CITIC Securities: NVIDIA's new product launch, the acceleration of optical module evolution is just in time

According to the report released by CITIC Securities on the Zhitong Finance APP, at the COMPUTEX 2024 conference, we once again saw the important position of network equipment in the development of AI. We also witnessed the huge opportunities brought by larger cluster sizes and faster product iteration speeds for optical modules and other network equipment. CITIC Securities believes that with the rapid development of AI and the accelerated introduction of high-end network equipment, top manufacturers with strong R&D capabilities will have higher competitive barriers. The industry landscape will continue to optimize, and high-end products will bring higher profits to leading manufacturers. It is recommended to focus on leading companies that are strategically positioned in AI high-speed optical modules, coherent optical technology, AEC, and other directions.

NVIDIA CEO delivers a speech and launches the new GPU platform Rubin.

NVIDIA CEO Jensen Huang delivered a speech at 7 p.m. on June 2nd at the COMPUTEX 2024 conference. During the event, Huang introduced the application of AI basic technology in industries such as robotics and climate prediction, and further introduced the Blackwell platform and released the next-generation GPU platform Rubin. Additionally, NVIDIA showcased the development roadmap for the next generation of GPUs, switches, and network cards at a pace of one generation per year: it is expected to launch the Blackwell/Blackwell Ultra/Rubin GPUs and corresponding network supporting devices in 2024/2025/2026.

NVIDIA once again emphasizes the importance of AI networks, with the possibility of million-card clusters approaching.

Huang reiterated the importance of network equipment at the event: for a $5 billion AI data center, if the network utilization rate drops by 40%, resulting in a 20% increase in training time, it would waste $1 billion in costs, showing the significant impact of network performance on the overall efficiency of AI data centers. NVIDIA also introduced its Ethernet platform Spectrum-X and provided a product roadmap at a pace of one generation per year:

1. 2024: Launching the Spectrum-X800 switch (exchange capacity 51.2T) and the BF3 400G SuperNIC network card, mainly used for a million-card cluster;

2. 2025: Launching the Spectrum UItra-X800 switch (exchange capacity 51.2T) and the CX8-800G SuperNIC network card, mainly used for a hundred-thousand-card cluster;

3. 2026: Launching the Spectrum-X1600 switch (exchange capacity 102.4T) and the CX9-1600 SuperNIC network card, mainly used for a million-card cluster.

In NVIDIA's outlook, we see the trend of GPU computing power clusters expanding towards million-card levels, which is expected to drive the continuous growth in demand for network equipment The next generation optical modules are expected to land in 2026, driving continuous industry demand growth.

According to NVIDIA's Ethernet platform Spectrum-X development blueprint, the Spectrum-X1600 switch and the CX9-1600 SuperNIC network card are set to be launched in 2026. The switch capacity and network card speed will both double, potentially driving the upgrade of optical module demand from 1.6T to 3.2T. This roadmap once again proves that AI development is accelerating the iteration progress of optical modules, shortening the cycle from the previous 3-4 years to 1-2 years (which is consistent with Marvell's statement at the Marvell AI Day conference that "the cycle of AI interconnection rate upgrades has been halved from 4 years to 2 years"). Previously, there were concerns in the market about the sustainability of optical module industry demand in 2026, but NVIDIA's conference clearly outlined the future development blueprint of network equipment. The launch of Spectrum-X1600 and the CX9-1600 SuperNIC network card are expected to directly drive the demand for 3.2T optical modules, thereby fueling continuous industry demand growth