OpenAI and Broadcom's $18 Billion Chip Financing Stalemate

According to an exclusive report by The Information, OpenAI and chip design giant Broadcom announced a joint development of custom AI chips last fall, characterizing the cooperation as "a done deal." However, an internal memo obtained by The Information and two informed sources revealed that after several months, the financing arrangements for this deal remain unresolved, with the core issue being Microsoft's delay in making purchase commitments.

Financing Structure and the Root of the Problem

According to the aforementioned memo, the first phase agreement being negotiated between the two parties involves approximately **$18 billion** for chip production corresponding to 1.3 gigawatts of data center capacity. Based on this calculation, the full 10-gigawatt "Nexus" plan could cost up to **$180 billion** for chip production alone, equivalent to the power supply capacity of five Hoover Dams.

The core of the financing obstacle lies in the fact that Broadcom demands Microsoft's agreement to purchase about **40%** of the new customized chips before it's willing to finance the initial production. Under this arrangement, Microsoft would install the chips in its data centers and then sublet them to OpenAI for use.

However, according to an individual familiar with the negotiations, although Microsoft has reserved some data center space for these chips, it has **not yet made a purchase commitment**. This uncertainty has prompted OpenAI to begin exploring emergency plans to find alternative buyers.

The memo cites internal statements from OpenAI's executives stating: "The persistent project risk lies in the fact that Microsoft may eventually not purchase OpenAI's chips."

Microsoft's "Leverage"

The negotiation details reveal Microsoft's ongoing influence over OpenAI—despite the two parties having signed multiple agreements in the past year, granting OpenAI greater freedom to collaborate with other partners. The memo acknowledges that Microsoft "has always had leverage over the chip project," with OpenAI having spent "a significant amount of energy and time" striving for a clearer purchase commitment.

In an internal communication last month, Sachin Katti, OpenAI's executive in charge of cloud business and chip projects, questioned the deal, pointing out that the financing structure based on Microsoft's purchase commitment "will seriously tie our hands in the long run," and that the related business structure "lacks economic attractiveness." "We will still move forward, solely for its strategic value," he wrote.

Project Schedule Delays

OpenAI's first custom chip, codenamed "Jalapeno," is designed specifically for inference workloads, aiming to replace NVIDIA's general-purpose GPUs and reduce operational costs. Originally planned to be put into use in the second half of this year, it is now expected by several informed sources that most Jalapeno chips will only be **ready by 2027**. The design work for the second-generation chip, codenamed "Serrano," has also been launched in tandem.

The memo points out: "Broadcom's TSMC capacity constraints and the need to support 2027 demand constitute a sense of urgency."

Broadcom Concession, Negotiations Still Advancing

Despite the obstacles, negotiations are still moving forward. Broadcom, which had long insisted on OpenAI's 1:1 co-funding ratio, has recently decided to break this "long-standing firm demand," willing to invest more capital up front. The two sides are currently working hard to sign a "conditional agreement" so that Broadcom can "confidently" reserve the corresponding capacity with TSMC.

Broadcom's stock price fell by about 3% that day, then partially recovered lost ground. Over the past year, driven by the demand for custom chips from companies such as OpenAI, Meta, and Microsoft, Broadcom's stock price has more than doubled.