Anthropic's $36B Debt Deal Shows TPUs Are Now Cheaper Than GPUs for Inference

The $36 billion Apollo-Blackstone facility works like this: capital flows through an SPV to purchase Google TPUs, Anthropic leases them back, and Broadcom's residual-value guarantee means senior-tranche lenders are really underwriting Broadcom's credit quality. This lets Anthropic keep hardware debt off its books while lenders bet on contracted inference demand rather than corporate earnings. The TPU choice matters—Google's Ironwood generation runs 30-44% cheaper than Nvidia for inference, which is where most AI spending is headed by 2030.

The Market Noticed What This Says About Nvidia

Alphabet and Broadcom shares rose in after-hours trading. Nvidia didn't move. The deal closes days after Anthropic raised $65 billion in equity at a $965 billion valuation, giving the lab a capital structure that keeps hardware debt separate while it builds out data centers in New York, Texas, Louisiana, and Indiana.

| What people are saying | The evidence | How it's changing behavior | What I think | |----------------------|--------------|---------------------------|-------------| | TPUs work at scale now | Broadcom guarantees $31B in senior notes; 30-44% cost advantage over GB200 | Labs treat custom silicon as real collateral, not experiments | TPU adoption at this scale will push inference costs down faster than training-focused forecasts assume—bad news for Nvidia margins | | Multi-vendor compute is standard | Anthropic still uses AWS Trainium and Nvidia GPUs alongside the new TPU commitment | Sourcing from multiple vendors is operational necessity | Labs locked into Nvidia-only face rising relative costs and supply constraints | | Private credit is reshaping AI capex | SPV leaseback structure similar to Meta's Beignet bonds | Infrastructure debt now competes with equity for frontier lab funding | The market is underestimating how fast non-dilutive hardware financing becomes normal | | Nvidia's inference position is weaker than it looks | Bank of America projects inference will hit 75% of AI data-center spend by 2030 | Training dominance doesn't automatically transfer | Nvidia has maybe 18 months to close the TCO gap or lose share in the biggest spend category |

• Broadcom's backstop is the real innovation here. It gives the notes de-facto investment-grade status without Broadcom taking on direct liability. Other hyperscalers can copy this.
• The tranche split ($6B A1 / $25B A2 / $4.5B B) shows institutional demand concentrated at the senior, guaranteed level.
• Delayed-draw mechanics mean capital calls align with actual chip deliveries—lenders aren't paying for idle assets.
• Anthropic's revenue run-rate above $30 billion gives lenders the cash-flow anchor they need. Without that trajectory, this structure wouldn't have worked.

The idea that Nvidia owns training and will therefore dominate inference doesn't hold up here. Inference economics already favor purpose-built chips, and capital markets just validated that at record scale.

Significance: High
Categories: Market Impact, Partnership, Industry Trend

Verdict: Investors and enterprise buyers reading this as confirmation that TPU economics now have a structural advantage are ahead of the curve. Those assuming Nvidia's GPU moat extends unchanged into inference are behind.