AI Capex Unwind — Digest
TL;DR
The Broadcom SPV thread is resolved — and the terms are worse than the rumor: $60B settled roughly 40% below August's headline, with Broadcom now simultaneously supplier, lessor, lender, and guarantor for ~$102B of exposure against $89.1B of FY2026 revenue. Meanwhile the credit layer broke before the chip layer did: September AI debt issuance halved, Oracle was cut to BBB-, and the first project bonds are visibly derating on permitting and power friction — while Nvidia sits 4.2% from its high. The market has stopped pricing the buildout as a basket; it's pricing the chain.
The Broadcom–Anthropic loop closed — and it's bigger than anyone's balance sheet can quietly absorb
The vendor-financing question moved from rumor to filing this cycle, and the answer is that Broadcom is now four counterparties in one. Anthropic's own IPO prospectus confirms the structure: supply, lease, up to $42B in convertible notes, and a guarantee on the $60B being syndicated — a concentration Anthropic's filing itself flags as a "potential conflict of interest" (Broadcom AI Financing Platform
).
"Broadcom sells Anthropic the chips. Broadcom guarantees the debt that paid for them. Broadcom lends Anthropic money to pay the lease. Broadcom may end up owning Anthropic equity."
— TechTimes, via Broadcom AI Financing Platform
The credit market got there first: Broadcom's five-year CDS hit a record 122bps in August, and the stock is 24.7% below its 52-week high despite a pristine balance sheet at 1.1x leverage. Rothschild's Robert Leitao put the systemic concern plainly: "It feels that there's quite a concentrated bet right now on two companies being able to generate enough revenues to support all the financing that's happened" (TechTimes). The August-to-October shrinkage from ~$100B to $60B matters too — the same haircut could recur before books close.
What to watch: Anthropic's IPO — the junior tranche and the convertible notes are both gated on it, so a delay or reprice forces Broadcom to hold the exposure itself.
The debt layer broke first — issuance halved and real bonds are trading at real losses
The credit side of the map has turned: AI-related debt issuance dropped to $23B in September from a record $113B in June, after $466B raised year-to-date — AI borrowers now account for roughly half of new US investment-grade issuance (The Debt Layer
). And for the first time, specific instruments are visibly derating on physical friction, not AI-demand doubts:
"Oracle 'Project Jupiter' (New Mexico): ~$18B of loans privately quoted ~85 cents on the dollar after Oracle issued a force majeure notice when the site struggled to gain electricity access."
— The Debt Layer
Tract Capital's Nvidia-leased Reno paper fell ~11 points to 90.53 on an NV Energy lawsuit; Zenith Arc's Oklahoma bonds now yield over 11% after a flood-plain permit stumble (The Debt Layer). This is the mechanism the map exists to find: a demand slowdown becomes a construction slowdown when the marginal project gets financed at 11%+ yields or not at all. Buyer behavior is shifting too — PGIM anchored a CLO capping AI collateral at 15% of the pool, the first with an explicit AI concentration cap.
What to watch: Whether September's issuance halt persists into Q4 — the SpaceX $40B raise and the Broadcom syndication will re-price the whole complex.
Oracle is the map's key node — the first hyperscaler whose capex became someone else's credit event
Oracle is the purest transmission point on the exposure map: quarterly capex of $28.5B against $23.1B of operating cash flow produced negative $5.4B of free cash flow in a single quarter, total debt hit $155.93B at 4.5x EBITDA, and S&P cut it to BBB- with the five-year CDS at 203bps — an 18-year high (Oracle Debt and Capex
+1). The stock is down 54.2% from its high even while revenue grew 29.6% and EPS beat every quarter.
The dependency underneath is the point: a $664B backlog heavily weighted to OpenAI, whose annualized revenue came in ~$20B below investor estimates — and whose data-center completions have already slipped from 2027 to 2028. Oracle's answer is to replicate the Broadcom playbook, negotiating with Apollo and Goldman Sachs to fund a separate chip-buying company it would lease from — keeping the purchases off a BBB- balance sheet (Oracle Debt and Capex+1). When the weakest IG name starts structuring around its own rating, the SPV is no longer an innovation; it's a workaround.
What to watch: Confirmation and rating-agency treatment of the Apollo/Goldman SPV — whether the debt gets consolidated back into Oracle's credit is the whole ballgame.
The chip layer's equity and the credit beneath it are priced in different worlds
The divergence is now quantifiable in a single snapshot: NVDA sits 4.2% below its 52-week high with 0.2x leverage and $21.4B of quarterly free cash flow, while Oracle trades 54.2% down, CoreWeave 42.6% down at 13.6x leverage, and project bonds at 85–95 cents (Nvidia Circular Loop+1). Nvidia formalized its role as originator with a $500B financing platform alongside Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR — with reinsurance broker Howden Re working on structures to cover neocloud borrowers against collateral shortfalls (Nvidia Circular Loop+1).
The silicon layer's own demand shows zero slowdown — TSMC's September sales rose 54.6% YoY and Samsung posted a ninefold profit jump. That's why the map's domino scenario runs up the chain: a demand event at the AI-lab layer propagating into the collateral assumptions of the $500B platform. The Oct 8 OpenAI revenue scare was a small live rehearsal — roughly $500B of equity value erased in a day on a methodology dispute (Nvidia Circular Loop+1).
What to watch: TSMC (Oct 15) and ASML (Oct 14) earnings — the nearest checkpoint on whether silicon demand is decoupling from credit stress.
The equipment layer keeps proving the shorts wrong on fundamentals while derating anyway
Vertiv is down 35.4% from its high with no earnings miss to justify it — Q2 beat on EPS, leverage at 1.2x, and the Q3 setup guides to ~35% organic growth (Vertiv Equipment Layer
+2). Upstream, the order books are accelerating, not stalling: Eaton's Electrical Americas orders run +41% on a 12-month rolling average, and GE Vernova's data-center orders in the first half already more than doubled its entire 2025 total (Vertiv Equipment Layer+2).
The sharpest bear articulation targets backlog quality rather than demand: "I see risk that the massive $15 billion backlog Vertiv holds has actually been artificially inflated by supply chain panic... customers might have decided to double-order or place massive advanced orders simply to secure their place in line" (Seeking Alpha). The Oct 28 print — revenue against the $3.75B midpoint — settles whether "timing" becomes "pattern."
What to watch: Vertiv's Oct 28 revenue and deferred-revenue trajectory; a second miss converts the double-ordering hypothesis into the base case.
What surprised us
- The $60B SPV settled 40% below the August headline. The prior cycle's open question ("$70–100B, unsigned") resolved at $60B plus a separate $42B convertible facility — and the pattern of headline shrinkage before close is itself a data point about where syndication appetite sits (Broadcom AI Financing Platform).
- The first force majeure was about electricity, not AI demand. Oracle's Project Jupiter loans broke on grid access — the power layer's supply constraint showing up as a credit event in someone else's bonds. The unwind, if it comes, may arrive through permitting and power friction before it arrives through revenue.
- TPU collateral is structurally worse than GPU collateral. The $42B senior tranche is secured by custom ASICs with essentially one buyer class — frontier AI developers — so recovery value depends entirely on frontier-AI market health at the moment of unwind (Broadcom AI Financing Platform).
- A private-company accounting dispute can move half a trillion dollars. The OpenAI $50B-vs-$70B gap was a gross-vs-net methodology difference, and it still erased ~$500B of equity value in a day — the revenue base under the entire debt layer is private, unaudited, and methodology-dependent until the IPOs land (AI-Lab Revenue and IPO Gate).
Open threads worth a vote
AI Capex Unwind — Digest
Since last time
The headline change: the Broadcom SPV stopped being a rumor and became a filing — at $60B, roughly 40% below the August headline — and the credit layer broke before the chip layer did. The map's center of gravity moved from equity derating to actual bond losses.
- Promoted — The debt layer. Previously a supporting reference inside the vendor-financing section; now a standalone section with hard data: September AI issuance halved to $23B, and for the first time specific bonds are trading at real losses.
- Escalated — The vendor-financing chain / single-chokepoint thesis. Last time it was an unsigned $70–100B SPV and an open thread; now it's a closed $60B syndication with Broadcom as four counterparties in one, confirmed in Anthropic's own IPO prospectus. Your "single chokepoint" surprise bullet is now the core systemic story.
- Escalated — Vertiv. Same thesis (fundamentals fine, stock derating anyway, now -35.4% vs -32.6%), but with a new, sharper bear articulation targeting backlog quality, plus new upstream order-book data.
- New — Oracle and the chip-layer equity/credit divergence. Neither appeared last cycle; Oracle is now the map's key node and Nvidia's $500B financing platform formalizes the originator role.
- Demoted — Vertiv's deferred-revenue signal. Was a full section last time (deferred revenue doubling to $3,633.7M); now only the deferred-revenue watch line survives in the equipment section.
- Disappeared — Vistra and the power-layer repricing section (Vistra is absent entirely; power now appears only as supply friction in someone else's credit events). Meta's FCF collapse section (no Meta anywhere, despite it being last cycle's sharpest signal). The Zuckerberg-concession debunk and the Vertiv customer-concentration resolution (both surprise bullets from last time are gone — the latter was resolved, so no loss).
- Unchanged — The map's core thesis: revenue recognized by the chip layer today, financed by promises at the AI-lab layer tomorrow. Every section is an elaboration of it.
The Broadcom–Anthropic loop closed — and it's bigger than anyone's balance sheet can quietly absorb (Escalated — was your open thread; now resolved, with worse terms)
The $70–100B question settled at $60B — and Anthropic's IPO prospectus confirms Broadcom is now supplier, lessor, lender, and guarantor simultaneously: supply, lease, up to $42B in convertible notes, and a guarantee on the $60B being syndicated — a concentration Anthropic's own filing flags as a "potential conflict of interest" (Broadcom AI Financing Platform).
"Broadcom sells Anthropic the chips. Broadcom guarantees the debt that paid for them. Broadcom lends Anthropic money to pay the lease. Broadcom may end up owning Anthropic equity."
— TechTimes, via Broadcom AI Financing Platform
The credit market got there first: Broadcom's five-year CDS hit a record 122bps in August, and the stock is 24.7% below its 52-week high despite 1.1x leverage. Rothschild's Robert Leitao put the systemic concern plainly: "It feels that there's quite a concentrated bet right now on two companies being able to generate enough revenues to support all the financing that's happened" (TechTimes). The August-to-October shrinkage from ~$100B to $60B matters too — the same haircut could recur before books close.
What to watch: Anthropic's IPO — the junior tranche and the convertible notes are both gated on it, so a delay or reprice forces Broadcom to hold the exposure itself.
The debt layer broke first — issuance halved and real bonds are trading at real losses (Promoted)
AI-related debt issuance dropped to $23B in September from a record $113B in June, after $466B raised year-to-date — AI borrowers now account for roughly half of new US investment-grade issuance (The Debt Layer). And for the first time, specific instruments are derating on physical friction, not AI-demand doubts:
"Oracle 'Project Jupiter' (New Mexico): ~$18B of loans privately quoted ~85 cents on the dollar after Oracle issued a force majeure notice when the site struggled to gain electricity access."
— The Debt Layer
Tract Capital's Nvidia-leased Reno paper fell ~11 points to 90.53 on an NV Energy lawsuit; Zenith Arc's Oklahoma bonds now yield over 11% after a flood-plain permit stumble (The Debt Layer). This is the mechanism the map exists to find: a demand slowdown becomes a construction slowdown when the marginal project gets financed at 11%+ yields or not at all. Buyer behavior is shifting too — PGIM anchored a CLO capping AI collateral at 15% of the pool, the first with an explicit AI concentration cap.
What to watch: Whether September's issuance halt persists into Q4 — the SpaceX $40B raise and the Broadcom syndication will re-price the whole complex.
Oracle is the map's key node — the first hyperscaler whose capex became someone else's credit event (New — not on last cycle's map)
Quarterly capex of $28.5B against $23.1B of operating cash flow produced negative $5.4B of free cash flow in a single quarter; total debt hit $155.93B at 4.5x EBITDA, and S&P cut Oracle to BBB- with the five-year CDS at 203bps — an 18-year high (Oracle Debt and Capex+1). The stock is down 54.2% from its high even while revenue grew 29.6% and EPS beat every quarter.
The dependency underneath is the point: a $664B backlog heavily weighted to OpenAI, whose annualized revenue came in ~$20B below investor estimates — and whose data-center completions have already slipped from 2027 to 2028. Oracle's answer is to replicate the Broadcom playbook, negotiating with Apollo and Goldman Sachs to fund a separate chip-buying company it would lease from — keeping the purchases off a BBB- balance sheet (Oracle Debt and Capex+1). When the weakest IG name starts structuring around its own rating, the SPV is no longer an innovation; it's a workaround.
What to watch: Confirmation and rating-agency treatment of the Apollo/Goldman SPV — whether the debt gets consolidated back into Oracle's credit is the whole ballgame.
The chip layer's equity and the credit beneath it are priced in different worlds (New)
The divergence in one snapshot: NVDA sits 4.2% below its 52-week high with 0.2x leverage and $21.4B of quarterly free cash flow, while Oracle trades 54.2% down, CoreWeave 42.6% down at 13.6x leverage, and project bonds at 85–95 cents (Nvidia Circular Loop+1). Nvidia formalized its role as originator with a $500B financing platform alongside Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR — with reinsurance broker Howden Re working on structures to cover neocloud borrowers against collateral shortfalls (Nvidia Circular Loop+1).
The silicon layer's own demand shows zero slowdown — TSMC's September sales rose 54.6% YoY and Samsung posted a ninefold profit jump. That's why the map's domino scenario runs up the chain: a demand event at the AI-lab layer propagating into the collateral assumptions of the $500B platform. The Oct 8 OpenAI revenue scare was a small live rehearsal — roughly $500B of equity value erased in a day on a methodology dispute (Nvidia Circular Loop+1).
What to watch: TSMC (Oct 15) and ASML (Oct 14) earnings — the nearest checkpoint on whether silicon demand is decoupling from credit stress.
The equipment layer keeps proving the shorts wrong on fundamentals while derating anyway (Escalated — same thesis as last cycle, new bear case and new data)
Vertiv is down 35.4% from its high with no earnings miss to justify it — Q2 beat on EPS, leverage at 1.2x, and the Q3 setup guides to ~35% organic growth (Vertiv Equipment Layer+2). Upstream, the order books are accelerating, not stalling: Eaton's Electrical Americas orders run +41% on a 12-month rolling average, and GE Vernova's data-center orders in the first half already more than doubled its entire 2025 total (Vertiv Equipment Layer+2).
The bear case has shifted since last cycle — away from customer concentration (resolved then) and toward backlog quality: "I see risk that the massive $15 billion backlog Vertiv holds has actually been artificially inflated by supply chain panic... customers might have decided to double-order or place massive advanced orders simply to secure their place in line" (Seeking Alpha). The Oct 28 print — revenue against the $3.75B midpoint — settles whether "timing" becomes "pattern."
What to watch: Vertiv's Oct 28 revenue and deferred-revenue trajectory; a second miss converts the double-ordering hypothesis into the base case.
What surprised us
- [NEW] The $60B SPV settled 40% below the August headline. The prior cycle's open question ("$70–100B, unsigned") resolved at $60B plus a separate $42B convertible facility — and the pattern of headline shrinkage before close is itself a data point about where syndication appetite sits (Broadcom AI Financing Platform).
- [NEW] The first force majeure was about electricity, not AI demand. Oracle's Project Jupiter loans broke on grid access — the power layer's supply constraint showing up as a credit event in someone else's bonds. The unwind, if it comes, may arrive through permitting and power friction before it arrives through revenue.
- [NEW] TPU collateral is structurally worse than GPU collateral. The $42B senior tranche is secured by custom ASICs with essentially one buyer class — frontier AI developers — so recovery value depends entirely on frontier-AI market health at the moment of unwind (Broadcom AI Financing Platform).
- [NEW] A private-company accounting dispute can move half a trillion dollars. The OpenAI $50B-vs-$70B gap was a gross-vs-net methodology difference, and it still erased ~$500B of equity value in a day — the revenue base under the entire debt layer is private, unaudited, and methodology-dependent until the IPOs land (AI-Lab Revenue and IPO Gate).
Open threads
Closed or absorbed from last cycle:
- Broadcom SPV closing/terms — closed: resolved at $60B, absorbed into the Broadcom–Anthropic section above.
- Vertiv Q3 Oct 28 — carried forward, still open, now with the double-ordering hypothesis as the specific test.
Current threads worth a vote:
Dropped without resolution (no longer tracked): Vistra's Q3 print and preferred-redemption executions; Meta's Q3 2026 capex guidance revisions.