Add preferred source
South Korean AI startups Upstage and Rebellions have each held meetings with Nvidia executives, exploring potential collaboration. Upstage CEO Kim Sung-hoon met with Madison Huang, senior director at Nvidia and eldest daughter of CEO Jensen Huang, to discuss LLM collaboration. Rebellions CEO Park Sung-hyun held his first meeting with Jensen Huang. Industry observers are focusing on the possibility of technical cooperation combining GPUs and NPUs rather than mergers and acquisitions. The South Korean government is also pursuing heterogeneous computing demonstrations linking AMD GPUs with domestically developed NPUs. However, significant hurdles remain, including how much core technology to disclose and national core technology regulations.
Key Elements
South Korean artificial intelligence (AI) startups are making successive contacts with Nvidia, exploring potential collaboration. With confirmation that the heads of AI model developer Upstage and AI semiconductor designer Rebellions have each met with Nvidia executives, attention is focused on how a cooperative framework centered on the GPU ecosystem will unfold.
According to industry sources on the 14th, Upstage CEO Kim Sung-hoon recently met with Madison Huang, senior director of Omniverse Robotics, at Nvidia’s headquarters in California. Madison Huang is the eldest daughter of Nvidia CEO Jensen Huang. The meeting was arranged when Madison Huang invited Kim, who was on a business trip to the United States, to the headquarters.
At the meeting, Kim said, “Nvidia asked what they could do to help Upstage, which excels at large language models (LLMs). We discussed collaboration, and Nvidia said they would review it positively.”
Upstage recently passed the second-stage evaluation in the South Korean government’s independent AI foundation model project. With a focus on refining production-ready models and expanding overseas ahead of the third-stage evaluation, collaboration with Nvidia could be a critical variable for model performance improvement and global expansion. Upstage is currently in the process of raising pre-IPO equity investment.
Rebellions Holds First Meeting with Jensen Huang
Prior to Upstage, Rebellions CEO Park Sung-hyun of the AI semiconductor startup reportedly met with Jensen Huang last month. The two met at Nvidia’s headquarters in Santa Clara, though specific details of the discussion were not disclosed.
Rebellions stated, “This is the first time Park has met with CEO Huang,” adding, “We are cautious about disclosing the purpose of the meeting.” While the market has even raised the possibility of Nvidia acquiring Rebellions, industry sentiment leans toward the likelihood of technical cooperation over immediate mergers and acquisitions (M&A).
One industry insider said, “They’re probably in the exploratory stage with each other,” adding, “Discussions could include examining whether system performance or efficiency can be improved by pairing Rebellions’ Neural Processing Units (NPUs) alongside Nvidia’s Graphics Processing Units (GPUs).”
Rebellions is a fabless AI semiconductor design company founded in 2020. The company name “Rebellions” means “revolt,” and it has been regarded as a leading contender for the “K-Nvidia” title. In March, it was selected as the first direct investment target of the National Growth Fund, securing 250 billion won (approximately $185.9 million) from the Advanced Strategic Industry Fund.
GPU and NPU: Coexistence, Not Competition
The potential for collaboration between Rebellions and Nvidia is drawing attention amid recent shifts in the AI server market. While GPUs broadly handle large-scale computation, NPUs excel at processing specific operations such as AI inference with relatively lower power consumption. Data Processing Units (DPUs) manage networking and data movement, reducing the burden on GPUs.
Nvidia itself does not insist on GPUs alone. Its latest rack-scale AI system, the “Vera Rubin NVL72,” incorporates 72 Rubin GPUs, 36 Vera CPUs, and BlueField-4 DPUs together. Late last year, Nvidia signed a non-exclusive licensing agreement with Groq, a U.S. company with fast AI inference technology, and at GTC 2026 in March, it unveiled the Rubin GPU alongside the inference-focused “Groq 3 LPU.”
Industry observers believe this architecture could be applied to Rebellions as well. Just as Groq LPUs are paired alongside GPUs, Rebellions NPUs could be used together, with workloads distributed based on use case. For Rebellions, experience operating services alongside Nvidia GPUs could serve as a reference for entering U.S. big tech data centers.
South Korean government policy aligns with this direction. In July, the Ministry of Science and ICT partnered with AMD to launch a “heterogeneous AI computing” demonstration connecting AMD’s CPUs and GPUs with domestically developed NPUs in a single system. Five South Korean NPU companies are participating, including Rebellions, FuriosaAI, DeepX, Mobilint, and HyperExcel.
Challenges: Technology Disclosure and Regulation
If actual technical cooperation begins, there are hurdles to overcome. Running GPUs and NPUs in a single system requires aligning software (SW) and interconnect methods. Determining which operations go to which chip and coordinating data exchange protocols raises the question of how much core technology each side must disclose.
An industry insider noted, “To work together, each side needs to understand the other to some degree, but ultimately the question is how much information to open up. You can’t show everything, including your core technologies.” Nvidia provides CUDA to developers but retains platform control. South Korean NPU companies are also building their own SW as a competitive advantage, making boundary-setting a key issue.
Regarding M&A possibilities, regulatory variables also exist. If Rebellions’ technology falls under national advanced strategic technology or national core technology designations, the approval process of the Ministry of Trade, Industry and Energy and technology protection issues could become obstacles in any acquisition by a foreign company. The case of FuriosaAI CEO Baek Joon-ho rejecting an acquisition offer from Meta last year is being raised again in the same context.
Industry observers attribute Nvidia’s strengthening collaboration with South Korean startups to the expansion of the GPU-based AI ecosystem. From Nvidia’s perspective, participation from related companies is essential for ecosystem building, while South Korean startups can expect tangible benefits including stable GPU supply and access to global networks. However, no specific technical cooperation or acquisition between Nvidia and the two companies has been confirmed as currently in progress.
Once added, BigGo Finance appears first in Google Search Top Stories, so you get the broadest, most up-to-the-minute, and most comprehensive global financial news first.
