NVIDIA "GTC 2025" Hardware Forecast and Investment Trend Analysis

Wallstreetcn
2025.03.18 04:07
portai
I'm PortAI, I can summarize articles.

Guo Mingqi from Tianfeng Securities stated that investors are concerned whether the conference can address three major doubts: the effectiveness of the AI expansion law, the mass production issues of GB200 NVL72, and geopolitical impacts. Among them, the hardware update of the B300 series chips is crucial. In addition, the conference may announce multiple data center server solutions and network solutions, but the AI PC solution is expected to be revealed at Computex

NVIDIA's upcoming GTC 2025 conference will directly impact the direction of the AI chip market.

Guo Minghao from Tianfeng Securities recently expressed the view that investors should focus on the sustainability of the AI scaling law and the mass production status of new AI servers. Among them, the B300 series chips will be the focus of hardware updates at this conference, with HBM memory significantly upgraded from 192GB to 288GB, and computing performance improved by 50% compared to the B200. For investors, although the conference may serve as a catalyst for short-term stock price increases, sustainability will depend on whether the conference can effectively address the aforementioned concerns.

Additionally, the conference may announce multiple data center server solutions and networking solutions, but the AI PC solutions are expected to be unveiled at Computex.

Three Major Concerns for Investors

Guo Minghao stated that the current investment trend in AI servers dominated by NVIDIA faces three major concerns:

First, as low-tier AI devices begin to "rise," the market questions the continued effectiveness of the AI server scaling law. If NVIDIA can present new insights on this at the GTC 2025 conference, it will help alleviate market concerns.

Second, whether it can ease market worries about the poor mass production of the GB200 NVL72. The poor mass production status of the GB200 NVL72 has become a consensus in the market. If NVIDIA can showcase data center construction cases using the GB200 NVL72 and focus on the conversion benefits between the B200 and B300, increasing the transparency of B300-related mass production, it will help the market start to pay attention to the investment theme of B300. The B300 is expected to enter trial production in the second quarter of 2025 and mass production in the third quarter.

The B300 offers both Dual-die (CoWoS-L) and Single-die (CoWoS-S) options, with the biggest highlight being the significant upgrade of HBM from 192GB to 288GB, and a 50% increase in computing performance compared to the B200 (FP4).

Finally, the impact of geopolitical factors on NVIDIA's business. Given the sensitivity of geopolitical issues, it is expected that NVIDIA will not discuss related topics at GTC 2025.

It is worth noting that while end-user AI is also a long-term key trend for NVIDIA, Guo Minghao expects that GTC 2025 will focus on AI servers, and the anticipated AI PC solutions N1X and N1 are more likely to be announced at this year's Computex.

Data Center AI Server Solutions: Product Lineup Exposed

Additionally, Guo Minghao listed various data center server solutions and server models currently being developed by NVIDIA, including:

  • B300 chips: GB300 NVL72, HGX B300 NVL16 (air-cooled), HGX B300 NVL16 (liquid-cooled), and the lower-tier B300 NVL
  • B200 chip: HGX B200 NVL8, GB200 NVL4.
  • Workstation equipped with RTX PRO 6000 Blackwell Server Edition chip.
  • Next-generation AI server equipped with VR (Vera Rubin) chip, with new 144/288 options.

At the GTC 2025 conference, the following server models are most likely to be released:

  • GB300 NVL72: Replacing GB200 NVL72. Due to similar cabinet size and power specifications, data centers can seamlessly upgrade to GB300 NVL72. The PS (Pre-build sample) schedule is June 2025.
  • HGX B300 NVL16: Replacing HGX B200 NVL8 and HGX H200 NVL8. Since it uses a Single-die, the actual total GPU count remains unchanged. The PS schedule for air-cooled and water-cooled versions is June and September 2025, respectively.
  • NVL288/144: Designed for the new VR architecture, but since Vera and Rubin have not yet reached mass production, if announced, the cabinet sample will temporarily use GB chips. It mainly promotes NVIDIA's Scale-up design advantages. Not much detail about Vera and Rubin may be disclosed, as the mass production timeline is still far off (expected to have small-scale production in 2–3Q26).
  • Workstation equipped with RTX PRO 6000 Blackwell Server Edition chip, used for AI and visualization applications, equipped with GDDR7 96GB (1.6 TB/s), TDP 400–600W. Expected to reach mass production in 2Q-3Q25.

NVIDIA will also launch a series of data center networking solutions, including:

  • Quantum-3
  • Quantum-X800
  • Spectrum-5
  • ConnectX-8 (CX8)

The speed of CX8 is twice that of CX7, integrating SuperNIC and PCIe Switch (supporting PCIe Gen6), thus improving power consumption by 30%. It supports GB300 NVL72 and the latest Quantum platform.

In addition to the core content mentioned above, the GTC 2025 conference will also cover other key AI applications and solutions, including robotics, autonomous driving, quantum computing, and more

Source: Wallstreetcn The copyright of this article belongs to the original author/organization.
The views expressed herein are solely those of the author and do not reflect the stance of the platform. The content is intended for investment reference purposes only and shall not be considered as investment advice. Please contact us if you have any questions or suggestions regarding the content services provided by the platform.