OpenAI's much-hyped custom silicon partnership with Broadcom has run into the messy reality of who actually pays for it. According to a report published Thursday by The Information, the first $18 billion tranche of OpenAI-designed AI chips is in limbo because Broadcom does not want to put up the financing alone — and Microsoft has not yet agreed to buy the share of processors that would unlock the deal.
Broadcom shares fell roughly 3% to 4% on the news before recovering part of the move, a reminder that any wobble in OpenAI's spending plans now ripples directly through the AI hardware complex.
Inside Project Nexus
The collaboration, internally referred to as "Project Nexus," was unveiled in October 2025 as a 10-gigawatt buildout of OpenAI-designed accelerators, with deployments targeted to begin in the second half of 2026 and complete by the end of 2029. The custom inference chip at the center of it carries the codename "Jalapeno" and is engineered to run ChatGPT and other OpenAI workloads more efficiently than general-purpose Nvidia hardware.
The stalled tranche covers the first 1.3 gigawatts of capacity, which The Information reports carries an $18 billion price tag.
The Microsoft Sticking Point
The holdup is structural, not technical. Broadcom is reportedly willing to bankroll the initial wave only if Microsoft commits to purchasing about 40% of the chips, or if OpenAI brings in other major buyers. The proposed arrangement would have Microsoft buy the accelerators, install them in its own data centers, and then rent the capacity back to OpenAI.
Microsoft has not signed a firm purchase agreement. Reporting also points to a deeper architectural disagreement: OpenAI wants specialized data centers tailored to its custom silicon, while Microsoft prefers more standardized, versatile facility designs. According to The Information, Broadcom "is not particularly happy about the risk that Microsoft might not choose to buy the chips and thinks it is OpenAI's mandate to ensure Microsoft does."
Broadcom has reportedly already softened its traditional dollar-for-dollar matching requirement, agreeing to invest more capital upfront than OpenAI — a notable break from what insiders described as a "long-held hard-line requirement."
Why It Matters
The snag lands at an awkward moment for the AI capex narrative. Hyperscaler infrastructure spending is widely projected to exceed half a trillion dollars in 2026, and OpenAI's projected operating costs through 2029 run into the hundreds of billions. Custom silicon is central to bending that cost curve down and reducing dependence on Nvidia.
For now, OpenAI's options are narrow. It can restructure the financing tranche, recruit alternative anchor buyers, or trim the rollout — each of which would reshape one of the most ambitious chip programs ever attempted by an AI lab.
