The Buildout Goes Off-Balance-Sheet — Broadcom Backstops a Customer's $29bn Compute Lease, Amazon Reaches for a Delayed-Draw Loan, and Oracle Beats a Record and Falls 13%

The Lead

Last issue closed on a thesis that had been hiding in plain sight: the AI buildout went on credit, and a record-printing semiconductor complex was selling off because the trade had quietly become a credit story, not an equity one. This week the filings turned that thesis structural — and specific.

Three balance sheets that did not need the help reached for engineered financing in the space of a single week. Broadcom’s 10-Q disclosed a roughly $29bn, five-year non-recourse-like partnership with Apollo whose purpose is to backstop a customer’s AI compute-lease obligations — Broadcom is now arranging the financing that lets a large customer consume Broadcom’s own custom accelerators. Amazon — the company that has spent two decades funding itself from operating cash flow and the public bond market — took a $17.5bn delayed-draw term loan with a three-month commitment window. Ciena raised $2.875bn in zero-coupon convertibles. Vendor backstop, delayed-draw, zero-coupon: three different instruments, one direction of travel. The capital intensity of the buildout has outrun the income statements that are supposed to carry it, and the financing is migrating onto structured paper — leases, guarantees, SPV-shaped facilities, converts — where the obligations are real but the exposure is harder to see from the outside.

And the bellwethers told you why the structures are necessary. Dell printed AI-server revenue of $16.1bn, up 757% — and in the same 10-Q disclosed a hiring freeze and a $20.8bn purchase-obligation balloon. The demand is undeniable; the margin and the working capital are the problem. Broadcom’s backlog is now $164.6bn of contracted RPO, with the inventory already built to ship it. This is no longer a demand-forecasting exercise. It’s a financing one.

The tape ratified it. Forecast [2026-06-05-001] — that at least one of CrowdStrike or Oracle would beat consensus revenue and still close lower — confirms. Oracle reported a record quarter (revenue $19.2bn, +21%, ahead of the ~$19.1bn consensus; OCI +93%) and the stock fell roughly 13% over the next two sessions. CrowdStrike, by contrast, rose. “Great isn’t enough” held — and it held on the name with the most aggressive cloud-capex story in the index.

Threads

The buildout went off-balance-sheet — a $29bn vendor backstop, a $17.5bn delayed-draw, and the credit story turning structural

The single most important disclosure of the week is buried in Broadcom’s 10-Q: a financing partnership with Apollo, structured as non-recourse-like, with maximum exposure around $29bn over five years, whose function is to backstop a customer’s obligations under a five-year lease of Broadcom’s custom AI accelerators. Read that mechanism slowly. The merchant-silicon winner is no longer just selling chips; it is arranging the credit that allows a hyperscale customer to take those chips as a leased, financed asset rather than a cash purchase — with a third-party balance sheet (Apollo) absorbing the risk. That is vendor financing of compute at private-credit scale, and it tells you the customers want the capacity faster than their own cash flows or shareholders will comfortably fund.

It did not happen alone. Amazon’s 8-K disclosed a $17.5bn delayed-draw term loan with a three-month draw window and a three-year maturity — a structure built for lumpy, scheduled capital deployment (i.e. data-centre build phases), and a notable departure for a company that has historically scorned bank debt. Ciena termed out and raised through a $2.875bn zero-coupon convertible, repaying a term loan and explicitly earmarking the residual for supply-chain capacity.

The read-through: when the cash-flow king (Amazon), the merchant-silicon champion (Broadcom), and a profitable optical supplier (Ciena) all reach for engineered financing in the same week, the marginal dollar of the AI buildout is no longer coming off the income statement. It is coming off structured credit — which is exactly where stress, when it comes, will show up first and be seen last. Issue 11 said follow the debt. This week the debt went off the balance sheet.

Broadcom’s 10-Q is the tell, not the 8-K — $164.6bn RPO and a $2.1bn inventory build for custom accelerators

We covered Broadcom’s headline number last issue — AI revenue +143%, the stock −13%. The 10-Q filed this week is where the durable signal lives. Two line items: remaining performance obligations of $164.6bn, which converts the custom-silicon thesis from a growth-rate argument into a contracted backlog with multi-year visibility; and an inventory build of roughly $2.1bn over two quarters, explicitly earmarked for custom AI-accelerator shipments — not a vague “demand strength” comment but committed working capital against named future deliveries. Stack those against the Apollo structure and the picture is coherent: Broadcom has the orders ($164.6bn), has built the inventory to fill them, and is helping finance the customer that will take them. The custom-silicon ascendancy the system has been tracking since Issue 8 is now visible in three places on one balance sheet. The equity can still de-rate on margin and multi-sourcing (Google remains the swing risk); the business reality is no longer in question.

Dell’s tell — +757% AI-server revenue and a hiring freeze in the same filing

Dell is the cleanest bellwether in the set, and its 10-Q is a study in tension. AI-optimized server revenue went from $1.9bn to $16.1bn year-on-year (+757%), now 55.6% of ISG and 36.8% of total Dell revenue. That is the demand signal in its rawest form. In the same filing: purchase obligations ballooned to $20.8bn, with $17.3bn due inside twelve months — a working-capital commitment plainly tied to GPUs and HBM; a hiring freeze and workforce reorganisation despite 88% revenue growth; and a fresh $10bn buyback authorisation (≈$15.2bn total, around 20% of the company’s market cap). Management is simultaneously absorbing explosive top-line, rationing headcount, pre-committing billions of working capital, and returning capital. The translation is the one the convergence engine is screaming (DELL flags bearish despite the growth): this is real demand attached to a thin-margin, capital-hungry business, and the market is pricing the margin, not the revenue. One buried detail worth a footnote for later: Dell booked a $631m gain on a single early-stage strategic investee (a ~$0.6bn one-quarter revaluation) — a quiet mark-up in Dell’s venture book that nobody is discussing.

The physical layer cashes in — Vertiv buys ThermoKey, Ciena converts $2.875bn, the optical complex bifurcates

The picks-and-shovels are using the boom to integrate and add capacity. Vertiv completed its acquisition of ThermoKey, an Italian heat-exchanger manufacturer — vertical integration into the thermal layer that increasingly gates how fast GPUs can actually be deployed. Ciena’s convert (above) is, in part, a supply-chain-capacity raise. But the optical complex is splitting: convergence flags CIEN bullish while LITE and ANET read bearish — the demand is real and the names that can fund capacity and move up the stack (coherent optics, intra-data-centre) are separating from the ones the crowd already owns at full price. The lesson of Issue 11 — the sector holds while the crowded name bleeds — is now playing out within the optical and networking sub-complex, not just across it.

The Pre-IPO Watch — SpaceX files to go public, as an AI-compute company

The most consequential filing of the period is one the watchlist’s price engines can’t score, because the issuer is still private: SpaceX filed its S-1. Read for what it discloses rather than the rocket headline, it is an AI-infrastructure document. It confirms SpaceX’s merger with xAI earlier this year (over 440m shares issued; Tesla took a stake), a ~$1.25bn/month cloud-compute agreement with Anthropic (the COLOSSUS capacity), and an explicit strategic pivot to selling AI compute externally (“expect to enter into additional similar services contracts”) — alongside a buried ~$10bn Cursor-related termination liability and the $19.6bn EchoStar spectrum commitment. It belongs in this issue’s lead, not a sidebar: SpaceX/xAI selling monthly compute capacity to Anthropic is the same off-balance-sheet, compute-as-a-leased-asset economy that Broadcom’s Apollo structure and Amazon’s delayed-draw are financing — now arriving in a registration statement. The pre-IPO tripwire exists to catch exactly this; it caught it.

Since Issue 11 (5 June)

• Forecast [2026-06-05-001] — CONFIRMED. Predicted: at least one of CrowdStrike (~9 Jun) or Oracle (~10 Jun) beats consensus revenue and still closes lower. Oracle beat (record $19.2bn vs ~$19.1bn consensus, OCI +93%) and fell ~13% over the next two sessions; CrowdStrike rose, so it was Oracle that carried the call. The beat-and-fall regime is intact and it bit the most aggressive cloud-capex name in the group — consistent with the credit-over-equity read now running through this issue.
• The custom-silicon thread (Issues 8→11) graduated. What was a revenue-growth argument is now $164.6bn of Broadcom RPO plus a committed inventory build. The thesis is contracted, not forecast.

Watchlist Updates

• APO (Apollo) enters the frame — not as a watchlist name but as the counterparty that the compute-financing trade now runs through. If structured AI-compute credit is the story, the private-credit arrangers are the under-covered beneficiaries. Worth tracking as a read-through, not yet a position.
• DELL flagged bearish by convergence (466) despite +757% AI-server growth — the sharpest demand/margin divergence on the board this week. The setup to watch is whether the working-capital balloon ($17.3bn due <12mo) compresses cash conversion next quarter.
• TSM convergence reads bullish (309) on the foundry read-through; the merchant-ASIC and packaging demand that shows up as Broadcom RPO ultimately lands at TSMC.

The Week Ahead

The reporting lull favours the financing thread over the earnings one. Micron (MU) reports ~24–25 June — the cleanest memory read on whether the HBM/DRAM cost pressure building across the complex is a genuine sector-wide squeeze; FedEx (~23 Jun) and Accenture (~18 Jun) bracket the macro. The catalyst worth pre-positioning is structural, not a print:

Forecast [2026-06-13-001]: within 60 days, at least one more mega-cap or merchant-silicon name discloses a structured compute-financing arrangement — a vendor backstop, a customer compute-lease guarantee, or a delayed-draw / SPV facility explicitly tied to AI capex — beyond the Broadcom–Apollo and Amazon structures disclosed this week. Confidence 0.62, horizon 60 days. If it confirms, the off-balance-sheet financing of compute is a pattern, not three coincidences, and the private-credit arrangers (Apollo and peers) are the under-priced node. If 60 days pass with no further such disclosure, this week was idiosyncratic and the credit-structuring read is overstated.

Methodology and Disclosures

Threads are written from a structured dossier the system generates off the SEC-filing database (this issue: 20 filings, 29 signals across the 6–13 June window, 15 convergence setups). Figures are extraction-sourced from the underlying 10-Qs, 8-Ks and 6-Ks and should be verified against the primary filings before any trading decision; the Oracle result was cross-checked against the company release. Forecasts are pre-registered, dated and falsifiable, and graded in the following issue — the discipline is to be early and right, and to keep the score. As of this issue the system also maintains a machine-graded calls ledger (relationship-flip, divergence and radar calls — 39 currently open, awaiting their resolution windows) that runs the same falsify-and-grade loop beneath the editorial layer. Nothing here is investment advice.