Broadcom and Amazon test investor demand for AI capital expansion

Building the infrastructure for artificial intelligence has become so expensive that the companies doing it are now essentially running their own shadow banking operations. Broadcom agreed on Oct. 1 to lend Anthropic up to $42 billion to support a $125.2 billion TPU lease arrangement, a deal that could eventually let Broadcom convert that debt into an equity stake in one of the world’s most valuable AI startups.

Meanwhile, Amazon has issued over $62 billion in bonds this year alone to fund its own AI buildout.

The Broadcom financing machine

Broadcom’s lending agreement with Anthropic is structured to span five years, and it’s designed to position Anthropic as Broadcom’s largest custom-chip client by 2027. The debt-to-equity conversion option makes this more than a standard financing arrangement.

In June 2026, Broadcom formed strategic partnerships with Apollo Global Management and Blackstone to finance over 20 gigawatts of compute capacity for AI labs. The first tranche of that effort closed at $35 billion, projected to support nearly 1 gigawatt of compute for Anthropic alone.

By August 2026, Broadcom was reportedly in negotiations for upwards of $60 billion in debt financing, potentially escalating to $100 billion, channeled through special-purpose vehicles dedicated to AI chip financing.

Broadcom’s AI semiconductor revenue for fiscal 2026 is estimated at approximately $58 billion, driven largely by custom silicon and strong backlog visibility. The company is forecasting that figure to reach approximately $115 billion in fiscal 2027, with expectations to double it the year after.

Amazon’s bond binge

Amazon issued $37 billion in bonds in March and another $25 billion in July, bringing its total bond issuance this year past $62 billion. All of it is aimed at bolstering capital expenditure for AI infrastructure.

Total hyperscaler spending on AI infrastructure is projected to exceed $600 billion in 2026, with tech giants like Alphabet, Microsoft, and Meta all contributing to the arms race.

What the financing structures reveal

Broadcom’s special-purpose vehicles for AI chip financing resemble the kind of off-balance-sheet structures that have historically drawn scrutiny in other industries. They allow companies to take on significant leverage without it appearing directly on their corporate balance sheets.

The debt-to-equity conversion embedded in Broadcom’s Anthropic deal adds another layer of complexity. If Anthropic’s valuation continues to climb, Broadcom gets a windfall. If it doesn’t, Broadcom is left holding a very large receivable from a company that burns cash at industrial scale.

For Anthropic, the arrangement solves an immediate problem: funding the compute required to train and deploy frontier AI models without diluting existing shareholders through traditional equity raises. The $125.2 billion TPU lease is a staggering commitment, and having a chip supplier willing to finance a significant chunk of it provides breathing room during a period when AI companies are competing fiercely for compute access.

Broadcom’s targeting of over 20 gigawatts of compute capacity by 2028, in partnership with Apollo and Blackstone, represents a bet that demand for AI inference and training will continue growing at rates that justify building power-hungry data centers at unprecedented scale.