Apollo, Blackstone, and Broadcom Built a $35 Billion Machine to Finance Anthropic’s Compute

A special-purpose vehicle, three debt tranches, a vendor backstop, and a five-year lease: structured finance has arrived at the AI compute layer — and the risk architecture deserves a careful read.

What Happened

According to Financial Times reporting on the deal’s tranche structure — corroborated by Apollo’s own release and subsequent coverage from Axios — Apollo Global Management and Blackstone, coordinating with a group of global banks, arranged roughly $35 billion in private credit, one of the largest such facilities ever assembled. The vehicle at the center: a compute special-purpose vehicle that purchases approximately one gigawatt of Google-designed, Broadcom-co-designed TPU pods and leases them to Anthropic on a five-year term. The hardware sits on the SPV’s books, not Anthropic’s — a structurally deliberate choice as the AI lab moves toward an IPO.

The SPV is funded through three tranches. A senior “A1” slice of $6 billion prices at roughly one percentage point over Treasuries. A second senior tranche of $24 billion carries a 5.75% coupon. A junior, first-loss tranche of $4.5 billion prices at 8.5% — anchored by Apollo and Blackstone, who absorb losses before any senior noteholder is touched. Broadcom provides a residual-value and deficiency guarantee on the approximately $30 billion of combined senior debt: if Anthropic defaults and the chips are resold, Broadcom covers any remaining shortfall on the senior notes. It is a contingent guarantee — not funded cash on Broadcom’s balance sheet, and not a standalone interest guarantee — triggered only by both an Anthropic default and a resale shortfall.

The pods are scheduled to deploy at Fluidstack data-center sites beginning mid-2026, representing the first slice of what Broadcom is calling the “Broadcom AI XPV Platform” — an aspirational program targeting more than 20 gigawatts of capacity through 2028, with OpenAI named as another prospective tenant. That 20-gigawatt figure is a stated ambition, not a contracted commitment, and the gap between the two matters for how one reads the risk.

The Structural Read

Strip this deal to its bones and it is infrastructure securitization: an SPV owns the asset, an operator leases it, and the cash flows are carved into tranches sold to yield-hungry private-credit investors — exactly how toll roads and power-purchase agreements get funded. What Apollo and Blackstone have done is import that architecture into AI compute, which is the most capital-intensive layer of the AI stack and the one operating companies most need kept off their books. Anthropic’s earlier Volta compute arrangement showed the direction of travel; this deal is the full institutional expression of it.

The load-bearing piece is the Broadcom backstop, and it is worth being precise about what it does and does not do. The contingent guarantee covers any senior-note shortfall remaining after the TPU pods are resold following an Anthropic default. It is a dual-trigger instrument — default plus resale gap — not a standalone coupon guarantee and not funded reserves sitting on Broadcom’s balance sheet. What it does, structurally, is convert speculative AI hardware into paper that institutional investors can hold at roughly one percentage point over Treasuries. That compression of yield is the whole mechanism: Broadcom’s creditworthiness standing behind fast-depreciating chips makes the senior notes look like infrastructure debt even though the collateral behaves nothing like a bridge or a power line. This is the vendor-financing pattern the AI Capex Map has been tracking — chip suppliers underwriting their own customers’ demand — now rendered in full structured-finance detail.

The flywheel of aligned incentives is real: Google sells the pods, Broadcom earns on the co-design and collects on the guarantee premium, Apollo and Blackstone collect 5.75–8.5%, Fluidstack operates the sites, and Anthropic gets a gigawatt of compute as a lease expense rather than a $35 billion liability. Every party is rationally positioned — which is precisely why the risk sits where no single party is forced to confront it directly.

The honest risk accounting runs as follows. A TPU pod that is best-in-class in mid-2026 may be a generation behind by 2028 and two generations behind by the time the five-year lease expires. Infrastructure securitization has a long track record precisely because bridges and transmission lines do not become obsolete — their residual values are stable and estimable. Used TPU resale markets are thin and effectively untested at this scale. If Anthropic were ever to default and the pods went to market, the gap between book value and realized proceeds is genuinely unknown, which means the gap Broadcom would be required to cover under the guarantee is genuinely unknown. At one gigawatt that is a manageable, if novel, tail risk. At twenty gigawatts — if the platform reaches its aspirational scale — the same contingent guarantee structure begins to concentrate tail risk on a single vendor’s balance sheet in a way that warrants systemic attention. This is how idiosyncratic risk becomes structural: not through any single actor’s recklessness, but through a rational structure that scales faster than the market’s ability to stress-test it.

The off-balance-sheet mechanics deserve acknowledgment without over-dramatizing them. Keeping $35 billion of compute obligation off Anthropic’s books is a legitimate tool — operating leases have been standard corporate finance for decades — and it does not make the obligation disappear. Sophisticated IPO investors will model the lease commitment. But it does mean that Anthropic’s pre-IPO financials present a cleaner capital structure than the underlying compute dependency would suggest, which is worth holding alongside the structural elegance. The compute floor Anthropic has been building is real and strategically important; this deal locks it in at institutional scale while distributing the financing risk outward.

Three Implications