OCF” and “they need to borrow a third of the build” are both true at once.

Deal flow: who printed, and how big

The 2025 second-half wave rewrote US investment-grade league tables. Meta’s $30 billion October 2025 financing was among the year’s largest corporate bond deals anywhere. Oracle, Alphabet, and Amazon followed with multi-ten-billion packages; MUFG’s year-end note observed that four of the five biggest US IG financings in 2H 2025 were hyperscalers. Public tech IG issuance as a broader set surpassed $200 billion in 2025 in some dealer tallies, with hyperscalers supplying a large share of the late-year surge.

2026 opened even louder. Amazon’s March package alone combined roughly $37 billion of USD bonds with a €16.8 billion euro tranche. Subsequent Amazon prints (including a later ~$25 billion deal that was still oversubscribed, but at a lower cover ratio than March) showed the market’s appetite cooling from “infinite bid” to “selective.” Alphabet’s ~$84.75 billion June 2026 equity raise — not a bond — is the complementary signal: when primary credit books thicken and rating headroom tightens, issuers reopen the equity channel.

Oracle is the credit-market stress case inside the cohort. S&P’s July 2026 cut to BBB- (still investment grade) after surging capex, negative free cash flow, and customer concentration is exactly the differentiation FactSet flagged: the five names no longer trade as a homogeneous “cash-rich tech” bloc.

Spreads, concessions, and index weight

Supply is not free. Reuters’ LSEG-based analysis (via CNA, late July 2026) shows median new-issue spreads for Amazon, Alphabet, Meta, and Oracle widening across the curve: 2–4 year tenors from about 30 to 40 basis points, 5–7 year from 50 to 60, and 20-year-plus from roughly 108.5 to 118. Secondary performance worsened too — 78 of 91 comparable 2026 hyperscaler bonds were trading wider than at issuance by late July, with a median increase near 22 bp. Deal-level concessions rose as cover ratios fell. Investors are still buying; they are demanding more compensation to absorb the calendar.

On the index side, tech’s weight in broad US corporate IG benchmarks has drifted toward roughly 10% in dealer commentary (up from the high single digits), while the hyperscaler subset’s weight inside Bloomberg US Corporate IG roughly doubled over the year to April 2026 in Breckinridge’s reading (about 2.2% to 4.1% for Meta, Alphabet, Amazon, and Oracle collectively). Concentration risk is no longer a boutique credit-committee talking point. It is a benchmark construction problem: passive IG portfolios are mechanically more exposed to a handful of AI builders.

Funding mix: bonds are not the only residual

The dashboard’s funding mix lens keeps IG bonds in context. Even in 2026–27 research paths, operating cash flow remains the largest single channel — often modeled near half of the stack. The residual splits among:

1. Investment-grade corporate bondsthe transparent, index-eligible wave Goldman sizes at roughly a third of capex.
2. Primary equityback in the mix after years of buybacks-only capital returns (Alphabet’s June raise; Oracle’s planned debt-plus-equity FY27 package).
3. Leases, ABS/CMBS, project finance, joint ventures, and private creditthe off-index valves Goldman and Breckinridge both flag as likely to grow if traditional IG books saturate.

That third bucket is easy to undercount in bond-only dashboards. GPU capacity purchased under operating leases, campus-level project vehicles, and power-side financing can move economics off the hyperscaler’s PP&E and debt lines without shrinking the physical build. Treat the IG share as a lower bound on external financing intensity, not a complete map of every dollar of AI infrastructure capital.

The equity-market side: ETFs still fund the narrative

Issuers raise capital in primary markets. Public investors express the AI trade in secondary markets — and there the story is paradoxically concentrated in broad proxies. FactSet’s 2025 US ETF summary put ~$43.5 billion of net inflows into thematic ETFs after outflows the prior year, with robotics/AI among the leaders. Yet Invesco QQQ alone took in about $21.7 billion, and wealth-management commentary still treats the Nasdaq-100 — with Mag 7 near 40% weight — as the default AI vehicle. Dedicated robotics/AI funds remain a satellite sleeve.

That matters for financing because equity valuation support and primary equity capacity are linked. A market that funds the AI story through QQQ and SOXX can still choke on simultaneous mega-IPO supply (OpenAI / Anthropic expectations) and mega-issuer secondary raises. Hyperscalers that raise equity early, as FactSet argued, are partly booking capacity before that collision.

Caveats and methodology

1. Cohort and scope differ by source. Goldman’s IG issuance path is global IG-rated bonds for the five hyperscalers. BofA’s five-company US series and Reuters’ four-name YTD tally will not match dollar-for-dollar. We label the Goldman path as the spine and annotate deal-level Reuters/LSEG prints separately.
2. Capex ≠ AI-only spend. Company gross PP&E mixes warehouses, offices, network, and accelerators. Debt-funded share of capex is not identical to debt-funded share of AI infrastructure. See the spend-scope map and Goldman Tracking Trillions breakdown for perimeter discipline.
3. Funding-mix pie is research synthesis, not a single audited ledger. Channel shares are constructed to be consistent with Goldman’s debt-share path and disclosed equity raises; structured/lease slices are directional.
4. 2026 YTD is not a run-rate. Front-loaded issuance can overstate full-year annualization if books pause after rating actions or spread spikes.
5. ETF flows fund secondary ownership, not issuer capex directly. They appear here as the public-markets sentiment and capacity channel alongside primary equitynot as a substitute for bond proceeds.
6. Ratings and CDS (especially Oracle) can gap faster than cash-bond indices; spread charts based on new-issue medians understate idiosyncratic stress.

Primary sources: Goldman Sachs credit strategy as reported in financial press (mid-2026), FactSet hyperscaler external-financing analysis (July 2026), Reuters/LSEG deal and spread tallies via CNA, MUFG “Financing the AI supercycle” (December 2025), Breckinridge balance-sheet commentary, and FactSet / ETF.com 2025 flow summaries.

What to watch next

Three coincident signals will tell you whether the financing regime stabilizes or tightens further: (1) full-year 2026 IG issuance versus Goldman’s ~$250B path — and whether Microsoft remains a light issuer relative to peers; (2) new-issue concessions and cover ratios on the next mega deal after Oracle’s downgrade; (3) equity vs debt mix — another Alphabet-scale primary raise would confirm that credit saturation is binding. If debt share settles near a third while spreads stop widening, markets will treat leveraged AI infrastructure as a new steady state. If issuance keeps climbing toward the $400B / 35% 2027 sketch while secondary bonds gap wider, the constraint stops being GPU supply and starts being credit market plumbing. Either way, the question is no longer only how large the build is — it is who funds it, at what price, and on whose balance sheet.