NVIDIA GTC 2025 Summary: Jen-Hsun Huang foresees the "AI Agent revolution" entering household consumer level, Blackwell Ultra may spawn a large number of solo startups.

Huida held the GTC 2025 developer conference, Huang Jenxun's speech was an AI agent from beginning to end, and it is expected to launch a consumer-grade Blackwell Ultra, targeting personal consumer users such as studios and families, and it can be said that it is foreseen that AI agents may replace manpower in many industries. (Synopsis: China's crazy grab RTX5090 "8 cards can train DeepSeek R1" Taiwan single-card gift package rose to NT$400,000) (Background supplement: Huida RTX 5090 exploded the first meltdown incident, what happened?) Graphics card scalpers "speculated to a sky-high price of 250,000 yuan" did not hit the gas) Huang Jenxun, founder of graphics card design manufacturer Nvidia, held a speech at the GTC 2025 developer conference held last night, in addition to announcing new GPU product lines, but also launched more open source and AI telecommunications, autonomous driving related new software architectures, and preached to developers that in the Nvidia cuda ecosystem, Nvidia GPUs have become an all-round productivity tool, developers can do it in GTC as well Nvidia related resources have been a lot, the following are a few face-oriented conference highlights organized by the dynamic area: Blackwell Ultra: Magic change for AI agent The architecture of the graphics card determines the transmission bandwidth of the graphics card, and the bandwidth of the graphics card determines the capacity supported by the graphics card memory, Huang Jenxun announced on GTC2025 that Blackwell's improved architecture "Blackwell Ultra" will be launched in the second half of the year, and the single display chip does not pass NV-LInk links can support 288 GB direct connections, which is three times that of the Blackwell architecture, which means that the scale of single-card scalable models is also greatly improved on AI inference models. Huang Jenxun also specifically mentioned the part of the AI agent in his speech, Blackwell Ultra's large display memory, enough to run multiple sets of high-efficiency model training, post-inference, and AI agents need multiple sets of different models to be connected together, the operation of agent AI needs to be orders of magnitude higher than the current one, Blackwell Ultra is born for this, this platform can easily perform training and inference. In addition to this year's architecture Blackwell Ultra, Huang Jenxun also announced the architecture names for 2026 and 2028, respectively, "Vera Rubin" and "Feynman", and the expected corresponding consumer products are RTX 6090 and 7090. Blackwell Ultra class on personal computers In the past, chips that support large display memory (video memory) are usually designed for commercial products used in server cabinets and AI water cooling kits, but Huang Jenxun specially announced the launch of DGX AI products on stage, and announced that these are Blackwell Ultra products for home and individual users, which will challenge and crush high-end MACs in terms of performance, and these products will be developed by Asus, Dell, HP and other cooperative manufacturers OEM. The details of DGX AI are not yet known, but according to Huang, NVIDIA's intention to create a high-end product that can enter the personal, home, and studio settings inevitably leads to speculation about his thoughts on the future AI market. If AI agents cause a revolution in labor and production, then it is very likely that the recent "ChatGPT" moment is spreading to the main body dominated by small companies, then replacing some traditional manpower with high-budget, small-volume computers may be a predictable market-oriented conclusion, to know that before Nvidia could not launch a small home computer with an asking price of millions of dollars, but it was for home individuals, the idea can be seen. AI operating system: Nvidia Dynamo In addition to the new high-computing hardware, Huang Jenxun also released the "AI operating system" Nvidia Dynamo that can optimize multi-graphics card scheduling, which will be provided to CUDA users in the form of open source, Dynamo can evenly distribute different computing needs through system design, strengthen coordination to each GPU, and achieve the characteristics of unchanged hardware and optimized performance. These are essentially AI factory operating systems, on which you can call CUDA resources very well. In addition, Huang also announced at the same time that it will open source the Llama Nemotron series of inference models based on AI agent capabilities, which will enable developers and enterprises to build AI agents in an open source environment. Huida Robot Prototype Isaac GR00T N1 In addition to AI agent software, Huang Jenxun also announced Nvidia's self-developed robot prototype Isaac GR00T N1, which is the world's first AI model and physics suite developed for "humanoid robots", including open source training models, common hardware specifications, and related development tools. At the same time, it was also announced that Nvidia will cooperate with Google and Disney to develop the "Newton Physics Engine" to optimize the robot's learning efficiency in physical tasks, it is understood that Newton is compatible with simulation architectures such as Google DeepMind and Isaac Lab, and the Newton engine will be launched later in 2025. Quantum-X Photonics Hardware: Photonic Switches In addition to the traditional silicon chip moat, Nvidia also announced their new breakthrough in the field of optoelectronic engineering, and TSMC's COUPE SoIC-X packaging technology to manufacture 65nm optoelectronic chips Spectrum-X Photonics, Quantum-X Photonics can increase power efficiency by 35 times and improve the signal by 63 times while the number of lasers is greatly reduced. At the same time, with Taiwan factories such as silicon products, Hon Hai, as well as foreign factories Corning, Sumitomo Electric, etc. to build a complete silicon photoelectric supply chain, it is expected that the market will launch a photonic Ethernet switch based on Nvidia chips in 2026, when the speed can reach more than 100 times that of today's mainstream routers. New self-driving partner: General Motors Since Jensen Huang first announced at CES in January 2025 that he will jointly develop self-driving software with Toyota, Jensen Huang also needs to announce that General Motors (GM), the second largest car company in North America, has also joined the Nvidia autonomous driving software AGX ecosystem. Nvidia is also strengthening NVIDIA Omniverse in NVIDIA Cosmos, creating a unique virtual physics engine environment to provide GM's factory production, as well as road driving, robot services and other related equipment simulation, in addition to autonomous driving safety, Nvidia's software and hardware integration can help these traditional car manufacturers complete the modern production process improvement. Summary: Nvidia is developing across the board, stressing AI agents In addition to the above, Huang Jen-hoon also mentioned the development of the next generation of wireless networks (6G), which has been dominated by Korean telecom operators since 4G, and now NVIDIA has also announced that it will cooperate with major telecom companies to develop new 6G AI native networks. In general, GTC 2025 conference summary in one sentence: and the physical range that Nvidia chips can solve is expanding, which means that AI ...