OpenAI-Broadcom custom chip project stalls; Broadcom demands Microsoft purchase guarantee before funding

OpenAI's Project Nexus—a collaboration with Broadcom to design and deploy 10 gigawatts of custom AI inference chips (first chip codenamed 'Jalapeno') by 2030—has hit a critical financing roadblock. Broadcom has declined to fund the first phase (estimated $18 billion for 1.3 GW capacity) unless Microsoft commits to purchasing roughly 40% of the chips and installing them in its own data centers before leasing back to OpenAI.

Microsoft has reserved data center space but has not formally agreed to purchase, citing disagreement over infrastructure design: OpenAI envisions specialized centers optimized for custom silicon, while Microsoft prefers standard, versatile data center architecture. An internal OpenAI memo noted that Microsoft 'has always held the leverage' in the chip strategy. The full Nexus plan could cost up to $180 billion in chip production alone (not including data center construction or power).

For architects evaluating AI infrastructure strategy and capex planning, this delay signals that even well-capitalized labs face financing constraints on custom silicon at scale. The absence of a Microsoft purchase agreement exposes OpenAI's dependency and raises questions about timeline: Jalapeno is not expected until 2027, adding pressure to Nvidia's near-term margin outlook if OpenAI defers custom silicon investments.