Nvidia presented its "Agentic AI Stack" on Monday, November 9, at AI Day Seoul 2026 in Seoul — a stack of technologies for AI agents that carry out real work. The announcement was made by Jonathan Cohen, the company's vice president of applied research, in a keynote at COEX.
Update on AI Day Seoul 2026, which opened today focused on physical and agentic AI.
According to Cohen, agentic AI should be seen as a new computing platform. The stack brings together the prompt, the agent harness, model routing and data integration — and the central issue is inference cost, not just model capability.
Token use doubling every month
Cohen said the volume of tokens generated by language models "is doubling every month", in both proprietary models and open-weight models. The statement is an executive's keynote claim, not audited data — Nvidia did not disclose the measurement basis.
To manage that cost, the company highlighted Switchyard, a tool that automatically picks the right model for a task based on difficulty and cost. In Nvidia's internal evaluations, Switchyard kept an 80% task success rate while cutting costs to roughly one third of using only high-performance models.
Nemotron and Korea's sovereign AI
Cohen also said SK Telecom, LG AI Research and Upstage — the three finalists in South Korea's Ministry of Science and ICT sovereign foundation-model project — are using datasets from Nvidia's open Nemotron models. The project is set to pick two winners in early 2027.
Naver Cloud, a member of the Nemotron Alliance, shares data, task assignments and reinforcement-learning environments for pre- and post-training, and uses Nvidia technology in its HyperCLOVA X model. Nemotron 3.5 Super VL, a 120-billion-parameter open model released on October 15, was described as about five times faster than comparable open models — its launch was already covered by Geekiki in October.
Cohen also addressed security: routing tools send easy tasks to cheaper models, and the OpenShell sandbox aims to stop agents barred from the internet from slipping their limits by exchanging credentials with one another.
Image credit: Endeavor building, Nvidia's Santa Clara headquarters, photo from the NVIDIA Newsroom.
Sources
- Biggo Finance — Nvidia Unveils 'Agentic AI Stack'
- MoneyToday — Nvidia vice president says company will expand AI model development partnership with Naver
- Aju Press — Korea's sovereign AI finalists tap Nvidia data
Transparency: this article was produced with the assistance of artificial intelligence, cross-checked against public posts on X and sources on the internet. Please consult the original sources to confirm the information.
By Muse
