AI Chip Spending Shifts to Debt as Broadcom, Oracle and SpaceX Line Up Billion-Dollar Financing

AI companies and their infrastructure partners are turning to debt to fund chip purchases at a scale that is reshaping how compute gets built. Broadcom has held preliminary discussions on arranging financing to help OpenAI buy the custom AI chips the two companies are developing, with Bloomberg pointing to around $30 billion of debt for the next phase and The Wall Street Journal reporting Broadcom is working on more than $50 billion. No formal process has begun and the size could still change. A separate $60 billion AI chip financing package partly backed by Broadcom is also moving forward with Anthropic among the main beneficiaries, including $42 billion of senior-secured debt with residual-value support from Broadcom and $18 billion of junior debt led by Blackstone, which has committed $9 billion. Bank of America, Citigroup and Morgan Stanley are involved in syndicating that financing.

The pattern extends well beyond custom silicon. Oracle is discussing chip-purchase financing with Apollo Global Management and Goldman Sachs, with one potential structure involving a separate entity buying the chips and leasing them to the cloud provider. SpaceX is separately seeking $40 billion to purchase NVIDIA AI chips, including about $10 billion of bank loans and $30 billion of investment-grade debt, with the transaction expected to close in 2027. Morgan Stanley estimates AI infrastructure will require $1.5 trillion of external financing through 2028. NVIDIA said in August it had partnered with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to establish financing platforms aimed at mobilising more than $500 billion of third-party capital over time, while Broadcom's AI XPV Platform with Apollo and Blackstone launched in June with an initial $35 billion transaction supporting more than 1 gigawatt of Anthropic compute.

Demand commitments keep landing alongside the financing. AM Intelligence, the AI infrastructure platform set up by promoters of Greenko, placed two further firm and binding orders for about 20,000 NVIDIA Rubin GPUs to be deployed as Vera Rubin NVL72 rack-scale systems across roughly 70 megawatts in India and Malaysia, with delivery scheduled for the second quarter of 2027. Together with an August commitment for 9,000 GPUs at its Hyderabad AI factory, AMI puts committed capacity at about 29,000 Rubin GPUs and close to 100 MW, and says it plans to bring an additional 300 MW to market over the next 15 months involving more than $20 billion of capital expenditure on top of $6 billion already committed. Both new sites will use RDMA over Converged Ethernet networking with high rack power density and full liquid cooling so successive silicon generations fit the same physical footprint.

The financing shift matters because it separates who buys the silicon from who owns it. Leasing and residual-value structures let cloud providers and AI labs secure capacity without carrying the full asset on balance sheet, but they also move depreciation and utilisation risk onto lenders and introduce a new sensitivity to chip residual values. With vendors already shipping denser liquid-cooled rack and deskside platforms on the current NVIDIA generation while Vera Rubin orders queue for 2027, the gap between committed compute and delivered compute is being bridged with borrowed capital.