NVIDIA 2026 GTC Taipei

Two months ago, I decided to drop out of the University of Melbourne and return to Taiwan to research Agentic AI at Academia Sinica. Throughout the process, I had been full of doubts about my choice, especially with all the online chatter saying software engineers are going to be unemployed, it made me wonder whether I should just finish my master’s degree and head back to Taiwan to catch the last wave.

But after listening to Jensen Huang’s talk today, I regained my confidence in the industry. I took notes while watching the keynote and used Gemini to reorganize them afterward. Here are the notes:

1. Agentic AI and a New Era of Software Engineering

• The Rise of Agentic AI and Demand for Software Engineers: AI is not replacing software engineers. On the contrary, because engineers’ productivity (output) is growing several-fold, they can generate enormous economic value (turning $3 trillion in wages into $9 trillion in productivity). As a result, companies will tend to hire more software engineers [30:01].

• Tokens as the New Unit of Profit: Compute power (Watts) is now directly equivalent to revenue. The essence of an AI factory is a token generator, and tokens have become the core unit for measuring demand and profit [30:51].

• The Four Pillars of Agent Architecture (LLM + Harness + Tools + Runtime): Future applications will no longer just call a single large language model, but will be “agents” composed of four components:

  • Brain (Model): The large language model responsible for thinking, reasoning, and planning.

  • Coordinator (Harness): The mechanism that helps link and coordinate paths.

  • Tools & Skills: NVIDIA is turning its most powerful treasure — the CUDA-X libraries — into tools that agents can call directly.

  • Runtime: The underlying infrastructure that allows agents to operate safely [41:25].

• Memory Management: Since agents need to continuously process observations, reasoning, and multiple tasks, the retrieval and compression of working memory (short-term), long-term memory, and KV Cache have become the key to revolutionizing storage systems [44:01].

2. Infrastructure Revolution: Vera Rubin and Vera CPU

• Vera Rubin Enters Full Production: This is a multi-rack supercomputer system purpose-built for the Agentic era [01:01:37].

• Why Do Agents Need a Dedicated CPU? Past CPUs were designed for “humans,” while future CPUs are designed for “billions of impatient agents that demand extremely low latency.” When an agent calls a tool or database, it needs nanosecond-level response times [01:13:30].

• The 4 Core Features of Vera CPU (4 things to remember):

  1. Extremely High IPC (Instructions Per Clock): Focused on single-threaded performance and low latency, capable of fetching, decoding, and executing 10 instructions per clock cycle [01:17:00].

  2. Extremely High Bandwidth per Core [01:17:21].

  3. Cross-section Bandwidth: Reaching the speed-of-light limit at 3.6 TB/s [01:18:15].

  4. No Chiplet Tax: No cross-boundary transmission latency between cores, ensuring data flows extremely efficiently between the CPU, storage devices, and GPU [01:18:22].

3. Enterprise AI and Chip Design Agents

• Agent Toolkit for Enterprise AI (NeMo): NVIDIA launched an enterprise agent toolkit that includes open models, a Harness, tools, and a high-security Open Shell sandbox runtime, enabling every enterprise to build its own dedicated agent operating system [01:29:56].

• Chip Design Super Agent: NVIDIA partnered with Cadence to develop a chip design super agent. By letting the agent read RTL code and automatically call simulators and formal verification tools, it dramatically reduces debugging and verification cycles from “weeks” to “hours” [01:33:01].

• Nemotron-3 Ultra (SSM Architecture): The latest open model released — the world’s first model combining the hybrid architecture of SSM (State Space Model) and MoE (Mixture of Experts). It is 5x faster, reduces inference costs by 30%, and is suitable for long-running reasoning tasks [01:38:30].

4. The Future PC and Localized AI Supercomputers

• Windows Partnership and RTX Spark: After 40 years of development, the PC is about to be reshaped. NVIDIA unveiled RTX Spark, a laptop chip developed in collaboration with MediaTek, redesigning the personal computer for agents while maintaining 100% compatibility with Windows and the CUDA ecosystem [01:45:30].

• MSI EdgeXpert-59STW Desktop Supercomputer: A showcased MSI AI desktop supercomputer that can run agents at home 24/7, with no “token billing anxiety,” and can link and control all the devices in the home.

  • Regarding your note’s viewpoint: Jensen Huang did indeed compare this type of device to the once-ubiquitous “Home Theater” or game console [01:56:15]. However, as you observed, the initial setup cost of such extremely high-compute hardware does reflect the social reality that early-stage technology development can exacerbate the digital divide and unequal resource distribution.

5. Physical AI and Autonomous Driving

• Computer is Data: In the field of physical AI, the biggest challenge is “data,” because most videos are from a third-person perspective rather than the robot’s first-person perspective [01:59:52].

• Cosmos 3 World Model: An open frontier model (World Model) purpose-built for physical AI. It can integrate pixels, actions, sounds, and language to synthesize videos and build closed-loop simulators — the foundation for robots to understand how the physical world operates [02:01:05].

• Alpamayo 2: The world’s first open model with reasoning capabilities purpose-built for “autonomous driving.” While driving, the system can semantically reason about road conditions and make decisions in real time, just like a human [02:04:56]. (The “aplamayo” in the notes should be Alpamayo.)

• Isaac GR00T: NVIDIA’s reference design platform for general-purpose humanoid robots, integrating a complete software-hardware stack for data generation, simulation, training, and deployment — so academia and industry don’t need to build environments from scratch [02:08:46].

(Note: The Taiwan: GMI mentioned in the notes refers to GMI Cloud, a regional AI cloud infrastructure partner supported by NVIDIA in Taiwan [54:25].)