Lisa Su courts 13 Korean AI firms to build inference systems around AMD chips

AMD already sells its own AI accelerators, the MI455X among them ^[2]. The systems discussed in Seoul would put another company's inference chip beside AMD's CPUs and GPUs ^[14]. The Korea Herald's assessment is that the partnerships could widen AMD's inference offerings and give the Korean chipmakers a route to global customers ^[22].

In the version AMD is pitching, the demonstration projects produce their blueprint and the paired systems ship abroad ^[16]. Buyers of inference hardware then get one more supplier outside Nvidia, and AMD sells CPUs and GPUs around a chip it did not design ^[14]. In a smaller version, some of the Korean firms port their software to ROCm and keep selling on their own. AMD would gain developers and little hardware revenue ^[15]. In the least ambitious version, the inference track stays a government program built on the cooperation agreement AMD signed with the science ministry in July, and the lasting business from the visit is memory ^[18].

For now I'd put the most weight on the third, because the dated commitments are in memory. Besides the HBM4 designation, AMD and Samsung signed a memorandum in March on memory supply that also opened foundry talks ^[4]. Samsung already ships HBM3E and HBM4 for AMD's latest accelerators ^[3]. On Wednesday, Su said AMD continues to explore memory and foundry partnerships with Samsung ^[5]. The inference systems are still at the talk stage ^[14][16].

Several of her remarks to reporters were about scarcity. AMD is "working with our customers to optimise their memory footprint so that we can satisfy more computing going forward," Su said, and she wants memory suppliers to "build faster, as fast as possible" ^[9][7]. "I think we very much rely on the Korean supply chain," she said ^[10]. Yonhap reported that shortages across a broad range of memory products are straining the expansion of AI infrastructure ^[11].

Beyond the research center, AMD will keep supporting graduate-level AI training and is investing in local computing capacity, including its first cloud infrastructure project with MangoBoost ^[20]. The reports do not describe any AMD investment in the 13 companies or any terms for the paired systems. Counting from Wednesday's meetings on Oct. 7 to the center's year-end deadline, AMD has 85 days to choose a site and open it ^[21][13][19].

The case against my weighting comes from Su herself. "I am a big believer in an open ecosystem," she said in Seoul ^[17]. On the sidelines of AMD's Gamer Day she said, "Semiconductors is a team sport. A team sport means there is collaboration" ^[6]. Thirteen companies across chips, software and cloud services make a long list of possible ROCm adopters for a platform up against CUDA ^[8][15]. Suppose one of the chipmakers ships ROCm support, or a paired system reaches a named customer outside Korea, before the center opens. Then the inference track is further along than the paperwork shows, and the memory-first reading is wrong.