← Back to feed

Broadcom is reportedly seeking to borrow up to $100 billion to finance new AI chip commitments — a figure that would make it one of the largest debt raises in the history of the technology sector.

Bloomberg reported on August 20 that the structure includes $60-70 billion in senior notes and approximately $30 billion in junior notes, with Blackstone and Apollo Global Management among the lenders. The proceeds are earmarked for Anthropic and at least one other unnamed company — likely OpenAI, which has its own active chip partnership with Broadcom.

The Structure

Senior notes are first-in-line debt: paid before other obligations in a bankruptcy. Junior notes are subordinated — they yield higher interest to compensate for the repayment risk. Running both tranches simultaneously suggests Broadcom is optimising for total raise size rather than cost of capital. The company may guarantee a portion of the debt.

Broadcom is not borrowing to fund its own operations. The raise finances infrastructure spend by AI labs that will run on Broadcom silicon. The chipmaker designs and supplies the network switches, host bus adapters, and custom AI accelerators that sit inside the data centres being built. More capital deployed into those data centres means more Broadcom silicon shipped.

What This Adds to the AI XPV Platform

In June 2026, Broadcom, Apollo, and Blackstone launched the AI XPV Platform — a joint investment vehicle that committed $35 billion to Anthropic specifically for data centre construction. That $35 billion is expected to put more than 1 gigawatt of computing capacity online in 2026, with a stated goal of 20+ gigawatts by 2028.

The new $100 billion raise is separate from that platform but draws on the same investor relationships. It implies the original $35 billion commitment is insufficient to meet demand — or that additional labs beyond Anthropic are being brought under the same financing umbrella.

Broadcom’s broader AI infrastructure relationships now span three fronts:

• Anthropic: AI XPV Platform ($35B, June 2026), ongoing network silicon supply
• Google: Custom TPU design partnership extended to 2031 (April 2026); multi-gigawatt TPU compute for Anthropic
• OpenAI: Jointly designed Jalapeño inference accelerator (June 2026), optimised to reduce data movement; OpenAI expects better inference throughput than current GPUs

The $100 billion raise, if completed, would sit alongside all three relationships as a new debt layer enabling each to scale faster than their own balance sheets allow.

Why Debt, and Why Now

AI infrastructure spending has moved beyond what any single company’s equity can absorb at the required pace. A frontier training cluster runs north of $10 billion to build and several billion more to operate annually. The labs need capital that can move fast; investors want exposure to AI infrastructure with contractual returns rather than equity upside.

Debt solves both problems. It provides predictable yield to investors (Blackstone and Apollo are infrastructure specialists who have built this model in telecoms and energy). It provides low-friction capital to Broadcom, which can then guarantee the loans are backed by real silicon shipments and engineering commitments.

Broadcom itself is not capital-constrained — it projects more than $100 billion in annual revenue. What it is doing is intermediating between AI labs that need capital and financial institutions that want to lend against real infrastructure assets. The chip commitments function as collateral.

Scale Context

The $100 billion figure is larger than the GDP of most countries. For reference:

• The US interstate highway system cost approximately $500 billion in today’s dollars, over four decades
• TSMC’s planned US fabs total roughly $65 billion across multiple sites
• NVidia’s market capitalisation at its 2024 peak was approximately $3 trillion

What is being financed here is not a single facility or chip line. It is an ongoing commitment to build and operate compute infrastructure at a scale that was not seriously contemplated in AI investment projections as recently as two years ago.

Whether the debt gets fully deployed, and on what timeline, will depend on lab demand for compute continuing to grow. Given that labs are announcing new training runs every quarter, and that model scaling still shows measurable returns, the demand picture looks stable through at least 2027.