AI capex / financing strain Credit-funded and circular — and the circularity is now measurable inside the vendors' own revenue lines

App Economy: the market has "stopped paying for backlog and started pricing the cost to deliver it." Oracle (Jun 13) beat and grew RPO +363% to $638B, yet fell 12% as the new CFO guided gross margins to "step down," ~$70B of FY27 capex and a ~$40B raise (FY26 FCF −$24B on $56B capex; net debt $124B) — "now a bet on execution." It rhymes with Alphabet's $85B equity raise (Jun 9): Big Tech is funding the AI build via self-funding, debt and stock all at once, and trailing FCF is collapsing under the capex (Amazon's to ~$1B on $200B planned capex). App Economy (Jun 6): the buildout is increasingly debt-financed into private labs — a $35B Apollo/Blackstone facility backstops Anthropic’s Google-TPU purchase, with Broadcom AI bookings >$30B (~3× shipments, "visibility to 2028"), OpenAI 10 GW (1.3 GW in 2027) and Meta 3 GW Broadcom deployments, and HPE cumulative AI bookings $16.4B. App Economy (May 30): on the demand side it "breaks the model" — Dell revenue +88% with a $60B AI-server guide ($165-169B FY27 vs $145B consensus) — but DRAM/NAND/memory supply is now the gate (Dell signing 5-yr supply deals, HP margins bitten, Best Buy seeing no demand pull-forward). Hay (May 28, via Deluard): the Fab Four (MSFT/META/GOOGL/AMZN) ~$700B 2026 capex sets up a depreciation spike on short ~4-5yr chip life — a widening gap between reported profit margins and the free-cash-flow margins Buffett weights more. Eisman (Jun 12): the regime shift is that the equity market now funds the build — Google's $85B all-equity raise (its first since the 2004 IPO) as AI capex jumps $80B→$180-190B beyond what cash flow covers, with Meta/Microsoft rumored to follow; "software companies have become capital-intensive hardware companies," so rotate to AI beneficiaries that don't need capital (alt energy, semis, networking). Polomny (Jun 19): the asset-light → heavy-industry drift is the tell — the hyperscalers (GOOG/META) are now building power plants and substations and buying gas turbines (low-margin businesses, not the 90%-margin software they were), and Google ran a ~$10B bond auction to fund it that Berkshire bought — "everybody's into this," a 1998–2000 redux. His play: "own the molecule toll-collector, not the hyperscaler" (the gas/uranium/copper the build must come to him for). Singh (Jun 28): the bill is now hitting the tape — the Mag 7 has shed ~$5.3T of market cap since October as the market punishes unproven AI capex (pure multiple compression, Mag 7 ~32×→~22×, with hyperscaler aggregate FCF falling from ~a third of a trillion toward ~zero, not weak earnings); only Google's raises land cleanly (a 100-yr bond, $85B equity) while Meta gets hammered on raise rumors and Oracle (worst FCF, rising CDS, ~30k layoffs) is the tell — 2027 capex consensus ~$920B, with some banks at $1.2–1.5T. Paulo Macro (Jun 29): the funding strain is now acute into quarter-end — dealer balance sheets behave like extreme year-end tightness, with the S&P AIR TRF future (AXW1, a dealer-leverage proxy) ~120 on the Sept roll while the July contract screams toward 180+, as leverage demand (levered ETFs/futures/margin) meets IPO/ADR paper supply; he expects partial relief after Jun 30 but a near-term 'reckoning,' especially around leveraged semiconductor ETFs. Hay (Jun 29): the capex-vs-returns recalibration has reached China — Alibaba posted its first operating loss since early 2021, non-GAAP net income −80%, and negative free cash flow on ~$56B of AI/cloud capex, with strong cloud growth (+38-40%) overwhelmed by the AI bill; the China mirror of the US hyperscaler problem (he moved BABA from Buy to Hold). Paulo Macro (2026-MAY-15, back-fill — "Mark to Myth Goes Public"): the "Great Circularity" climbed from the revenue line into earnings — hyperscalers mark up their private-AI stakes, book the gains as "Other Income" ($53bn for Alphabet+Amazon in 1Q, ~60% of net income), and book cloud commitments from the same companies they marked up (>$1.3tn of AI-infrastructure commitments per MS; ORCL >7× forward operating cash flow) — a mark-to-myth flywheel that reverses once OpenAI/Anthropic list (the "one true mark") or any sponsor blinks in the Prisoner's Dilemma; Micron's memory rally is the same flow, and the commitments "often underpin asset-backed private credit deals," leaking the circularity into private credit. Halftime committee (Jul 9): the premium is gone — Morgan Stanley GIC puts the Mag-7's valuation premium over the S&P 493 at ~10%, the lowest in a decade, and Brown says it's no mystery: with one exception (Apple) they now spend >100% of cash flow on CapEx (~$700B this year, $900B–$1T planned next), and the market historically hands spenders an industrial-conglomerate-style discount, not a premium — "we're in the building phase, not the harvesting phase of those profits." Lebenthal takes the other side: on a 3-month-plus horizon they're mostly buys (Meta, even NVIDIA, below the market multiple on P/E and PEG), and a year out — when the free cash flow reverses meaningfully positive — "we're going to look back on these prices and say what were we thinking not buying them here." Eisman (Jul 10): the transition is complete — the hyperscalers went "from no need for capital to massive need" (Google's $90B 2025 AI capex guided to $180B for 2026, Meta $135B, Amazon $200B+; Google's $85B June equity raise, Oracle raised, Microsoft/Meta rumored), and capital intensity without moats (Anthropic/Gemini leapfrogging weekly; corporates experimenting with cheaper Chinese AI) is a recipe for a price war, not returns; the token-subsidy reversal is biting — corporates that measured employees on AI usage blew annual budgets in months and are now curbing the all-you-can-eat buffet — so investors have migrated from the hyperscalers to semis / semi-equipment / AI-power, where Samsung's −7% on a +1,800% print shows the worry spreading to the suppliers. Link (Halftime, Jul 13): the numbers keep stepping up — ~$800B of hyperscaler capex this year and "easily $1.6T" next (Morgan Stanley same day went to $1.2–1.4T for '27–'28, raising Meta and Amazon), Meta's spend rumor moved $200B→$250B in a week, and capacity is going up "fourfold" to 2028 — she knows why they're spending but concedes returns aren't material this year (why she isn't adding to Meta and prefers Amazon's position of strength). The debt-market strain is visibly repricing the funders: Oracle hit a new 52-week low the same session, which Lebenthal pins on the "monumental" AMZN/SpaceX debt offerings going wide (7%+ yield-to-worst on the long end) — the market struggling to digest the AI-build issuance that Oracle's data-center financing rides on. Paulo Macro (Jul 21): ORCL — the hyperscaler most levered to OpenAI ($300bn+ Stargate/compute commitments, ≥$90bn FY2027 capex per Ed Zitron) — is where the AI-debt bill is coming due first: debt $100bn→$120bn since December (weighted coupon 4.4%→4.8%), market cap $500bn→$365bn, and its credit repricing to junk levels while still rated investment grade. Paulo Macro (Jul 23): the Great Circularity is now driving the earnings line, not just revenue — Google "beats" and raises capex but prints FCF -$6bn, with 87% of 2Q "income" booked as "Other" from marking up its own Anthropic/private stakes (chart via Charlie Bilello). Chanos (Risk Reversal, Jul 17): "much worse" than '99–2000 — then, the only two doomed models (CLECs + fiber) spent ~$100B over five years combined and the big spenders (telcos, enterprises like Coca-Cola/GE networking their computers) stayed profitable through the bust; now individual companies raise more than that for projects that "may be completely uneconomic from day one" — the physical structures, not the models — on off-balance-sheet and near-junk debt, committing 20-year assets against 1–2-year spot rental prices ("one of the biggest financial crimes… finance 101"). The accounting tell he harps on: construction in progress — GPUs sitting in warehouses aren't depreciated until revenue starts (he uses 10-yr chip life to be conservative and still can't pencil the data-center economics; hyperscalers/neo-clouds use 5–6), which is how S&P EPS runs 4–5× its 6% trend, exactly as in '99–2000 ($1 of one company's capex is another's profit but nobody's expense). Hyperscaler return on incremental invested capital has fallen 40%→~20% in 18 months and is heading toward 10% (Oracle the worst — the other C-suites are watching it), at which point "we could earn that on treasuries"; insider tells everywhere: Nebius' asset-light pivot ("a tremendous admission" from the most asset-heavy neo-cloud — $4.40 of capital per $1 of revenue), Coreweave exploring price hedges on GPUs whose narrative is they appreciate, Meta's 20%-owned/3-year-out Louisiana SPV, xAI's promotional 3-month-out spot deals, and legacy data centers suddenly listing portfolios as maintenance-capex reality dawns. Pencil-meets-paper moment: "some point in the next 12 months." Eisman (Jul 24): the capex regime has flipped from cheered to punished — Google raised 2026 AI capex $190B→$205B and printed negative free cash flow (−$5.9B despite revenue +24%, cloud +82%) → −7% Thursday; Tesla's $5.8B capex sent FCF to −$1.1B, its first red print in 2 years → −14.5%; the pair took Nasdaq down >2%. "A year ago… when companies raised their CapEx budgets, the market cheered… the terms of debate on AI have truly shifted." Even Intel's blowout (best revenue growth in 15 years, data center +59%) won't quell nerves — "it's the AI CapEx budgets themselves that are making investors nervous." And Blackstone now brags it is "the biggest financier of AI data centers" — "that may or may not prove to be a great bet." Morrison (Jul 23) — the dissent on memory: "I'm not in the memory bubble camp" — a two-to-three-supplier oligopoly (HBM had ONE supplier until a year ago) at high-teens/low-20s FCF yields (Samsung, SK Hynix; Micron mid-to-high teens), with no meaningful new supply for ~2 years (fabs are hard; "I need to call up ASML — yeah, get in line"). The bubble is "the fourth or fifth LLM model," already commoditising. But demand is the un-trackable half — company channel checks give "mixed reviews about AI ROI success stories" — and his acid test for the whole spend is whether it improves bank margins. "Stay tuned" for the hyperscaler capex calls. Hay (Jul 26): the financing leg is cracking — "Oracle bonds are trading almost like junk" on a "trashed balance sheet" (with the symmetry rule that its ~$200B one-day gain on the 2025 AI-spend announcement invites "the payback on the downside"); debt financing "is getting very costly," spreads "widened out drastically," and SpaceX's post-IPO bond has been "the space shuttle crashing back down to Earth." AI capex of ~$700B–800B this year toward ~$1T+ next raises "how are you going to get an ROI on that?" — hyperscaler backlogs lean on customers like OpenAI, whose >$1T IPO he thinks "gets pulled" or becomes "quite a fiasco" (already leaning toward a 2027 delay). Singh (Jul 26): Alphabet's record $112B Q2 net income was 87% mark-to-market gains on its SpaceX/Anthropic stakes (ex-markups, net income +2% q/q); free cash flow −$5.9B — its first negative quarter since going public; capex guided up to $195–205B; and contracted purchase commitments hit $811B, up $500B in three months ($201B short-term) — chips, data centers, electricity, content locked in ahead of the revenue. Consensus now has aggregate hyperscaler FCF negative in 2027 against >$1T of annual capex by 2028, funded from balance-sheet cash or new debt/equity — "but what if the cost of this debt and equity goes up? Spreads are widening." And Nvidia is in talks to guarantee ~$250B of OpenAI data-center financing (a 10-GW SoftBank build in Southern Ohio) — vendor financing as the marginal funding source. Eisman (Jul 27): on the Nvidia deal — backing the financing so OpenAI can buy Nvidia chips — "it's so incestuous… it's totally convoluted. You don't know if there's any real profitability in any of the machinations." The one concession: "there probably is for Nvidia. They're not doing it for free." Paulo Macro (Jul 30): the financing cap is now visible in primary. Data-center bond deals pulled "because the all-in yields are just not economical (can't make their IRRs work)"; a supposedly A+ credit (pseudonymized "Sopaipilla") had to price at T+285bps, repricing the existing data-center space ~50bps. "Capital is unable to keep up with the pace of the exponential spending… investors are running out of balance sheet room to continue to fund this, and that is really quite scary." NVDA's $250bn "Great Circularity" financing "has not been taken well and its CDS has joined the party" — widening on a risk-on day. App Economy (Jul 24/29/31): the hyperscaler FCF turn arrived — trailing-12-month operating cash flow across the majors grew 34% to $660B, yet free cash flow is now negative at Amazon (−$7.6B TTM on $169B trailing capex; the 2026 guide raised $200B→$220B on memory costs) and Alphabet, near zero at Meta, and −23% at Microsoft on $115.9B of FY26 capex (+80%). Debt is entering every plan — Amazon issued this year, Alphabet lists equity issuance, and even Tesla is arranging up to $30B of facilities against $43.5B of cash while its own FCF swung to −$1.1B. Amazon's counter is a two-bucket return profile: servers break even in <3 years against ≥5-year contracts, data centres monetize ~2 years out but run 30+ years across five or six server generations — with most 2027 AWS capacity already reserved, so "a timing mismatch, not a deterioration in economics." Eisman (Jul 31): a credit cycle is visible inside the trade — CoreWeave is raising $2.6B priced at "more than, get this, 9%… expensive debt," with lenders "discriminating between the large companies like Google that they know can pay the money back and smaller newer companies." And META's quarter showed where the spend hides: FCF a "mere" $784M on revenue +28% / expenses +55%, plus $279B of future AI lease commitments not yet on the balance sheet — up 53% in three months. Polomny (Aug 1): the tell arrives — Google reports its first quarter of negative cash flow, so "is this still a company that deserves a premium valuation?", and "the companies are shifting to debt to continue their buildouts… not sure this will end well." His pattern-match: "When has piling on debt to invest in a commodity-type investment worked out long term? I guess they will continue to borrow and build until the money runs out. We have seen this movie before: fiber, shale, housing." He frames the cycle as a "wealth transfer from hyperscalers to semiconductor makers" — "that which can't continue will not" — naming no chip ticker, keeping the receiving side at sector level. App Economy (Aug 1): the clearest single-quarter datapoint yet — Meta's Q2 put $31.9B of operating cash flow against $31.1B of capex and finance leases, leaving just $784M of free cash flow, −91% Y/Y, with buybacks halted and $24.9B of debt issued; the FY26 capex floor was raised a second straight quarter to $130–145B. The funding structure is the tell: the new 1 GW El Paso campus has BlackRock owning 80% while Meta keeps 20% and leases the entire site — "reduces the upfront cash burden without reducing Meta's long-term commitment." Offsetting it, AI is already paying inside the ad business (+8% ad clicks, +16% Facebook conversions, Advantage+ past a $75B run rate), and Zuckerberg floated selling paid model access and leasing excess compute at a "meaningful premium" to cost. 2026-AUG-02 (Jay Singh SSR call): the four largest hyperscalers are on course for >$725B of 2026 capex (Morgan Stanley: $779B 2026 → $1.23T 2027 → $1.4T 2028). Hyperscalers are ~20% of the S&P at −10% YTD, subtracting 2.0% from the index while the remaining 494 names returned +12% (+9.6% contribution) for a net +7.6% — "if it wasn't for the hyperscalers, the S&P 500 would actually be higher by 2%." The mechanism is structural: capital intensity cuts ROIC (worked case 26.7% → 18% when $200B of capex adds $10B of near-term earnings), Goldman models a ~700 bps mega-cap ROE decline next year, depreciation is a multi-year delayed charge, and buybacks get crowded out while $5T of IG debt funds AI projects. But the cohort no longer trades as one: MSFT held 2027 capex flat (~$117.5B) with +$19.6B FCF, Azure +43% past a $100B run-rate and >30M paid Copilot seats (+13%); AMZN was forgiven a $220B guide and −$7.6B trailing FCF on a fifth straight AWS acceleration (+37%, +15%); META raised the guide to $125-145B, printed $784M of FCF (lowest since 3Q22) and guided Q3 revenue below consensus (−10%). Cloud run-rates: AWS $169B (+37%), Azure $124B (+43%), Google Cloud $99B (+82%). Nvidia CDS doubled 40→80 bps on >$750B of circular financing tied to OpenAI ($250B guarantee, $350B vendor financing, >$500B SK Group). 2026-JUL-31 (Horizon Kinetics Q2 2026 commentary): reports that a recent HK client roundtable covered "the rapidly deteriorating business model of the five AI data-center hyperscalers, four of which are members of the Mag 7," with analyst/private-fund manager Fredrik Tjernstrom (author of HK's short-sale report The Devil's Advocate) and James Davolos — and that Tjernstrom's May fundholder letter was "essentially a sell recommendation" on those AI data-center companies. Amazon, used elsewhere in the letter as the model long-run compounder, is explicitly qualified: AI data-center capital spending is rising dramatically, it "has just about transitioned from free-cash-flow positive to negative," and it has begun large-scale borrowing — all departures from its prior capital allocation, against a current-year P/E of 36x. App Economy (Aug 4): SpaceX's first public quarter is the sharpest reading of the gauge yet — $18.4B of CapEx in a single quarter (from $10.1B in Q1 and $2.8B a year earlier), of which AI was $15.8B — 86% — against $2.6B of AI revenue and a $1.3B AI operating loss, i.e. roughly six dollars of segment CapEx per dollar of segment revenue and still steepening. Google and Anthropic are already paying cloud customers, so the revenue is real; ~$100B of post-IPO cash answers affordability, not return. 2026-AUG-10 — Paulo Macro & Le Shrub (Fly on the Wall): the funding mix has stepped down a rung — "last year the hyperscalers were funding everything with their own cash… and now this year, guess what? There's zero buybacks. They've used all their cash flows, and they're borrowing on top," now even "sprinkling some equity in there alongside, like Google… moving down the capital stack because the market's just not cooperating." Their coinage: "AIBS — Artificial Intelligence Backed Securities. Or bullshit." Consequence: equities and credit are now tethered through the same handful of levered names that also drive "half of economic growth" — "that's where the Minsky moment always goes" — into "a ton of issuance coming in in September" with yields breaking out. Dating tools (crediting Kevin Muir and a credit desk): for every "$20bn of credit" the flippers "have to start making room sooner," so watch the ladder — "a big pop and run… turning into a pop and fade… into a weak pop… into just break deal price on day one" — plus tails in the auction and the new-issue-versus-secondary tracker, which since "June, July, when momentum started to go down the stairs" has shown "existing bonds trade very poorly ahead of so much of this issuance… guys are starting to choke on the issuance." The reflexive loop: issuers hedging in Treasuries to back out the duration push yields higher, worsening the next deal. Singh (Aug 9): Morgan Stanley raised 2027 cloud capex to $1.2T (+30% y/y, $170B above pre-Q2 consensus, upside $1.4T). Hyperscalers will out-spend operating cash flow in 2027, so the spend depends on debt markets — "which is why CDS for companies like Oracle is blowing out"; >$100B of hyperscaler debt raised, excluding non-recourse off-balance-sheet vehicles. Only Microsoft still has positive free cash flow to equity. The ring: Nvidia weighing a $250B backstop of OpenAI → OpenAI = 40% of Microsoft's cloud backlog and 70% of its AI revenue → Microsoft = 22% of Nvidia's revenue → Nvidia = 17% of Micron's. "How circular is that?" 2026-AUG-11 — Contrarian Codex (Mart, Macro Update #19): "the AI buildout has become a credit product." Nvidia lined up $500bn in financing with a consortium of the biggest private-capital firms, compute itself pledged as collateral and NVDA backstopping up to 25%; Goldman is the only bank in the group, positioning as lead bookrunner on the public debt to come; Jensen Huang now calls compute "an investable infrastructure asset." That makes the whole edifice "more sensitive to credit conditions and stacks another layer of circularity onto a set of deals already drawing hard questions." Alongside it: the SEC cleared data-center securitizations to skip the post-2008 risk-retention rules "meant to keep issuers honest" — issuance $2.4bn in 2020 → $15.5bn last year, on pace for another record, "the post-crisis brake lifted for the fastest-growing corner of new credit" — and NYSE-parent ICE is paying $6bn for a corporate-bond venue, "more consolidation of the plumbing all this debt has to run through." The demand side still roars: Microsoft added almost half a trillion of market value in a single session, a record for any stock, on Azure +43%. 2026-AUG-06/10/11 — CNBC Halftime, the $500B consortium arc: Brown (Aug 6) frames the neocloud funding chain — ten neoclouds building data centers without their own cash flow, borrowing from Blackstone and peers ("if you're Jensen, do you want NVIDIA chips in there or Trainium chips?"), NVIDIA "the last of the mega-cap techs that's still massively free-cash-flow positive" at $48.5B; his consistency test on circularity: "if you believe in SK Hynix and Samsung, you have to believe in the other — it's like pulling the wrong Jenga piece out." Aug 10 (Picker/FT): Apollo, Blackstone, BlackRock's GIP, Brookfield, Goldman and KKR reportedly entering a partnership to provide NVIDIA ~$500B of AI financing; NVDA −2.6% on the headline; Amoroso's counter — hyperscaler FCF inflects back positive one-to-two years out, making them "probably the biggest undervalued trade right now." Aug 11, confirmed: MOUs with GS/APO/BX/BN/KKR/BLK to raise $500B of third-party capital so NVIDIA's customers can finance compute, the street reading revenue-sharing as alleviating circularity. Brown: "the world is structurally short compute and probably will be for at least the next five years. What if we can socialize that risk and that upside?" Sechan: NVIDIA retains only ~25% of the exposure, and "compute is now an investable infrastructure asset." Terranova: GPU pricing risk moves off NVIDIA's book, the compute is the collateral, and "in the next 5 years one of the largest futures markets in the world will be… compute" — with Wapner flagging the commoditization that implies. 2026-AUG-11 — Oakley (David Lin Report): declines the data-centre build-out outright — "not for us because we don't know where it's headed." His analogue is the 1999–2000 dark-fibre build-out: a "new paradigm" everyone had to own until "2½, 3 years later they're like, gosh, we'll never use all of this… and those stocks suffered. It wouldn't surprise me if we have a lot of that go on now." 2026-AUG-05 (Pomboy): The ~12 companies that made the S&P look cash-rich are now all free-cash-flow negative, issuing debt into widening hyperscaler spreads and rising CDS costs — credit markets show "a real increasing discernment" the equity market does not. 2026-AUG-12 (Hayes): Nvidia's $500B third-party compute-financing consortium (Apollo, BlackRock, Blackstone, Brookfield, Goldman, KKR) securitizes depreciating GPUs — "if you don't get paid back, you're basically not secured by anything useful" — and works only if rating agencies (paid handsomely) bless it investment-grade so it can be laid off onto insurers and retail; a mañana problem, decisive in 3–5 years. Meanwhile the complex has flipped from buybacks to record issuance (GOOGL's $85B, the largest US raise ever; INTC ~20% dilution), $1.5T of unfunded commitments rest on OpenAI + Anthropic, and productivity growth is the sole ROI test ("everything else… is noise") — no signs yet; he'd buy the AI trade when despondency hits, expecting the payoff in 18–24 months. 2026-AUG-13 (Polomny): The biggest bubble ever as a share of GDP — exceeding both the 1999–2000 fiber/telecom buildout and housing; the financing ladder is already two rungs down (cash flow → stock issuance → debt), the Nvidia-refresh arms race never converts to maintenance capex, and "the amount of capital that they're investing, they cannot get a return on" — so it busts into a deep recession, after which the Fed prints and hard assets rally. 2026-AUG-14 (App Economy): The neocloud cohort is the financing frontier, split two ways: CoreWeave funds the build with GPU-collateralized debt — a $49M operating loss becomes a $626M net loss on $640M of quarterly interest, Q2 capex $9.4B (3× revenue), FY26 raised to $35–39B — while Nebius funds it from the customer: >$9B of 2026 prepayments covering 50–60% of capex against Q2 capex of $5.7B (~10× revenue). The counterpoint to the strain narrative is pricing and payback: $20–25M/MW on new long contracts vs $40–50M/MW short-term, CoreWeave's ~25% July price rise, and paybacks compressed to ~22 months. 2026-AUG-14 (Eisman) — the counterpoint clause: reads the NVDA $500B consortium as constructive and conventional — "not a brand new concept… a tried-and-true way to create a financial infrastructure to support a capex hungry growth industry" (the CRE/toll-road template, likely securitized); hyperscaler FCF "could — I emphasize could — improve," and it proves "there is as yet no slowdown in AI capex." His risk sits elsewhere: Anthropic + OpenAI ≈ ~70% of hyperscaler AI revenue and 25–35% of total cloud revenue, half of Oracle's $600B backlog is OpenAI alone, both labs lose billions and live on capital raises — "a huge one-way bet… there is no hedge for the LLMs"; if they fail "the entire AI chain goes into reverse." 2026-JUN-16 (McDonald, historical): put a number on it in June — "over $500–600 billion of off-balance-sheet debt" at the hyperscalers, a private-credit meltdown already running in software ("in theory, Adobe is at risk"), CCC yields rising, data-center financing exposed to NIMBY and SpaceX's orbital share grab; "a Lehman-like credit crisis" kicking off Q3, first cracks KKR −42% vs JPM +23%, clients short the financials via XLF puts. 2026-AUG-15 — Rick Rieder: AI capex is effectively rate-insensitive ("you'd have to raise hundreds of base points" to break the hyperscalers' IRR), and its debt is now a second claim on the same savings pool as Treasury supply — "an immense amount of supply coming through that is A.I.-related," part of why real rates press higher. AUG-16 (Jay Singh, SSR): Jensen Huang answered the circularity charge in an op-ed and Singh accepts it — for now. The structure: financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilize over $500B of third-party capital, up to $125B backstopped by Nvidia, with Nvidia's exposure limited to residual-value support for at most 25% of any project — "the capital providers independently underwrite each project… NVIDIA provides the platform." Singh's translation: "he's effectively going to guarantee GPU purchases." The fundable side: hyperscaler capex accelerating +84% → +92% → ~+100% forecast in Q3 against big-three cloud revenue 23% → 35% → 43% and cloud operating margins 34% → 37% → 39%. Verdict: "I think in the short term, he is right" — with the unresolved caveat that the ROI "needs to become revenue." 2026-08-17 — the cash-flow scoreboard, stated on air: Meta $785M for the quarter; Microsoft ~$19B, "the best of anybody," but −25% y/y; Amazon negative over the trailing twelve months; Google's $85B equity raise "shocking given the fact that the last time they actually raised equity capital was in the 2000s"; "all the hyperscalers are issuing debt now. Oracle was just downgraded." Trennert: "the cash flow of the hyperscalers over the next couple of years looks like it's going to be negative… I don't see how you do it without hitting the bond and equity markets — it's not going to come out of cash flow." Eisman adds the depreciation ramp as the part that hits reported earnings for years: Meta's "went from like 4 billion to 6 billion, it'll go to 8, then 10, then 12. It's like a weight on your shoulders." (Trennert/Verrone on Eisman Ep 73, Aug 17) 2026-AUG-15 — App Economy Insights PRO: the bill is no longer a US-hyperscaler phenomenon. Tencent's capex surged 176% Y/Y and 65% Q/Q to ¥52.8B (~$7.8B), flipping free cash flow from +¥56.7B in Q1 to a ¥13.8B outflow in a single quarter — on revenue that accelerated to +11% with ads +22%. The defence is verbatim the American one (surplus AI capacity "could ultimately be rented through Tencent Cloud if its own products fail to consume it"): "Sounds familiar?" What differs is the offset: 1.4B WeChat users, a native WeChat agent in test, Hy3 token usage up 20-fold since April — but "monetization remains early." 2026-AUG-17 — Jay Singh (David Lin Report): ex-Microsoft (up 13% purely for guiding capex flat) "all of the hyperscalers are now burning cash flow" — Alphabet's $45B quarterly capex took non-GAAP quarterly cash flow to −$5.85B on +24% revenue. ~$1T to be raised for AI projects over the next year, Nvidia alone $500B to finance purchases of its own GPUs, Morgan Stanley at $1.2T of 2027 AI capex, and Goldman's count via the FT of $1.5T of hyperscaler lease commitments, $1T not yet started. Backlogs are circular (OpenAI/Anthropic funded in the same capital markets): "that circular financing works in the short term… the numbers will continue to beat for the next couple quarters. At some point next year, the market will say — how long is this sustainable?" 2026-AUG-27: Hay: reverse crowding out — the “deluge of hyperscaler debt raises” funding “$700 billion plus annual expenditures” on data centers is “driving up the government's borrowing costs,” a third force on the long yield alongside the deficit and the 8.4% issuance surge. (David Hay, Haymaker Daily 2026-AUG-27) 2026-AUG-28: App Economy on NVDA's Q2 FY27: supply and capacity commitments jumped $119B → $279B in three months (largely memory), NVIDIA has invested ~$50B in frontier AI labs and helped arrange >$500B of third-party capital, and labs on some form of its balance-sheet support “could account for roughly one-quarter of its business next year” — “financing customers does not make the underlying demand fake, but it increases NVIDIA's exposure if AI labs eventually struggle to monetize.” (App Economy Insights 2026-AUG-28) 2026-AUG-28: IG issuance near $1.5tn YTD, up ~36% y/y and on pace to eclipse the 2020 record; the biggest technology borrowers alone are ~$200bn of it — about 25% of Treasury's net note-and-bond issuance to private investors and ~5x last year — with a single multi-tranche deal near $53bn and "somebody sold a century bond. 100 years of duration, into this." Hyperscaler capex near $800bn this year and above $1tn annually from 2027, AI-related issuance estimated $300–570bn, and one estimate puts ~30bp of this year's 10-year move on corporate and mortgage supply combined. "So the term premium thesis picks up a second engine." (Contrarian Codex 2026-AUG-28) 2026-AUG-23 — Jay Singh (SSR call, 2026-AUG-23): the "big tech bond paradox." Hyperscaler capex "initially was free cash flow, but that dried up. So now it's by issuing bonds" — Google's 100-year bond, Meta's and Oracle's 6-8% long paper at 200-300bp over Treasuries — so institutions sell sovereign duration to buy it, pushing Treasury yields higher. "Big Tech needs to issue this debt in order to continue to spend on CAPEX, and that same CAPEX growth is what's supposed to solve our debt issues." Credit is repricing it: Broadcom seeking $60B+ (potentially ~$100B) with a spike in its CDS, Nebius pressuring its own stock with a $4.5B convert via arb hedging, and Goldman arranging a $1.15B junk bond for CoreWeave-linked Virginia data centres. Jeff Currie (The Trevor Rose Podcast, 2026-AUG-20) re-runs the commodity template on the AI names: producers spent 120% of cash flow in 2013–14 and "when you're at 120% and the thing dies, it's game over." The behavioural cycle then repeats in fixed order — bust, "we swear we will never do it again," capital discipline, prices scream, a new management team arrives and starts spending (of 13 companies at the 2014 LME dinner, "only one management team was still the same from 2012"). His conclusion for tech: "it's cyclical… they're going to have to learn how to get much lower multiples, more like what the energy guys did." And the repricing casualty is whoever owns the fixed physical asset: "those companies like Chesapeake who own the land got killed… probably the same thing with the AI guys — the guys who own the land are going to get in trouble because they can't handle the repricing." Plus the flow-of-funds line: "all the money is sitting in AI right now. What happens when that money leaves AI? It goes to oil." Kedrosky (Meb Faber #648, Aug 28) dates the crossover: in H1 2026 more than half of data-center financing stopped coming from internal cash flow and became external — ABS, private credit, sovereigns, SPVs. He calls this the Minsky tell, the point where the money divorces from the application: lenders underwrite the 12-year renewable lease and the prime credit behind it, not the asset ("there could be hide-and-go-seek competitions going on inside the data centers and they wouldn't give a…"), so construction is no longer disciplined by utilization. A JP Morgan figure puts data-center-related paper at 15–18% of the investment-grade market — larger, as a share, than financial services. Plus Michael Burry's maintenance-capex point: perpetual capex obligations "into eternity." Net: orthodox tech companies are becoming utilities with similar obligations but "wildly overvalued compared to an orthodox utility" — "who gets re-rated first? … I obviously think the latter." Note the rhetorical tell he flags: the same people who argued it was safe because it was cash-flow-funded now argue it is safe because sophisticated outsiders are funding it. 2026-AUG-30 (Jay Singh, SSR call): credit moved first. "Hyperscaler CDS spreads hit all-time wides last week… hyperscalers have 3.1 trillion of off-balance-sheet arrangements, not all of which are actually guaranteed" across Meta, Google, Microsoft, Amazon, Oracle, Nvidia and Broadcom — and it does not mean-revert: "these CDS spreads could calm down if they stop issuing more and more debt, but I don't think that that is the goal." Goldman's Delta One desk: credit "is starting to ask questions that equity largely ignored," with NVDA/Broadcom CDS widening on AI financing backstops, inventory builds in the semi chain and power bottlenecks plus policy pushback on data centres — "none of these kill the AI story, but the range of outcomes is clearly getting wider = lower multiples (SOX 24-month forward P/E from 21-22 to 15ish)." The contingency is now stated inside Nvidia's own guide: part of the 70%-through-FY28 forecast "is contingent on financing for Anthropic doing an IPO in October and OpenAI securing financing… a lot of Nvidia's growth is actually being financed by itself, through these circular capital arrangements" — Nvidia's CFO puts balance-sheet-supported AI-lab demand at ~25% of next year's business. His qualifier: "it's not the base case, but it's still a big uncertainty." Funding is also reaching further out — SoftBank's $6.3B retail bond at a 4.3-4.9% coupon against $60B+ of OpenAI commitments, after selling its entire Nvidia stake. Steve Eisman (The David Lin Report, 2026-AUG-28): cheap capital is not sufficient — asked whether low rates and available funding keep the train running, "No." The sequential arithmetic on OpenAI is the reason: June-quarter revenue ~$6.5B, up only 18% q/q, while costs went to $12.5B, "up 3 billion in 3 months… revenue in 3 months was up a billion dollars sequentially and its costs were up 3 billion sequentially. That's not the right direction." With token maxing over, "I think it's going to be harder for these guys to raise capital." He also flags an announced financing that does not exist: the $500B Nvidia third-party capital facility with KKR, Apollo, Blackstone and BlackRock is "not exactly" real — per Bloomberg the transaction was dragging, "so Jensen went out and basically just announced it anyway… right now there is no $500 billion." Anthropic is the healthier credit (~$11.5B in the June quarter, +100% q/q) but the H2 print is the test. (New Money clip, 2026-AUG-15) adds the transmission mechanism into listed names: "so much of the hyperscaler backlogs are from these two companies. For example, of Oracle's 600-plus billion backlog, around half is from OpenAI." 2026-AUG-28 — Paulo Macro: "IG credit spreads for what were previously considered 'asset light' businesses are pressing higher in a reflexive loop with the US Treasury curve with an expected onslaught of future issuance just as managers are fully committed to equities and carrying low cash." The reclassification is the tell — the hyperscalers are now heavy, debt-funded borrowers being repriced as such; issuance needs duration hedging, which lifts yields, which widens spreads, which raises the cost of the next deal. 2026-AUG-20 — AI is bidding for capital against the Treasury, and the Gulf money is Bessent's. "You've got essentially AI bidding for capital. Secretary Bessent bidding for capital. The two of them are competing with each other" — with hyperscaler issuance heading toward $500bn this year. The history: the Bloomberg series of five prior capex booms as a % of GDP back to the 1840 canal boom ("several US states defaulted, went bankrupt"), and "every single one of those prior five booms ended in a bust. And this one will end in a bust too" — at which point "Bessent's $4 billion in Treasury buybacks will probably turn to 40 or 400 billion, and the Fed will probably be helping." The rotation rule from the same work: two to three years into any prior boom you did better selling most of it and buying gold, which won over the full cycle every time (sell the Dow for gold in May 1929 and you're still +15% 103 years later). And the circular reference he says nobody has named: the UAE demanded swap lines after the missiles flew ("we're illiquid, and if you don't give them to us we're going to go to the Chinese") — so "if Bessent's given swap lines to the UAE to invest in us, who's actually investing in it? He's just creating the money." Related: insurers have swapped long Treasury duration for SOFR-linked private credit and did not return when the 10y hit 4.7 — "if they could sell it at a decent mark… they would have, and they didn't." Gromen (Monetary Matters, 2026-AUG-20) Wigglesworth (Monetary Matters, 2026-AUG-16) put a number on the footnote. Hyperscaler off-balance-sheet lease commitments went from ~$1trn to ~$1.5trn in a single quarter — of which roughly $500bn are leases already commenced (a right-of-use asset with the liability opposite: visible as payment obligations, never labelled debt) and ~$1trn are leases that have not started and exist only as a footnote. Goldman Sachs did the filing sweep on leases; Wigglesworth built the parallel purchase-commitment tally out of the 10-Qs himself and got the same shape — ~$1trn → ~$1.5trn of contracted chips, memory, cooling and power, with GOOGL alone at $800bn (almost half the total) and ~$200bn of it flagged short-term, which implies published capex expectations are too low. These obligations "walk, talk and quack a bit like debt but they don't actually appear as debt." The template is META's Hyperion JV with Blue Owl (OWL): 20% equity, a 20-year lease guarantee sized to cover the vehicle's costs, bonds sold to third parties, nothing on Meta's balance sheet — and the guarantee is "incredibly strong. I don't see how they can squirrel out of them." His diagnostic: "this is a debt cycle," not an equity cycle — equity-financed booms break benignly (the dot-com bust halved the market and was economically "a nothing burger"), while debt-fuelled capex cycles "quite often end in tears even when the underlying premise comes true," as the canals and railways did. Scaled to today's economy, 19th-century railways issued the equivalent of ~$10trn of bonds and still bankrupted Jay Cooke & Co. The kicker on optics: for an investment-grade issuer a plain-vanilla corporate bond is the cheapest route, so "flexibility is probably a convenient excuse" to obscure companies going "from being lean mean cash machines into being capex hungry utilities." On timing he claims none — the industry has become "remarkably incestuous with an incredible tangle of financing agreements, co-investments, supplier and customer relationships" that both worries him and "can keep the show going for a lot longer." (Wigglesworth, Monetary Matters, 2026-AUG-16.) 2026-AUG-31 — Jeff Keller (Other People's Money): credit-funded capex is "definitely a concern" and "a more dangerous way to finance this buildout" — buybacks turning into debt issuance at hundreds of billions heading toward trillions, crowding the Treasury market. But not decisive: after the Anthropic IPO the labs + SpaceX + Google will have raised ~$500B year-to-date plus heavy debt and "markets are broadly okay"; top lines still grow double digits and there's cost base left to cut (as in 2022). Unless yields really blow out, there's enough money for a couple of years. Key framing: hyperscaler capex is rational optionality-buying, not an ROIC calculation — defensive (not playing could wipe out the business) plus offensive, structured to stay optional (Microsoft buying powered shells before chips). A low-but-positive ROI is still rational, so a 10–20% drawdown won't force a capex brake — "there's been maybe a disconnect between the financial analysis of this and the religious analysis that maybe comes from the West Coast." 2026-SEP-01: Trader Ferg's article-of-the-week is Myrmikan Research, "AI Debt Failure Will Prompt Another Wave of Fed Bailouts" — the debt-financed-build-out-to-central-bank-rescue arc — with his only comment the method: "nothing like a good dose of history to help understand the present." (Ferg's Finds, 2026-SEP-01.) 2026-AUG-26 (Dowd/Phinance, WTFinance): the financing engine is visibly cracking — OpenAI's Altman/Friar floated government financing ("salvo number one") because "the engines of financing were starting to crack"; private credit, "a big source of funds for the AI buildout, started to freeze"; spreads widen with every new AI bond deal; Apollo's CEO says most AI-complex revenue "is coming from investors, not end customers." Headline financings are MOUs, not contracts — Nvidia's $500B from Blackstone and Stargate's $500B, "still yet to get funded," so they "can just vaporize overnight." Nvidia chips are also being pledged as collateral ("insanity to me"). (2026-SEP-03, Dan Niles, Excess Returns) The change is the funding, not the demand. Hyperscaler free cash flow has gone negative — "like for Google… for the first time since they went public" — whereas the dot-com build "we were funding through free cash flow. Now we're starting to use debt… some of the circular stuff." His daily screen is hyperscaler credit default swaps against the North American investment-grade index: levels are "at the highest they've ever been," and "even Nvidia… their credit default swaps were actually higher than the average North American investment grade credit default swap for 5 years." The rationale is structural — "credit investors are way smarter than equity investors" because their upside is capped at repayment, so they do more work and move first. End-game: one hyperscaler misses its commitments, and "they'll all probably go ahead and slam the brakes on at the same time." Dillian (2026-SEP-03): says the capital structure is the whole bubble — “in my lifetime, this is the first time I’ve seen tech being financed with debt. Usually you finance tech with equity … because the asset has a very short lifespan. It’s going to be obsolete in a couple years. You don’t want to take out 10 or 30 years worth of debt to finance an asset that’s going to be around for two or three years.” The 1999 contrast is explicit: “the dot-com bubble was all equity. Nobody was issuing debt and this time we have a lot of debt. And the leverage is what gets people into trouble.” The cost is now real even at a tight spread — Google’s paper is maybe 60–80bp over Treasuries but “they’re essentially paying a 6% coupon,” against 2–2.5% in 2021 — and neither he nor Farley expects it to slow the build: “are the CEOs going to say, oh my god, the 30-year Treasury just hit 6%, stop capex? It’s not going to happen.” 2026-SEP-07 (Excess Returns — Cameron Dawson clip): a second channel beyond the debt: "there's a bunch of stuff that's not even hitting the balance sheet or cash flow statements because it's all happening in these SPVs — super esoteric and very murky; we don't know what's going on in them, we don't know all the different players, and we know that they're private." Combined with the supplier-financing loop (Nvidia's receivables against hyperscaler payables), she reads the "clarion calls saying no, it's not circular financing" as itself the tell: the denials "raise a lot more eyebrows." 2026-SEP-08 (CNBC Halftime): the yield→issuance channel gets a single-name gauge. Lebenthal on Oracle: "I would like to see them not raise their CapEx expenditures… I think it's at 70 billion… the five year credit default swaps are over 200 basis points, don't want to see that going higher," with 33% y/y top-line growth the pass mark before margins matter and the Abilene, TX site the ramp to watch. 2026-SEP-08 (Larry McDonald, Julia La Roche): the off-balance-sheet leg, and the lenders' own verdict on it. Meta has gone from $30bn to "up to 600 billion" of off-balance-sheet financing in two to three years, and the hyperscalers are adding "an extra 5, 600, 700 billion" of debt over 12–24 months. Banks lend against double-A-rated future free cash flow and then buy credit default swaps on their own Mag-7 borrowers — which "rhymes a lot with 2006, 2007, 2008." The pivotal variable is the cash-burn-to-free-cash-flow turnaround date: "If that's in 2030, 29, okay. If that's in 2031, 32, 33, Houston, we have a problem." 2026-SEP-09 (Steve Eisman, Ed Zitron interview — free preview only, remainder paywalled): the circularity gets put on the revenue line rather than the balance sheet. The hyperscalers — Google, Microsoft, Amazon and Meta — have spent "over a trillion dollars in capital expenditure buying GPUs from NVIDIA"; OpenAI and Anthropic "are estimated to make up around 70% of hyperscaler AI revenue," and UBS projects the two "could account for 48% of all Google Cloud revenue next year." So "the hyperscalers are investing heavily in GPUs largely to capture revenue from two companies they themselves continue to fund" — "something of a closed loop," with both counterparties "deeply unprofitable, having lost tens of billions of dollars between them." The disclosed test case is Microsoft: fiscal-2026 AI revenue of $34.3bn of which $24.1bn came from OpenAI alone, leaving its arm's-length AI compute/software business "a single-digit-billion-dollar business." Eisman does not counter the case — he introduces Zitron as someone he "subscribe[s] to and read[s] regularly." App Economy Insights (2026-SEP-11): Oracle's Q1 FY27 CapEx hit $28.5B (~$5B negative quarterly FCF, -$29B TTM, from +$6B in Feb 2025), but customer prepayments covered $11B, netting the cash outlay to ~$18B; the $20B equity offering is complete and the latest $30B+ of AI contracts 'require no incremental increase to its planned capital raise.' ~$70B of net CapEx still to fund this year. Customers are becoming part of the funding stack. Lebowitz (Sep 10): hyperscalers 'going beyond their free cash flow' into debt and equity markets is now a crowding-out narrative pushing Treasury yields up. AI-driven memory-chip price spikes (Apple raising prices) feed the inflation story, though he expects AI to be disinflationary later. 2026-SEP-02 Pomboy: 'eerie echoes' of summer 2007, when asset-backed commercial paper was the first flash point of the housing bust — rising hyperscaler borrowing costs and CDS show credit demanding a risk premium while equities stay 'zippitydah'; even Mag 7 names are turning free-cash-flow negative, and the top 10 balance sheets look nothing like the other 490. CNBC Halftime (2026-sep-11): Oracle's print grades Lebenthal's own pre-set checklist — 30% top-line growth, free cash flow -$5B vs -$10B expected, NO new capex plans and no new fundraising, and the 5-year CDS in to 181bp from 215bp a few weeks earlier; RPO +$30B to a backlog approaching $700B. The stock fell anyway, which he attributes to rates, not the release: debt going from ~$80B toward $150B over two years. The sharper framing is Harrington's: "I don't think you need AI to come crashing down to have a problem... you could have AI contract at the margin," and her screen is balance-sheet experience rather than valuation — "I think of like the CoreWeaves more than I think of the Oracles... energy companies and real estate companies, they've lived through huge debt loads, super long dated, laddered out forever. They know how to deal with it." Steve Eisman (2026-SEP-11): grants a subscriber's long-run case (Nvidia at ~74-80% gross margin "extracting rent from hyperscalers" to finance their challengers; "sell the knives" to ASML/KLA/Lam/Applied/Micron/SK Hynix) but not the path - the current ecosystem is "almost completely dependent" on Anthropic and OpenAI, one could fail before replacements arrive (dot-com's failed first generation), and "if OpenAI fails within a year, there will be a massive correction." Oracle's +30% quarter came with RPO only $664B vs $600B, ~50% still owed by OpenAI. Anna Wong (2026-SEP-11): the second derivative of AI capex peaked this year. AI capex added ~1 pp to GDP growth in H1 2026 and could add ~0.5 pp or less next year even as spending rises: 'peak AI capex boom contribution to GDP.' 2026-SEP-10 Gundlach: AI-sector corporate spreads blew out (IG ~50 to ~125bp, HY ~180 to ~325bp) while ex-AI IG is unchanged and ex-AI HY sits near its tights - an issuance avalanche on top of Treasury borrowing that the market is struggling to digest; buyers may be PE-owned captive insurers. Luke Gromen & Darius Dale (2026-SEP-13): Dale is the AI bull — capital deepening at a record 22.3% implies a productivity boom and faster NIPA profits, and a stock/gold/Bitcoin bubble into end-2027/mid-2028, ended by redistributive taxes after the 2028 election ('they're gonna go Nvidia'). Gromen prefers electrical infrastructure over AI equities: AI borrows for hard assets meant to eliminate labor, undermining a tax base half made of employment. David Woo (2026-SEP-14): an endogenous loop — if stocks don't fall, Microsoft and Amazon keep issuing debt to fund AI capex and long rates rise "until the stock market pops"; higher long rates crowd out mortgages, housing "and possibly even AI." Nvidia's $125B into OpenAI's $500B data-center project and reported $10B Anthropic IPO anchor show the industry funding its own customers: "they have to keep this party going." 2026-SEP-14 — David Rosenberg: the leveraged AI boom 'needs low financing costs' and energy; push-back and executives signalling slower capex would shrink the debt binge and corporate issuance, easing long yields; ~45–50% of global equity lies outside the S&P 500 'where you don't have to take on all that AI concentration risk'. 2026-SEP-14 (Carlson): An inverse read of the capex debate: a coordinated AI pause would help most the heaviest spender with the least direct AI revenue. That is Meta, since AWS, Google Cloud and Azure rent their capacity out. Its capex would fall while ad monetization keeps compounding, lifting margins and FCF and rerating a 21 PE. He also argues the frontier labs' own commitments (Anthropic ~$500B) are a reason they want a pause. Dan Niles (2026-SEP-15): megacap tech CDS "have all been going up for some period of time," and some now cost more to insure than the North American investment-grade index. He "never thought I would see Google having to issue equity or debt." Neoclouds are debt-funded with no cash flow amid "off-balance sheet commitments of over $3 trillion" — "cash is king. Everything else is an opinion." 2026-SEP-14 (CNBC, Terranova): slower AI spend plus a 5% 10-year is "a disincentive to see more supply in the debt market" — his JOET AI-power names (GEV, VRT, PWR, ETN) are "in a much more perilous position"; Santoli: they live off forward order books "for which nobody has the money to pay for"; Weiss expects "a thinning of the herd" in private AI as unprofitable labs must raise more. Mike Taylor (Hedgeye, 2026-SEP-15): a top-down model says the math doesn't close — ~$1.5T of US AI capex in the ground plus ~$1.5T committed = ~$3T by end-2027; at 7-year depreciation that's ~$400B/yr, ~$1T/yr all-in with power, people, cooling and a 12% cost of capital, against OpenAI + Anthropic (~70% of demand) at ~$100B+ revenue, which at an 80% margin leaves ~$40B to pay for compute. Credit flagged it first: Oracle's bonds said no to the OpenAI backlog from March-April, and CoreWeave unsecureds trade ~12% ('meaningfully distressed') while street models triple its debt; OpenAI needs ~$600B over two years and 'the private money is out.' 2026-SEP-15 (App Economy Insights, SB Energy S-1): the circular loop from the landlord's side — OpenAI is anchor tenant and warrant holder (warrant revaluation alone = $2.6B of H1 loss); NVIDIA is compute supplier, ~$3B IPO-linked investor (only half fresh capital) and credit support on OpenAI's leases; SoftBank is sponsor, customer and >50% vote controller. The $439B backlog is 81% beyond year 8 and needs ~$178B of CapEx, so equity value rests on an 8–11% unlevered yield on cost vs 6–7% debt: 'the backlog may already lock in enormous future cash generation, but it does not lock in enormous shareholder value.' Eisman Ep 75 (2026-SEP-14): Nvidia 10-Q note 7 — top five direct customers = 70% of receivables; OpenAI June quarter $6.5B revenue (+18%) vs $12B costs ("revenue went up a billion and their cost went up three"); hyperscalers flipped from buybacks to debt and equity. Collins: "a different paradigm… it scares me." John Polomny (AIA weekly, 2026-SEP-05): hyperscaler and Nvidia debt issuance as a share of Treasury issuance — they've "consumed all their cash flow… going into debt"; "I've seen this movie before in the telecom bubble… this is far and beyond what that was." Joseph Carlson (Sep-15, 2026) dissents on Meta: a reported $180–200B 2027 capex that may turn FCF negative is rational because data-center land and permits near cities only get harder amid public backlash — "he's not operating the company for next year." Gundlach (2026-SEP-16): decomposed credit — non-AI junk and loans near their tights while AI junk is +50bp and AI loans +130bp; SpaceX bonds trade ~three notches below a BBB- he suspects was "persuaded"; Oracle wide after issue; hyperscaler deals +75 bid +200 within days; not yet in single-B. Gundlach (2026-SEP-16, post-FOMC): duration- and rating-adjusted, ex-AI IG went 70 → 78bp while AI IG widened 50bp; ex-AI high yield flat vs AI high yield +150bp, often days after issue — investors "do not believe the ratings" on AI bonds. Chris D'Agnes (Hamlin Capital, 2026-AUG-10): Nvidia and others are tapping the capital markets and Meta is raising a lot more debt - the free cash flow that once made hyperscaler dividends thinkable is now 'all getting used in capex'. Halftime (2026-SEP-17), Ethridge: the bond market is "in disarray" trying to absorb "$75 billion worth of issuances of hyperscaler debt just since July that it can't find takers for"; Oracle is "using their own equity as the thing that the debt is being offered against" — which only works while the share price rises. A first-half issuance rush ahead of the midterms could mean lower 2027 capex expectations, "the death knell for this particular trade"; Simpson: higher rates alone cool the appetite to borrow. 2026-SEP-19 (Jeff Weniger): the Silicon Valley giants have flipped from cash-rich (Google once had ~$100B earning ~5% in T-bills) to issuing debt and equity. That risks crowding out Treasuries and raising the discount rate on out-year (2030-35) cash flows, e.g. Oracle's 2032-33 earnings. The Mag 7 "got cold" once the raises became conversational. He calls 2027's theme "buybackers vs diluters": Ken French data since 1963 show expansion-era equity issuers fall 80-90% in bear markets. George Noble on The Real Eisman Playbook Ep 76, 2026-SEP-21: "it's all being funded with funny money." Nvidia having to announce a $250B package for OpenAI "tells me all I need to know"; the AI boom is "dot-com and subprime all merged in one." Eisman: the circular financing "tells me that there's probably something wrong"; Nvidia's 70% of receivables from five accounts is "very frightening"; and if lenders thought these were good credits "they'd make loans, but they're not doing it." Harley Bassman on MacroVoices #550, 2026-SEP-17: hyperscaler borrowing is price-insensitive ('6, 8, 10, 12, it doesn't matter') and the Fed can't stop it. But Meta, Google, Amazon, Microsoft and Oracle bonds are 'money good' even at +50-200bp because their core businesses cover the coupons. The risk is to the equity (Global Crossing redux: one or two winners), and to labs like Anthropic, which 'can go bust' as an equity, not a bond, problem. Contrarian Codex (2026-SEP-22): Amazon, Alphabet, Meta and Oracle issued ~$194bn of bonds in 2026 through early July (+79% on all of 2025; Goldman sees ~$250bn this year and $400bn in 2027), competing with the Treasury for the same capital. ~70% of federal receipts come from paycheck withholding, so "anyone pitching AI as the productivity miracle that saves the budget should explain how the budget survives the payroll cuts that miracle requires."