$71 Billion in Hidden Leverage: Anthropic's SPV Is the Oldest Trick in Finance

Here's a number that should stop you cold: $71 billion. That's the size of the hardware debt Anthropic has stacked inside a special purpose vehicle. It is not on the company's balance sheet. The lease obligations are structured so the model lab's financial statements stay clean ahead of a public listing that could value the company at $965 billion. The risk is real. It's just moved somewhere else. Let me walk through the mechanics, because this structure—not the AI models—is what actually matters. Anthropic isn't buying chips. It's renting Google TPUs through an SPV. The vehicle owns the hardware and leases it back to Anthropic. Broadcom—the chip design partner behind Google's TPU line—provides a residual value backstop. That's an insurance policy against the hardware depreciating into nothing when the lease matures. Apollo and Blackstone are the lenders. First tranche: $35 billion. Second deal in negotiation: another $36 billion. Total exposure: $71 billion and climbing. The economics are brutal, and revealing. Anthropic's annualized compute spending is roughly $45 billion. Annualized revenue: $19 billion. That's a 2.37x compute-to-revenue ratio. Traditional hyperscalers run this kind of cost at single-digit percentages of revenue. OpenAI sits around 1–1.5x. Anthropic is burning capital at a rate no AI player has attempted before. This is not a company paying for compute as an operating cost. This is a company financing its future on leveraged hardware. The term sheet tells you what the market thinks. The senior tranche gets investment-grade pricing—protected by Broadcom's backstop. The Class B tranche, $4.5 billion without backstop support, pays 8.5%. That's roughly eleven to twelve times leveraged credit pricing. The market is telling you exactly what it thinks of Anthropic's stand-alone credit risk. It's not flattering. I've audited capital structures that looked cleaner than this one. The 2017 ICO era taught me the simplest lesson about system vulnerabilities: you don't inspect the headline mechanism. You inspect what happens when the mechanism fails. Code is law until the audit reveals the trap. The trap inside this SPV isn't the vehicle itself. It's the triple assumption stacked underneath. Assumption one: AI application revenue keeps growing enough to justify a 2.37x compute-to-revenue ratio. Assumption two: TPU hardware retains enough residual value in five years to satisfy Broadcom's backstop. Assumption three: Broadcom's investment-grade rating holds through a stress scenario. Break any one, and the $71 billion chain reaction doesn't stop at Anthropic's doorstep. It floods the entire AI infrastructure ecosystem. Now let's talk about what nobody is addressing. Broadcom's new role is the biggest story hiding in this deal. The company has moved from chip supplier to AI credit intermediary. By backstopping the residual value, Broadcom is underwriting Anthropic's core business risk. That's not a silicon bet. That's a credit bet. Broadcom is betting its balance sheet on Anthropic's success. If the lab stumbles, Broadcom's exposure is far greater than what a single component supplier should ever carry. That's moral hazard, priced into a hardware contract. Google's position is equally tangled. Google is Anthropic's largest investor. Google provides the TPUs. Google also builds Gemini—Anthropic's direct competitor. The SPV locks Anthropic into Google's hardware stack for years. It's a strategic masterstroke dressed as a favor. Anthropic gets compute. Google gets a captive customer whose technical roadmap becomes inseparable from its own. The conflict is structural. Both sides know it. Neither side can move. Compare the players, and the divergence gets sharper. OpenAI depends on NVIDIA supplier financing—simpler, but binds the lab to one vendor's roadmap. Google funds its own TPUs and data centers—flexible, but capital-intensive and slow. Meta builds internally and buys externally, constrained by revenue capacity. Small startups rent GPU time by the hour—maximum flexibility, worst unit economics. Anthropic's SPV structure sits in a different universe entirely: off-balance-sheet, credit-enhanced, and larger than any lab's own financial statements would ever support. This financing paradigm is not isolated. Volta Infra is signing $10 billion contracts. SpaceX is planning $100 billion in compute. NVIDIA carries $600 billion in supplier financing exposure to OpenAI. Samsung and Broadcom have a $200 billion supply chain deal. The pattern is clear: AI infrastructure has stopped being a corporate capital expenditure story and become a financial engineering story. The people who win are not the best model builders. They're the ones who can structure the best credit deals. Here's the part that should make you uncomfortable. That 8.5% Class B yield is DeFi's old promise wearing a suit. High yield for taking risk. Yield is the bait; exit liquidity is the hook. The SPV creditors already priced the risk—that's what the 8.5% coupon means. The IPO investors have not. When the S-1 filing drops in 2026, analysts will see a balance sheet with minimal hardware debt, because the obligations live in a vehicle that consolidated reporting can legally ignore. Apollo and Blackstone negotiated with full information. Public equity investors won't have that luxury. That information asymmetry is the most dangerous part of this deal. I lived this pattern in 2022. When Terra's peg started sliding, I shorted LUNA exposure while most retail money was still aping into yields nobody had audited. Liquidity doesn't disappear gradually. It dries up when the music stops. If AI growth stalls—if the killer app doesn't materialize, if enterprise adoption misses the curve—the SPV's lease cash flows stretch. Residual value support triggers become real obligations. The credit crunch that follows doesn't respect narratives. The TPU residual value question deserves more scrutiny. Google's hardware cycles move fast. If Google ships a new TPU generation before the lease term expires, the old hardware loses value. Fast. And this is custom silicon optimized for Anthropic's workload class. It's not like an NVIDIA GPU with a liquid secondary market and a thousand potential buyers. A TPU's plausible buyer list is roughly Google and Anthropic. That's not a market. That's a bilateral negotiation with one dominant counterparty. The state-level footprint—New York, Texas, Louisiana, Indiana—tells a different story. These locations were chosen for power infrastructure, not just compute density. Data centers take three to five years to build. Transmission lines take even longer. SPV money solves the hardware problem. It doesn't solve the plug-it-in problem. Financing moves faster than electrons. That timing mismatch is an underappreciated execution risk. And the regulatory angle? The SEC hasn't deployed regulation-by-enforcement here. It doesn't need to. The disclosure rules around off-balance-sheet vehicles are vague enough that Anthropic could plausibly file its S-1 without revealing the full texture of this leverage. The 2008 comparison to structured investment vehicles is uncomfortable but precise. SIVs existed to hide leverage. Regulators looked at clean balance sheets while the shadow system grew underneath. Then the housing market stopped appreciating. Nobody said the same thing can't happen to AI compute. Patience is for traders; timing is for killers. That's the frame for what comes next. Watch three things. First: whether the second SPV deal closes. If it does, total exposure passes $71 billion. Second: what Broadcom discloses about contingent liabilities in quarterly filings. That's the earliest signal of stress. Third: the S-1 text. The language around this SPV tells you exactly how management wants the market to understand the risk. Read the footnotes. The footnotes are always where the truth lives. If the IPO hits the $965 billion anchor, this SPV model gets cloned by every AI lab with a term sheet and a prayer. If the IPO fails, the 8.5% yield is the least of anyone's problems. The hardware gets repossessed. The backstop gets triggered. The whole AI trade reprices overnight. This is the new infrastructure era: not algorithmic breakthroughs, but capital structure design. The technology story is the product. The leverage is the engine. We don't buy narratives. We buy structures. This one is leveraged as far as a software company has ever gone. We built the machine. The only question left is whether it survives the drawdown.