OpenAI Taps Samsung Foundry for Next-Gen AI Chips Alongside TSMC Partnership

Samsung Foundry appears poised to join TSMC in manufacturing OpenAI's forthcoming artificial intelligence accelerators. During remarks at a Seoul press conference this week, Harrison Kim, General Manager of OpenAI Korea, disclosed that the company intends to deepen its ties with Samsung beyond existing memory and software arrangements. The disclosure indicates OpenAI plans to engage Samsung Foundry for production of certain next-generation processors, a move that would establish dual sourcing and underscore the organization's substantial manufacturing capacity needs.

One of the areas where we have made the most progress and gained the most recognition with Samsung Electronics is our joint production and research on the next-generation chips we are developing

Harrison Kim, General Manager of OpenAI Korea

OpenAI maintains an established artificial intelligence ASIC initiative that leverages Broadcom's design expertise alongside TSMC's wafer fabrication and sophisticated packaging capabilities. The organization's inaugural inference accelerator, branded Jalapeño, underwent development by OpenAI's internal team, received co-design support from Broadcom, and entered production through TSMC—all within an 18-month window.

The specifics of Samsung's involvement remain unclear. OpenAI has not clarified whether Samsung would serve as an alternate manufacturer for the existing Jalapeño line, handle production of a separate processor currently in development, participate in a jointly designed chip, or contribute to other manufacturing and design functions. While Samsung and SK hynix already furnish memory components for OpenAI's Stargate data center project, collaborative chip design and fabrication represent a distinct undertaking from memory supply arrangements.

The original Jalapeño generation will probably not receive dual sourcing across TSMC and Samsung, given that this chip is presently undergoing mass production at TSMC and OpenAI's statements reference 'next-generation chips' rather than current production models. Development of Jalapeño's successor is progressing toward tapeout, positioning it for imminent mass manufacturing. Given OpenAI's explicit reference to 'next-generation chips,' the organization may indeed route its second-generation inference ASIC through Samsung Foundry.

A strategic alliance announced in late July between Samsung Electronics and Broadcom, valued in excess of $200 billion through 2030, targets cooperation on cutting-edge foundry, memory, and packaging infrastructure for AI systems. Since OpenAI maintains an arrangement with Broadcom for 10GW of custom AI accelerators, the company could manufacture these systems at both Samsung and TSMC facilities—contingent on Broadcom handling the necessary design adaptation work for Samsung production.

OpenAI might adopt a strategy comparable to Tesla's approach by securing Jalapeño's successor from both TSMC and Samsung to achieve greater manufacturing throughput. The volume dynamics between Tesla and OpenAI, however, may differ substantially.

Tesla requires exceptional AI5 processor volumes because the company intends to deploy the chip across three distinct high-volume product categories: AI infrastructure, automobiles, and Optimus robots. Tesla could potentially require millions of AI5 units for vehicles alone, independent of robot and data-center requirements. Establishing separate TSMC and Samsung physical designs provides Tesla both supply-chain stability and the cumulative capacity to serve multiple product lines.

Data center accelerators, however, demand substantially more silicon per unit relative to ASICs destined for vehicles or robots. Should OpenAI and Broadcom's forthcoming ASIC represent a substantial leading-edge processor incorporating multiple dies, and OpenAI requires gigawatts of capacity, the silicon demand could exceed TSMC's available production—particularly given the foundry's current commitments to companies like AMD and Nvidia. Under such circumstances, OpenAI might require a second foundry partner to procure sufficient quantities. These remain speculative considerations at present.

Source: Tom's Hardware