When OpenAI and chip designer Broadcom announced their collaboration last fall to develop custom artificial intelligence chips, they presented it as a finalized agreement. The companies stated the deal would bring enough chips online by 2030 to consume 10 gigawatts of power—equivalent to the output of five Hoover Dams—aiming to reduce OpenAI's costly reliance on NVIDIA hardware.
What was not disclosed was that the financing mechanism for the project remained unresolved. Months later, according to an internal memo and two individuals involved in the negotiations, the companies are discussing an arrangement where Broadcom would fund the initial phase of chip production. This stage would require 1.3 gigawatts of data center capacity at an estimated cost of $18 billion. Extrapolating from this, the full 10-gigawatt project, codenamed Nexus, could entail chip production costs alone reaching $180 billion, excluding data center construction and other expenses.
Securing financing is critical for OpenAI, as the ChatGPT developer anticipates its operations will consume over $200 billion by 2029. Utilizing its own chips is central to OpenAI's strategy of reducing server costs and improving gross margins.
However, negotiations have encountered a potential obstacle. An OpenAI executive informed colleagues in a memo last month that Broadcom indicated it would only finance the first phase if Microsoft agreed to purchase approximately 40% of the chips. Microsoft would install the chips in its own data centers and then lease capacity back to OpenAI.
One person involved in the talks noted that Microsoft, with its decades of data center experience and top-tier creditworthiness, would provide Broadcom the confidence needed to recoup its investment. Yet, the memo suggested Microsoft might ultimately decline to purchase the chips, which would alter the project's financing terms.
Draft agreements for the chip deal reportedly include a clause requiring OpenAI to find alternative buyers if Microsoft's purchases fall short. This scenario appears plausible. Although Microsoft has reportedly reserved some data center space for the chips, it has not yet committed to purchasing them, according to a source familiar with the discussions.
As of last month, however, Broadcom executives seemingly held a different understanding of the situation. The memo stated that because OpenAI and Microsoft "had not disclosed the full nuances of our relationship, Broadcom did not fully recognize this risk and believed requesting a purchase commitment from Microsoft was within [OpenAI's] control." The memo highlighted the "ongoing programmatic risk that Microsoft ultimately does not follow through" on purchasing the chips.
Verifying a purchase order from Microsoft for OpenAI-designed chips would presumably be straightforward for Broadcom, given the existing direct business relationship between Broadcom and Microsoft involving custom AI chips.
The memo indicated that, as of last month, Broadcom and OpenAI were working towards signing a "conditional" agreement to allow Broadcom to confidently allocate manufacturing capacity at Taiwan Semiconductor Manufacturing Company (TSMC) for OpenAI's chips. "Broadcom faces timing pressure for securing scarce TSMC capacity to support 2027 demand," the memo noted. TSMC also manufactures advanced AI server chips for NVIDIA.
OpenAI has a precedent of announcing landmark partnerships before finalizing details. For instance, one month prior to the Broadcom announcement, OpenAI stated that NVIDIA would provide up to $100 billion in funding for OpenAI to build its own data centers powered by NVIDIA chips. That headline-grabbing deal ultimately did not materialize, though NVIDIA later made a $30 billion equity investment in OpenAI.
Concurrent with that investment, OpenAI agreed to utilize 5 gigawatts of NVIDIA's advanced systems. This commitment could strain the data center capacity available to OpenAI's partners, potentially limiting its ability to deploy its own custom chips.
In January 2025, OpenAI announced Stargate, a joint venture with SoftBank and Oracle to develop data centers with a projected spend of $500 billion. However, that effort has stalled due to disagreements among the partners and lender hesitancy to finance multi-billion dollar projects directly tied to a company whose business model remains unproven.
