Nvidia GTC Conference Preview: Blockbuster B300 on the Way, Ultimate Goal to AGI

NVIDIA is set to hold the GTC 2025 conference in San Jose, California, from March 17 to 21, where it will unveil the latest breakthroughs in AI chips, deployment, and applications. With DeepSeek making waves recently, the focus will be on how the GPU maker navigates AI to advance further toward Artificial General Intelligence (AGI).

This year's GTC will feature over 1,000 sessions covering fields such as LLM model training and deployment, AI applications, AI-driven pharmaceutical research, content creation, AR/VR, conversational AI, cybersecurity, edge computing, robotics, physics, and more. The event will showcase cutting-edge AI technologies across multiple domains.

The most anticipated event will be nvidia CEO Jensen Huang's keynote on March 18, where he will discuss what's next in agentic AI, robotics, and accelerated computing. There is strong speculation that he will unveil the new B300 chip, the GB300 AI server, and the next-generation Rubin computing architecture, laying out NVIDIA's product roadmap for the next two to three years.

This product launch comes at a time of turbulence in the U.S. stock market, with concerns over Trump's tariff policies and fears of AI compute oversupply weighing on investor sentiment. NVIDIA's stock has already dropped nearly 14% this year, retreating 25% from its all-time high in January. Historically, NVIDIA's stock has gained an average of 4.8% during the past five GTC events, and investors are eager to see whether this trend will continue or if it will help reverse the company's recent downturn.

B300: 50% Computing Boost and a Full Transition to Liquid Cooling

At last year's GTC, NVIDIA introduced the Blackwell B200 chip and the GB200 server. While initial overheating issues led to delivery delays and some order losses, the company still managed to ship $11 billion worth of Blackwell chips in the fourth quarter of 2024.

Now, Huang is set to unveil the much more powerful B300, also known as Blackwell Ultra, which is expected to begin shipping as soon as May, with broader availability in the second half of 2025.

The B300 will continue using the Blackwell architecture and feature eight 12-Hi HBM3E stacks, increasing total memory to 288GB, a 50% jump from the B200. FP4 floating-point computing performance is also expected to improve by 50%.

To support these GPUs, NVIDIA will introduce the Mellanox Spectrum Ultra X800 Ethernet switch, which features a 512-radix architecture and supports up to 512 ports. This upgrade will significantly enhance AI and high-performance computing (HPC) systems.

However, higher performance will come with increased power consumption. The B300 chip inside the GB300 server is projected to reach a TDP of 1,400W, compared to 1,200W for the previous generation. To address this challenge, NVIDIA is expected to fully transition to liquid cooling, equipping each chip with its own dedicated cooling plate. 

Next-Generation Rubin Architecture: A Leap Toward AGI

In addition to the B300, Huang is expected to unveil NVIDIA's next-generation Rubin architecture, which is scheduled for release in 2026. This will mark another major step toward AGI.

Rubin will feature eight stacks of HBM4E memory, offering up to 288GB of capacity. It will integrate a Vera CPU, NVLink 6 switches with a bandwidth of 3,600 GB/s, CX9 network adapters with speeds of 1,600 Gb/s, and X1600 switches. Huang has described Rubin as a groundbreaking leap in performance and capabilities.

Huang may also provide a glimpse into post-Rubin products, such as Rubin Ultra, which is expected to launch in 2027. Rubin Ultra is rumored to feature 12 stacks of HBM4E memory, increasing total capacity to 576GB. HBM4E technology is set to deliver unprecedented bandwidth and memory capacity, which will be critical for handling increasingly complex AI models and computational tasks.

The introduction of Rubin raises concerns about AI compute oversupply. DeepSeek recently built its R1 inference model at a fraction of OpenAI's cost, demonstrating that AI inference is evolving beyond sheer compute expansion.

Huang has emphasized that AI is shifting from generative AI to inference-based and logical reasoning AI. Unlike one-shot inference, long-term reasoning AI may require 100 times more compute per task. This shift reinforces NVIDIA's argument for sustained demand for AI hardware.

Agentic AI and Robotics: The Next Frontier

Beyond accelerated computing, Huang's keynote will also highlight the potential of agentic AI and robotics.

Agentic AI, or AI agents, is an emerging trend where AI systems are designed to autonomously handle complex problem-solving and automate repetitive tasks. Many major tech firms are heavily investing in this field, and Huang is expected to showcase NVIDIA's advancements in simplifying toolchains and improving efficiency.

Discussions around robotics have gained momentum recently, but high costs and commercialization challenges remain significant hurdles. At CES 2024, NVIDIA introduced the Isaac GR00T Blueprint, a platform that enables developers to generate large-scale synthetic datasets with minimal human demonstration, facilitating imitation learning for humanoid robots.

At present, NVIDIA's focus remains on AI training rather than mass-producing robots. However, Huang's decision to include robotics in his keynote suggests that there may be new developments in the field.

NVIDIA's First Quantum Day: How Soon Will "Useful" Quantum Computing Arrive?

Another key highlight of GTC 2025 will be NVIDIA's first-ever Quantum Day on March 20, featuring leading figures from the quantum computing industry.

Huang previously predicted that a very useful quantum computer might still be 15 years away, a statement that sparked debate among industry experts. The Quantum Forum will explore how soon quantum computing will achieve real-world applications and its implications for human progress. This discussion is expected to attract significant interest.

With the B300, Rubin, and agentic AI in focus, NVIDIA is gearing up for another monumental GTC. However, given the current concerns about an oversupply of AI hardware and market uncertainty, investors will be closely watching whether this event can help reignite NVIDIA's stock momentum.