- The Information reported that Broadcom won’t finance production of OpenAI’s first custom AI chip (“Jalapeno”) unless Microsoft commits to buying about 40% of them.
- The first phase costs roughly $18 billion; Broadcom is using Microsoft’s creditworthiness as a safety net for the manufacturing commitment. Microsoft hasn’t agreed yet.
- OpenAI manager Sachin Katti called the Microsoft dependency “financially unattractive” and a long-term drag in an internal message, though he noted the arrangement would only apply to the first chip.
- The full project, codenamed “Nexus,” targets 10 gigawatts of data-center capacity and could cost up to $180 billion in chip production alone. Jalapeno is designed to run OpenAI’s models more efficiently than Nvidia’s hardware but isn’t expected until 2027.
What Happened
Broadcom won’t build OpenAI’s custom AI chip unless Microsoft buys 40% of them, The Information reported, summarized by The Decoder on May 9, 2026. The first phase with Broadcom costs around $18 billion. Broadcom won’t finance production unless Microsoft commits to purchasing about 40% of the chips, essentially using Microsoft’s creditworthiness as a safety net. Microsoft hasn’t agreed yet.
Why It Matters
The disclosure is the most concrete public detail to date on OpenAI’s custom-silicon ambitions, and it surfaces the same “private-company-can’t-be-priced” problem that drove SoftBank to slash its OpenAI-backed loan from $10 billion to $6 billion (covered separately today). Broadcom’s reluctance to finance manufacturing on OpenAI’s credit alone — requiring Microsoft as a counterparty backstop — illustrates how OpenAI’s pre-IPO status complicates major capital-intensive partnerships. The “financially unattractive” framing from internal OpenAI manager Sachin Katti is unusually candid for an AI lab pursuing a strategic priority.
Technical Details
The full project, codenamed “Nexus,” targets 10 gigawatts of data-center capacity and could cost up to $180 billion in chip production alone. The first chip, “Jalapeno,” is designed to run OpenAI’s models more efficiently than Nvidia’s hardware. Jalapeno isn’t expected until 2027. The first-phase $18 billion cost covers manufacturing setup, fabrication contracts, and initial production runs through Broadcom’s ASIC design and supply-chain organization.
The Microsoft 40% commitment requirement: Broadcom needs assurance that the chips will be purchased even if OpenAI’s commercial trajectory wavers. Microsoft, as OpenAI’s largest cloud customer and equity investor, is the natural counterparty. Microsoft hasn’t agreed yet — the negotiation is ongoing. Sachin Katti’s internal message called the dependency “financially unattractive” and a “long-term drag,” though he noted the arrangement would only apply to the first chip. The team is “pushing ahead for the strategic upside.”
The pattern parallel: it’s the second time OpenAI’s status as a private company has complicated a major deal in a single week. SoftBank reportedly slashed a planned loan — originally backed by OpenAI shares — from $10 billion to $6 billion because lenders couldn’t reliably value a private company like OpenAI. Both situations resolve cleanly with a public listing; OpenAI is reportedly considering an IPO later this year.
Who’s Affected
OpenAI faces a constrained custom-silicon program contingent on Microsoft’s commitment. Microsoft gains additional leverage in its OpenAI relationship at a moment when Microsoft has been reportedly diversifying its AI partnerships beyond OpenAI. Broadcom maintains the Microsoft-as-counterparty discipline its credit committee requires; the company gains $18 billion in first-phase manufacturing if the Microsoft commitment closes. Nvidia gains time: every quarter Jalapeno’s 2027 timeline slips means continued OpenAI Nvidia GPU consumption. The broader custom-silicon market — Google TPU, Meta MTIA, AWS Trainium/Inferentia — gains a more complicated competitive picture if OpenAI’s chip program proceeds slowly.
What’s Next
The Microsoft commitment decision is the immediate strategic milestone. Watch for whether OpenAI’s reported IPO progress accelerates as a way to resolve the credit-pricing problem that’s affecting both Broadcom and SoftBank deals. The “Nexus” 10-GW program scale ($180 billion) is comparable to OpenAI’s reported $725 billion combined hyperscaler AI commitments — meaning custom-silicon represents a meaningful but not dominant piece of total AI infrastructure spend if it proceeds at full scale.
