KEY POINTS
- Broadcom dropped 4% on May 7 after The Information reported that its $18 billion custom AI chip deal with OpenAI is stalled over Microsoft's refusal to commit to buying 40% of the first production run.
- The standoff exposes a deeper tension in AI infrastructure buildout: OpenAI wants custom silicon optimized for its models, but Microsoft wants standardized data center designs it can repurpose.
- Watch Broadcom's next earnings call and any Microsoft statements on AI capital allocation—if the deal restructures or collapses, it could reset market expectations for the entire custom silicon thesis.
Broadcom shares fell 4% on Wednesday after The Information reported that the company's blockbuster custom AI chip partnership with OpenAI has hit an $18 billion financing wall. The stumbling block is straightforward: Broadcom will only finance the first phase of chip production—consuming 1.3 gigawatts of data center capacity—if Microsoft agrees to purchase roughly 40% of the output. Microsoft has not made that commitment.
The deal, unveiled in October 2025 and originally sketched as a $500 billion multi-year hardware buildout encompassing 10 gigawatts of custom AI accelerators, was supposed to mark OpenAI's pivot away from dependence on Nvidia's general-purpose GPUs. Custom chips optimized specifically for OpenAI's model architecture could deliver better performance-per-watt and lower inference costs at scale. The economics made sense on paper. The financing, evidently, is proving harder.
The Microsoft Problem
An internal OpenAI memo, reported by The Information, laid out the core tension. Under the proposed structure, Microsoft would install Broadcom's custom chips in its own data centers and rent the compute capacity back to OpenAI. That arrangement requires Microsoft to build data centers to OpenAI's specifications—specialized facilities optimized for custom silicon rather than the standardized, fungible designs Microsoft prefers for Azure.
Microsoft has spent the past two years pouring over $100 billion annually into cloud infrastructure built around general-purpose Nvidia GPUs that can serve Azure customers, OpenAI workloads, and Copilot inference simultaneously. Locking 40% of a new chip run into OpenAI-specific facilities limits that flexibility. For Microsoft's capital allocation team, the math does not obviously work.
The standoff also reflects a broader unease among lenders. Banks and institutional investors who would backstop the financing are growing cautious about AI infrastructure returns arriving on schedule. The gap between AI capital expenditure and AI revenue generation remains the single biggest unresolved question in tech investing, and an $18 billion bet on custom chips for a company that has yet to turn a profit at scale tests the limits of even the most optimistic underwriting.
What It Means for the Custom Silicon Thesis
Broadcom's custom chip business—which also serves Google, Meta, and two undisclosed hyperscalers—has been one of the strongest narratives in semiconductors. The company's AI revenue grew over 200% in fiscal 2025, and its customers are planning to add roughly 10 gigawatts of data center capacity through 2027. A collapse or significant restructuring of the OpenAI deal would not destroy that business, but it would remove the largest single incremental revenue driver from forward estimates.
For traders, the immediate read is that Broadcom's stock had priced in significant upside from the OpenAI partnership. The 4% decline is modest relative to the deal's size, which suggests the market is treating this as a negotiation hiccup rather than a deal-breaker. If Microsoft agrees to a smaller commitment—say 20-25% of the first run—Broadcom could still finance the project, though at lower margins.
The bigger question is whether this signals a ceiling on AI capex. If the world's best-funded AI company cannot secure $18 billion in chip financing without its primary backer balking, what does that say about the dozens of smaller AI infrastructure projects competing for capital? Watch for Microsoft's next earnings commentary on AI spending and any updates from Broadcom's June earnings call. The custom silicon thesis is not dead, but it just got its first serious stress test.
