South Korea’s Upstage in talks with AMD to buy 10,000 AI chips

South Korean AI startup Upstage is in talks with Advanced Micro Devices (AMD) to buy 10,000 MI355 AI accelerators as it looks to add more computing capacity in Korea.

CEO Sung Kim said he discussed the potential purchase with AMD CEO Lisa Su during a meeting in Seoul, and added that Upstage wants to diversify beyond Nvidia chips.

The discussions come as Upstage competes in a government-backed contest to pick Korea’s national AI foundation models, dubbed the “AI Squid Game.”

Models are assessed every six months under the Ministry of Science and ICT, with two finalists due by early next year, and winners receiving more Nvidia GPUs.

Kim said Upstage is preparing a roughly 200 billion-parameter model for the next round and is targeting Vietnam and the United Arab Emirates with in-country AI deployments.

🔗 Source: Bloomberg

🧠 Food for thought

Implications, context, and why it matters.

Investment-backed partnership already in motion

These discussions build on an ongoing relationship rather than starting a new one.
Advanced Micro Devices (AMD) invested in Upstage’s 62 billion won (about $41.58 million) Series B round in August last year, with other backers including Amazon in a separate $45 million Series B bridge round reported elsewhere 1.
The two CEOs met to “materialize and speed up cooperation” after that funding 2.
Upstage plans to deploy AMD’s Instinct MI355 AI accelerators over the next year to improve its proprietary Solar language model and create document-processing AI products 1.

Sovereign AI efforts and competitive chips reshape the market

The talks give AMD a path to compete with Nvidia in national AI infrastructure builds.
Upstage plans to use the accelerators for Korea’s “sovereign AI” push through a government-backed contest to select national AI foundation models, dubbed the “AI Squid Game,” which positions AMD inside a state-linked program 3.
More buyers can now consider alternatives since AMD’s hardware competes closely.
MLPerf Training v5.1 results place the MI355X platform close to NVIDIA partner submissions on B200-SXM and B300-SXM for FP8 training; on Llama 2-70B LoRA fine-tuning AMD cited gaps of about 3% versus B200 and 6% versus B300 4.
A 10,000-accelerator rollout would test AMD’s ROCm open software platform at scale and signal readiness to more of the industry.