Nvidia GTC 2025 hardware forecast and investment trend analysis

Key Observations from an Investment Perspective on GTC 2025:

1. Currently, there are three major concerns regarding the investment trend in Nvidia's AI servers: the continued validity of the Scaling Law, the mass production status of new AI servers, and geopolitical factors.

2. With the growth of end/lower-tier AI devices, if Nvidia can provide new insights into the validity of the Scaling Law for AI servers, it would help alleviate market concerns.

3. The poor mass production status of the GB200 NVL72 is already a market consensus. If Nvidia can present data center case studies of GB200 NVL72 deployments, focus on the transition benefits between B200 and B300, and improve visibility into B300 mass production, it would help the market react to the B300 investment theme.

4. Nvidia is not expected to address geopolitical issues.

5. Edge AI is also a long-term key trend for Nvidia, but GTC 2025 is expected to focus on AI servers. The market-anticipated AI PC solutions N1X and N1 are more likely to be announced at this year's Computex.

6. As AI-related stocks have recently corrected, GTC should provide a short-term catalyst for stock price increases. However, whether the momentum can significantly continue depends on whether the GTC conference content can address the aforementioned investor concerns.

New AI Server Chips and System Solutions Are Key to GTC Hardware Updates:

1. The new AI chip B300 is the key highlight of the launch, including Dual-die (CoWoS-L) and Single-die (CoWoS-S), with the biggest upgrade being HBM increasing significantly from 192GB to 288GB, delivering a 50% performance improvement over B200 (FP4).

2. B300 is expected to enter trial production in Q2 2025 and mass production in Q3 2025.

3. Provides reference designs for Scale-up and Scaling-out solutions that offer stronger computing power and lower average token costs.

Data Center AI Server Solutions:

- Servers currently under development include:

1. B300 chip: GB300 NVL72, HGX B300 NVL16 (air-cooled), HGX B300 NVL16 (liquid-cooled), and the lower-tier B300 NVL.

2. B200 chip: HGX B200 NVL8, GB200 NVL4.

3. Workstations equipped with the RTX PRO 6000 Blackwell Server Edition chip.

4. Next-gen AI servers equipped with VR (Vera Rubin) chips, adding 144/288 options.

- Server models likely to be announced at GTC 2025:

1. GB300 NVL72: Replaces GB200 NVL72. Due to similar rack size and power specifications, data centers can seamlessly upgrade to GB300 NVL72. Pre-build sample (PS) timeline is June 2025.

2. HGX B300 NVL16: Replaces HGX B200 NVL8 and HGX H200 NVL8. Since it uses Single-die, the total number of GPUs remains unchanged. PS timelines for air-cooled and liquid-cooled versions are June 2025 and September 2025, respectively.

3. NVL288/144: Designed for the new VR architecture, but since Vera and Rubin are not yet in mass production, if announced, the rack samples will temporarily use GB chips. The main focus is on promoting Nvidia's Scale-up design advantages. Few details about Vera and Rubin are expected to be disclosed, as mass production is still far off (expected in Q2-Q3 2026).

4. Workstations equipped with the RTX PRO 6000 Blackwell Server Edition chip for AI and visualization applications, featuring GDDR7 96GB (1.6 TB/s) and TDP of 400-600W. Expected to enter mass production in Q2-Q3 2025.

Data Center Networking Solutions:

1. Includes Quantum-3, Quantum-X800, Spectrum-5, and ConnectX-8 (CX8).

2. CX8 doubles the speed of CX7, integrates SuperNIC and PCIE Switch (supporting PCIe Gen6), improving power efficiency by 30%. Supports GB300 NVL72 and the latest Quantum platform.