[alexbsr]

https://www.patreon.com/posts/nvidia-unveils-124754224

On March 18 (U.S. time), NVIDIA hosted its annual GTC (GPU Technology Conference) in San Jose, California. As one of the most highly anticipated tech events of the year, GTC 2025 attracted approximately 25,000 in-person attendees, with an additional 300,000 viewers tuning in online.

NVIDIA CEO Jensen Huang opened his keynote speech by emphasizing the rapid expansion of GTC due to the explosive growth of AI. "Last year, they called GTC the ‘Woodstock of AI.’ This year, we’ve moved into a stadium—GTC has become the ‘Super Bowl of AI,’" Huang remarked.

At this year’s event, NVIDIA introduced a series of groundbreaking new products, including the Blackwell Ultra GPU, silicon photonics switches, and advanced robotics models. A key message from Huang’s speech was that while AI model training demand is slowing, the AI inference era is accelerating, driven by innovations like DeepSeek’s advancements in model inference.

Despite the exciting announcements, NVIDIA’s stock closed down more than 3.4%, settling at $115.43 per share, with an additional 0.56% decline in after-hours trading.

Blackwell Ultra GPU: The "Computing Nuke" for AI Inference

The highlight of GTC 2025 was NVIDIA’s announcement of the Blackwell Ultra GPU, the next-generation AI GPU for data centers. Industry speculation had suggested that NVIDIA planned to rename the Blackwell Ultra to B300 by late 2024, but the company ultimately retained the original branding.

Compared to its predecessor, the B200, the Blackwell Ultra GPU delivers a 50% performance boost, reaching approximately 15 PetaFLOPS (using FP4 precision). It also features the most advanced HBM3E memory, with capacity increasing from 192GB to 288GB.

For enterprise and cloud computing clients, NVIDIA introduced two system integration solutions based on Blackwell Ultra:

Blackwell Ultra NVL72: A rack-mounted solution connecting 72 Blackwell B300 GPUs with 36 NVIDIA Grace CPUs (ARM-based) in a single data center rack.

NVIDIA HGX Ultra NVL16: A server system linking eight Blackwell Ultra GPUs via NVLink for high-speed interconnectivity.

Unlike AI GPUs such as the A100 and H100, which were primarily designed for AI model pre-training, the Blackwell Ultra is explicitly positioned for AI inference while maintaining strong training capabilities. The GB300 NVL72 and HGX B300 NVL16 systems have been optimized for large-scale AI inference workloads, with the HGX B300 NVL16 boasting an 11x improvement in inference speed compared to the previous-generation Hopper architecture.

NVIDIA also addressed concerns that AI model training demand might decline following DeepSeek’s success in optimizing inference efficiency. Huang reassured that AI inference still requires substantial GPU resources and high-performance networking. As the AI industry shifts from the "Scaling Law" of pre-training towards inference-driven demand, "DeepSeek’s innovation actually proves that the market needs even more AI chips."

Huang further confirmed that the Blackwell series has entered full production, with record-breaking output and customer demand. "AI has reached an inflection point where inference systems and intelligent agents require exponentially more computing power."

Looking Ahead: Rubin and Feynman Architectures

Following the Blackwell Ultra announcement, Huang previewed NVIDIA’s upcoming Rubin architecture, set to launch in 2026:

Rubin GPU: Capable of 50 PetaFLOPS (FP4), 3.3x the power of Blackwell Ultra.

Rubin Ultra GPU: Reaching 100 PetaFLOPS (FP4) with HBM4 and HBM4E memory.

Vera Rubin NVL144: A 144-GPU system launching in late 2026.

Rubin Ultra NVL576: A massive 576-GPU system debuting in late 2027.

Additionally, Huang revealed that NVIDIA’s Feynman architecture—named after physicist Richard Feynman—will arrive in 2028.

The Rise of AI Agents and Robotics

This year, NVIDIA devoted significant time to explaining the concept of Agentic AI, an evolution from Generative AI towards intelligent agents capable of complex reasoning and autonomous decision-making.

Huang outlined AI’s development in three stages:

Generative AI – Focused on text, image, and content generation.

Agentic AI – AI systems that understand tasks, plan, and execute multi-step operations autonomously.

Physical AI – Embodied AI for robotics and real-world applications.

Unlike Generative AI, which mainly relies on large language models, Agentic AI requires substantial inference capabilities. As AI agents generate and process vast amounts of data, computing demand will scale even further. To illustrate, Huang demonstrated that the Blackwell Ultra NVL72 cluster can generate DeepSeek-R1 671B responses in just 10 seconds—compared to 90 seconds on the previous H100 architecture.

NVIDIA also introduced Dynamo, an AI inference software that optimizes GPU usage across thousands of units. By decoupling model processing and generation stages across different GPUs, Dynamo ensures maximum efficiency for large-scale inference tasks.

Huang estimated that AI inference will demand computing resources 100 times greater than current expectations, predicting that global data center investments will exceed $1 trillion by 2028.

Silicon Photonics, Robotics, and Quantum Computing

Beyond GPUs, NVIDIA announced advancements in networking, robotics, and quantum computing:

Silicon Photonics Switches: NVIDIA unveiled the Spectrum-X (Ethernet-based) and Quantum-X (InfiniBand-based) silicon photonics switches, integrating optical communication directly into switch hardware for the first time.

Robotics & AI Agents: NVIDIA’s Cosmos, GROOT N1 humanoid model, and Omniverse simulation platform represent major steps toward intelligent robotics. The GROOT N1 model, inspired by human cognition, has now been open-sourced.

Quantum Computing Research: NVIDIA announced the NVIDIA Accelerated Quantum Research Center (NVAQC) in Boston, aimed at advancing quantum computing architectures and algorithms.

While competitors like Google and Microsoft have made breakthroughs in quantum error correction, Huang remains cautious, stating that "practical quantum computing is still 20 years away." However, he emphasized that quantum computing will complement, rather than replace, traditional AI supercomputing.

Conclusion

GTC 2025 underscored NVIDIA’s vision for the AI future—shifting from model training dominance towards an era driven by AI inference, intelligent agents, and robotics. With Blackwell Ultra GPUs in production and Rubin architecture on the horizon, NVIDIA is positioning itself at the forefront of next-generation AI computing. As demand for AI inference scales exponentially, Huang believes that the next computing revolution has only just begun.

PS: When upgrading your GPU, don’t let your old one sit idle. Sell your old GPU for a great price to professional buyers like BuySellRam.com, a leading global GPU supplier. Get a free quote today! If you want to learn more about where and how to sell GPU, click the link.