Broadcom has started raising $60 billion in debt to finance the development and supply of AI chips for Anthropic, according to a report by Quartz. The figure ranks among the largest debt financings ever mounted in the semiconductor industry, and it ties the company's balance sheet directly to a single AI customer relationship.

The financing plan emerged from reporting by qz.com, which states that Broadcom is beginning the process of raising the $60 billion to back AI chip production for the San Francisco-based AI developer. Neither Broadcom nor Anthropic has publicly detailed the terms, tranches, or timeline of the debt issuance at the time of writing.

The scale of the number matters. Sixty billion dollars approaches the cost of an entire leading-edge fab cluster. For comparison, TSMC has committed roughly $65 billion to its greenfield manufacturing complex in Arizona across three phases. A debt raise of this size to support chips for one customer signals that custom AI silicon programs have moved from engineering projects into capital-intensive commitments on par with physical infrastructure builds.

Broadcom has built its AI business around designing and supplying custom accelerators — application-specific integrated circuits — for large cloud and AI companies, alongside its networking chips that move data through AI clusters. Anthropic, best known for its Claude family of large language models, competes with OpenAI and Google in frontier AI development. Demand for training and inference compute at these companies has outstripped conventional procurement, pushing them toward dedicated chip designs rather than merchant GPUs alone.

The debt-based structure of the financing deserves attention. Raising capital through debt rather than equity or internal cash flow suggests Broadcom expects the Anthropic program to generate predictable, contract-backed revenue streams large enough to service the borrowing. That assumption rests on the durability of Anthropic's compute demand — and, by extension, on sustained growth in frontier AI model training and deployment.

It also concentrates risk. A $60 billion commitment to a single customer relationship exposes Broadcom to the commercial fortunes of one AI lab in a market where model-training roadmaps, cloud strategies, and chip-supplier allegiances have shifted repeatedly. Any material change in Anthropic's spending plans, or in the competitive position of its models, would weigh directly on Broadcom's ability to recoup the investment.

For the wider supply chain, the move points to where capacity will flow. Custom accelerators designed by Broadcom are manufactured at leading-edge foundries, most notably TSMC on advanced nodes, using large-diameter wafers and advanced packaging technologies such as CoWoS-class substrates that remain among the tightest bottlenecks in AI hardware. A financing of this size implies long-term volume commitments extending through those constrained stages of production.

The competitive picture sharpens as well. OpenAI has pursued custom-silicon arrangements across multiple suppliers, and Google designs its own TPU accelerators. Anthropic securing dedicated chip capacity through Broadcom aligns it with the same logic: own or control the compute layer, and reduce dependence on merchant GPU supply. Each multi-billion-dollar commitment locks up design resources, wafer allocations, and packaging capacity that competitors cannot then claim.

The report does not specify how quickly Broadcom will complete the raise, which banks are leading it, or how the proceeds will be phased against chip development and delivery milestones. Those details will determine how quickly the capital translates into shipped silicon — and into revenue Broadcom can book against the borrowing.

If the raise proceeds as reported, Broadcom will have staked a substantial portion of its financing capacity on the thesis that custom AI accelerators for individual labs will become a durable, multi-year market rather than a concentrated bet on a handful of model builders whose spending power remains untested at this scale.