Last reviewed 15 July 2026 · our own structural assessment, not a price target
The thesis
The AI buildout is increasingly funded by debt and private credit, with circular vendor-financing dynamics and a maturity question: short-lived compute hardware (GPUs and servers, roughly 3-6 year useful lives) sits inside data-centre projects financed with much longer-dated debt and leases (15-20 years) matched to the long-lived facilities and land, not the chips. If AI revenues fail to service that debt, the fragility could force a buildout slowdown — the risk overlay sitting under every other thesis on the board.
The case for
Scale and structure both point to fragility. Consensus estimates put the five largest US hyperscalers (Microsoft, Alphabet, Amazon, Meta, Oracle) on track for ~$690B of capex in 2026 (PIMCO, 22 May 2026, citing Bloomberg / S&P Capital IQ consensus) — about 4.3x the $162B they spent in 2022 (2022 actuals compiled from SEC filings by Epoch AI) — and PIMCO projects capex absorbing 94% of their operating cash flow in 2026-27, versus 40% in 2023 (PIMCO, 22 May 2026, using consensus estimates). Funding is shifting to debt: the five hyperscalers issued a record ~$121B of investment-grade bonds in 2025, roughly 4.3x their ~$28B five-year average (reported via Quartz / BofA Global Research, 2026), and Morgan Stanley projects total AI-linked debt issuance — hyperscalers, chipmakers and data-centre developers, in dollar and non-dollar markets — to reach nearly $570B in 2026, up from ~$236B sold by end-May (Morgan Stanley, via Reuters, 10 Jun 2026). Private-credit lending to AI-related sectors has grown from near-zero to over $200B outstanding globally (BIS Bulletin 120, 7 Jan 2026, using PitchBook data). The circularity is explicit — Nvidia holds ~11.5% of CoreWeave's Class A stock (Nvidia Schedule 13G/A, filed Jan 2026), up from ~7% at the March 2025 IPO, and in September 2025 announced an intention to invest up to $100B in OpenAI (a letter of intent, not a binding commitment — subsequently reported to be scaled back to a ~$30B equity stake in February 2026, per news reports) tied to Nvidia GPU deployments. And concrete stress has arrived: S&P cut Oracle to BBB- on 9 July 2026, citing an AI-capex-driven free-cash-flow deficit.
The case against
Fragility is not a crash. Hyperscalers still hold large liquidity buffers — Meta reported $81B of cash, equivalents and marketable securities at 31 March 2026 — though the binding constraint is free cash flow, which is compressing (Alphabet's Q1 2026 free cash flow fell 47% YoY to $10.1B as capex more than doubled; Amazon's trailing free cash flow collapsed toward ~$1.2B). Offtake and lease structures (e.g. Meta's Hyperion SPV) genuinely move risk off balance sheets — redistributing rather than eliminating it. Demand is real and cloud revenue is growing fast (AWS +28% YoY in Q1 2026; Microsoft's Azure +40% in its quarter ended 30 Sept 2025). The BIS itself judges financial-stability risks as "appearing moderate", flagging circular financing and off-balance-sheet leverage as items to monitor contingent on AI earnings delivering (BIS Bulletin 120, 7 Jan 2026) — not an active systemic warning.
What would change our mind
Re-scored 55 -> 58 on 2026-07-14 after a fresh source and counter-case review, which corrected several stats (five not six hyperscalers; ~4.3x not 6x; total AI-linked debt reframed to Morgan Stanley's ~$570B 2026 estimate rather than a hyperscaler-only ~$400B figure that could not be sourced; Nvidia's CoreWeave stake ~11.5% of Class A per its Jan 2026 13G/A; an unverifiable "7-10% of CMBS/ABS" claim removed; the BIS "systemic risk" language softened to match the bulletin's "appear moderate"). Nudged up because the fragility is now producing observable, dated stress — the S&P Oracle downgrade (9 Jul 2026) is a concrete credit event tied to AI capex, and 2026 bond issuance is running ahead of forecasts. Held below 60 as the deliberate bear call: no default has occurred and liquidity buffers remain large.
Where the exposure sits
Mostly a risk overlay on the power and equipment theses: if it triggers, it weakens demand for both.
Show our work — the evidence
The key evidence behind the score, dated and sourced — supporting and disconfirming. This is how the
conviction is built, and what we re-check as it moves.
AgainstBIS judges financial-stability risks from the AI boom as "appearing moderate", flagging circular financing and off-balance-sheet leverage as items to monitor, contingent on AI firms meeting earnings expectations.Official dataBIS Bulletin 120 · 7 Jan 2026
ForS&P downgraded Oracle to BBB- (one notch above junk), citing a widening AI-capex-driven free-cash-flow deficit — a concrete credit-stress event tied to the buildout.Official dataS&P Global Ratings · 9 July 2026
ForPIMCO: hyperscaler capex to absorb 94% of operating cash flow in 2026–27, versus 40% in 2023.Analyst estimatePIMCO — AI Credit Expansion · 22 May 2026
ForPrivate-credit lending to AI has grown from near-zero to over $200B.Official dataBIS Bulletin 120 · 7 Jan 2026
Tags mark how each fact is sourced: Company disclosure (the company said it) · Official data
(a government or agency figure) · Analyst estimate (a third party's number) · Reported (via the press).