The AI company is reportedly negotiating a debt package that would fund about 1.5 gigawatts of Google-developed TPUs through a special-purpose vehicle
Anthropic is offering Blackstone a yield above 10% to help pay for its next big batch of AI chips. The offer sits inside a much larger negotiation: a $60 billion debt financing package reportedly known internally as “Project North Star.”
Under the proposal, Blackstone would put $9 billion into the deal’s junior debt tranche.
The negotiations were reported on October 16-17, 2026. The financing is structured in layers, like a cake where some slices carry more risk than others.
Senior debt sits at the top of that cake. It gets paid first if things go sideways, so it earns less: roughly 7%, according to the research findings.
Junior debt sits underneath. It absorbs losses before the senior lenders do, which is why Blackstone’s slice is priced at more than 10%.
The senior lenders have an additional cushion. Their portion benefits from residual-value support provided by Broadcom , a form of backstop tied to the value of the underlying hardware.
The money has a specific shopping list. Funds are earmarked for roughly 1.5 gigawatts of tensor processing units, the custom AI chips developed by Google .
Those chips would be acquired through a special-purpose vehicle, or SPV. Think of it as a separate corporate box that owns the hardware and carries the debt, so Anthropic can use the compute without piling the full obligation onto its own balance sheet.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
This is not Anthropic’s first trip to the private credit counter. In June 2026, the company struck a $35 billion arrangement with Apollo and Blackstone to fund about 1 gigawatt of compute.
Around the time of that first deal, Anthropic’s credit profile sat one notch below investment grade.
The bigger ambition behind all this borrowing is scale. Anthropic’s expected compute capacity target is over 20 GW by 2028 .
The SPV approach is becoming a defining feature of how AI labs pay for compute. Moving the hardware into a separate vehicle lets lenders lend against the chips themselves, not just against Anthropic’s promise to pay.
That framing helps explain the Broadcom residual-value support on the senior side. Lenders want to know that if the borrower stumbles, the collateral still holds value, which makes the senior tranche easier to sell at around 7%.
For Blackstone, the trade-off is straightforward. It gets a yield that beats the public high-yield market by a comfortable margin, in exchange for sitting behind senior lenders in an AI company that has not yet gone public.
For Anthropic, the yield is the cost of speed. Raising $60 billion in debt without diluting equity holders lets it buy compute now. The SPV structure also keeps the obligation at arm’s length from its core balance sheet.
The research findings frame the package as a sign of strong investor appetite for AI-related debt, noting that lenders are reportedly becoming more selective about which capital spending plans they will fund.
The chips here are Google-developed TPUs, not the GPUs that dominate much of the AI conversation. A 1.5 GW commitment is a meaningful vote of confidence in Google’s custom silicon as a serious option for large-scale training and inference.