South Korean AI startup Upstage is in talks with Advanced Micro Devices Inc. for a possible purchase of 10,000 MI355 AI accelerators. The planned deal comes as the company expands computing capacity for large language model training in South Korea.
It also comes at a time when Korean AI firms are trying to secure more hardware options for domestic projects.
Upstage Chief Executive Officer Sung Kim said he spoke with AMD Chief Executive Officer Lisa Su in Seoul last week about the possible chip purchase. The discussion focused on AMD’s latest MI355 accelerators, which are built for demanding AI workloads.
Kim said, “We have a lot of NVIDIA chips in Korea, but we want to diversify to other chips, including AMD’s.” The comment made clear that Upstage wants a broader supplier base as demand for AI infrastructure rises.
The proposed 10,000-chip order would give Upstage added capacity for model development. At the same time, it would give AMD a chance to expand its position in South Korea’s fast-growing AI hardware market.
Upstage is one of four teams in a government-backed contest to build South Korea’s leading national AI foundation models. The competition has been described as the “AI Squid Game,” linking it to the popular Netflix drama.
Under the program, a panel overseen by the Ministry of Science and ICT reviews each team’s progress every six months. The process removes weaker entries over time, while the final two teams are expected to be selected by early next year.
The winning groups will receive more NVIDIA graphics processing units. Even so, Upstage is trying to secure other hardware in parallel, which may help it avoid depending on a single supplier during a period of high demand.
Kim said Upstage is building a large language model with about 200 billion parameters for the next round this summer. The company is aiming to improve performance while keeping costs under control.
He said Upstage’s approach combines scale with efficiency. The goal is to remain competitive against lower-cost AI offerings from China while continuing to build advanced models for commercial and national use.
This strategy reflects a wider trend in the AI sector. Companies developing large models are looking for more ways to manage hardware costs without slowing training and deployment.
Also Read: US AI Chip Export Rules Push Nvidia, AMD, and AI Tokens Lower
Upstage is also targeting markets such as Vietnam and the United Arab Emirates with sovereign AI systems. These systems are designed to operate within national borders, which can appeal to governments and institutions that want local control over data and computing resources.
That international push adds to the company’s need for more processing power. A larger accelerator base would support both domestic model training and overseas deployments tied to sovereign AI projects.