Compute is being tested as financeable collateral

Edition 11 argued that technical and economic obsolescence have begun to diverge — a chip generation can keep earning monetisable cash flow past its frontier-performance life. That mattered because lenders ultimately underwrite cash flows and recovery values, not chip architectures. The closing line: compute scarcity is the signal; the capital structure is the risk.

This transaction is among the first large-scale tests of whether capital markets are willing to make that same distinction — whether lenders, not just chipmakers, will take a view on what a used accelerator is worth.

On 5 June 2026, Apollo Global Management and Blackstone closed a US$35 billion debt package to buy Google's custom TPUs (Tensor Processing Units, co-designed with Broadcom) and lease them to Anthropic [REPORTED — FT/Bloomberg]. It is among the largest private credit transactions on record. But the size is not the story. The structure is.

This financing keeps the acquisition debt at the SPV rather than on Anthropic's balance sheet. An SPV (special-purpose vehicle), formed by Apollo's Atlas SP Partners, raised the debt and equity, bought the chips, and leases them to Anthropic. The lease payments service the debt. Ownership has been separated from usage — a financing pattern familiar from aircraft leasing, telecom towers, and fibre infrastructure, now applied to compute through structured capital-markets execution at a scale, and with a vendor residual backstop, not previously seen. GPU-backed lending already exists — CoreWeave, Lambda, and Crusoe have all raised financing secured against accelerator assets. What is new here is the combination: transaction scale, tranched capital-markets execution, vendor residual support, and ownership separation occurring together on this asset.

It is worth asking why the chips were financed externally rather than held on Google's own balance sheet. One plausible economic rationale is the same one that has historically moved ownership in aircraft, towers, solar, and rail: capital efficiency, alongside tax, regulatory, and specialization factors. A compute provider does not necessarily want every deployed accelerator sitting on its own balance sheet, tying up capital that could fund the next generation.

The load-bearing detail sits in the pricing. The US$35B was split into three tranches. The two senior tranches — US$6B of A1 notes and US$24B of A2 notes — carry residual-value support from Broadcom, the chips' co-designer, and priced accordingly: A1 to a group of banks at roughly 100 basis points over Treasuries, A2 at a 5.75% coupon [REPORTED]. The third tranche — US$4.5B of B notes, with no Broadcom support — priced at 8.5% [REPORTED]. The pricing gap suggests investors assigned meaningful value to the residual-value support, in addition to the differences in seniority, structure, and other credit protections.

Hold the infrastructure analogy honestly, because it is doing a lot of work and it does not fully hold. Aircraft, towers, and fibre became financeable for specific physical reasons: they are long-lived, depreciate slowly and predictably, and trade in deep secondary markets, so a lender can observe what the asset is worth if the borrower fails and the asset must be sold. That observable residual is what lets a lender underwrite the asset rather than the borrower.

TPUs are the opposite asset on every one of those dimensions. Service life is short. Depreciation is fast — and, per the last edition, economic and technical obsolescence are now diverging, which makes residual value harder to model, not easier. Custom silicon has a narrow set of potential resale buyers, so there is no established secondary market and therefore no observable residual to underwrite against.

The residual-value support is a synthetic substitute for the secondary market that does not exist. Broadcom effectively substitutes its balance sheet for the price discovery the secondary market cannot yet provide, and the senior debt prices off the vendor's credit instead of the chip's resale value. That is what separates this transaction from traditional equipment finance: the market is not yet confident enough to price the collateral on its own.

The pricing differential suggests investors distinguished counterparty risk from residual-value risk, rather than treating compute demand as the only variable. Spread reflects many things — duration, seniority, the attachment point, covenants, liquidity — so no single inference should be overloaded. But the structure is built around one specific protection: residual-value support on the senior tranches. The structure suggests recovery value has become a central underwriting question alongside demand for compute — demand has not disappeared; recovery is the variable layered on top of it.

This structure, if it scales into an asset class, is being built onshore in the US first. The maturation pattern is global in principle — compute may follow aircraft and towers toward financeability — but the conditions that made this specific deal financeable are not evenly distributed, and the APAC obstacle is not what it first appears.

The bottleneck is probably not the absence of a physical regional resale market — residual support could come from Nvidia, Broadcom, the hyperscalers, cloud marketplaces, or remarketing specialists, none of which has to be local. The harder constraints are legal and financial. The first is enforceability against export controls: in a default, can a lender repossess the chips and sell them across jurisdictions for enough to repay the notes? Advanced AI chips remain subject to evolving US export-control, licensing, and end-user compliance rules — and notably, the Biden-era tiered framework that once formally classified markets like Singapore, Malaysia, and India was rescinded in May 2025, with a replacement still being worked out [REPORTED]. Export-control regimes can materially narrow the pool of lawful buyers in a default scenario, potentially lowering recoverable value — which works against the very residual the deal is built to protect. The second is institutional credit appetite for technology-risk paper; the third, investor depth sufficient to absorb tranches of this size. Geography, in the end, is merely the visible expression of those underlying constraints — enforceability, credit, and depth are the variables that actually bind.

Two forward signals are worth watching. The proposed Chip Security Act, advanced toward a House vote in March 2026, would shift enforcement from licensing paperwork toward on-chip location tracking [REPORTED] — which could, paradoxically, make financed chips easier to locate and repossess even as it polices their movement. And the Affiliate Rule — stayed from November 2025 until November 2026 [REPORTED] — means the beneficial ownership of any SPV holding the chips is itself an export-compliance variable. The Uptime Brief's standing read holds: financing is not the same as access. Capital is willing; the legal and institutional plumbing is not yet built.

So the honest framing is narrow on purpose. Capital markets are beginning to test whether compute can be financed the way infrastructure is — and they are doing it, so far, only where a vendor will stand behind the residual and the security is enforceable. That is a US-first experiment, not a settled global asset class.

This architecture was built during a period of unusually strong private-credit liquidity. What happens to a compute SPV when the residual backstop holds but the financing market that priced it doesn't — when the question is no longer whether the chips hold value, but whether anyone will refinance the paper that bought them?

Edition 11 asked how compute should be underwritten. Edition 12 asked how compute can be financed. Edition 13 asks what happens when those financing assumptions meet a tighter capital market.

Edition 13 picks that up.