AMD, FuriosaAI and MangoBoost sign MOU on heterogeneous AI infrastructure in Korea

U.S. chipmaker AMD, Korean AI chip startup FuriosaAI and AI infrastructure company MangoBoost signed a memorandum of understanding on November 4 at AMD's Advancing AI event in Seoul to explore building open, heterogeneous AI infrastructure: AMD CPUs and GPUs as the compute base, linked with FuriosaAI neural processing units (NPUs) and MangoBoost data processing units (DPUs) and system software, so chips from different vendors can work together in one system. Under the MOU, the three companies will build a reference deployment at MangoBoost's Seoul facility, targeting operations in 2027. (Cover photo is a file photo of AI chips, courtesy of CSET.)

Division of labour: GPUs, NPUs and DPUs each play a role

Under the MOU, AMD supplies CPUs and GPUs; FuriosaAI supplies NPUs for AI inference, built around its Renegade accelerator; and MangoBoost supplies DPUs and system software and will run the Seoul AI infrastructure. MangoBoost is also advancing its Korea-based Mango AI Factory, targeting deployment in the first half of 2027 on AMD Instinct GPUs. The signing was attended by MangoBoost CEO Jangwoo Kim, FuriosaAI CEO June Paik and Keith Strier, AMD's senior vice president for global AI markets.

Korea's science ministry met the same day on computing and a research centre

On the same day, South Korea's Ministry of Science and ICT held a meeting of its AMD-MOU operations committee at the same venue, reviewing progress on heterogeneous AI computing infrastructure, an AMD AI research centre and support for universities and AI talent. The locations and number of research centres remain subject to further discussion. It was the first time AMD has held its flagship Advancing AI event in South Korea, with LG Electronics and other Korean companies also taking part.

An MOU for now: performance and compatibility still to be proven

The agreement is a memorandum of understanding — an exploratory step, not a confirmed order or product. Once chips from different vendors are connected, performance, power consumption and software compatibility still have to be verified one by one. The selling point of heterogeneous computing is cost: MangoBoost and AMD have previously said multi-vendor chip combinations can cut the total cost of AI infrastructure to as little as a tenth of competing setups, a claim that has not been independently verified.

(Sources: MangoBoost announcement, November 4, 2026; Aju Press, November 4, 2026. The collaboration is at the exploration stage; the 2027 target for the reference deployment is the companies' goal, not a confirmed date.)