AI hardware industry | Expected to generate excess returns in the future

ETFWorldSavior wrote a column · Jul 19, 2023 01:27

Core points:

1. The current AI investment is mainly concentrated in semiconductor companies, involving the provision of artificial intelligence rack solutions for high-intensity computing workloads. Artificial intelligence opportunities ignored by investors, including AI servers, optics, and cable connectors, will generate excess returns in the future.

2. The dense configuration of artificial intelligence servers will promote the demand for expanding network bandwidth, which in turn will drive the development of optics companies such as Coherent, Fabrinet, etc.

3. The development potential of active electrical cables (AEC) is huge, and it has not yet been fully valued. Credo is in the leading position in the early stage, accounting for more than 95% of the share, but it is expected that APH and TEL will also benefit.

Based on the following analysis, we believe that the AI hardware industry has a large profit margin and potential opportunities in the future.

1) The AI server expands computing power and significantly improves performance

Servers are the powerhouses of the data center, responsible for all the processing. General-purpose servers are usually 1U to 3U in size and consist of processors (usually CPUs), memory, storage, network cards, power supplies, cooling, and more. AI servers, on the other hand, can range from 1U for inference applications to 6U for AI training applications, and consist of numerous processors, optimized for specialized processors with specific workloads or optimized for computing, and can reach Higher computing requirements.

Figure 1: Comparison of server shipments

AI hardware industry | Expected to generate excess returns in the future

Figure 2: AI server ecosystem

2) AI deployment drives the increase in bandwidth demand, promoting the development of optics

The dense configuration of AI servers will drive the need to expand network bandwidth. Specifically, by 2026, the switching bandwidth of artificial intelligence applications is expected to grow at a compound annual growth rate of +100% to more than 10% of the total bandwidth, far exceeding the growth of the total bandwidth of the market and driving the vast majority of traditional data centers. Specific scenarios for most bandwidth requirements, such as DCI, L3 and L2. Therefore, cloud service providers will seek to deploy the fastest available data communication transceivers on the market, thereby driving faster network upgrade cycles, and the commercialization of 800G and 1.6T data communication transceivers will be realized faster than expected, driving optical computing related business development.

Figure 3: Bandwidth shipments under different application scenarios

Figure 4: Proportion of bandwidth application scenarios

3) Traditional cables are limited, and AEC will become a breakthrough point in the future

Datacenter racks are interconnected by cables, currently low-cost direct-attach copper cables (DACs) with no additional power requirements. However, as network densities increase, DAC losses lead to limited length, and large gauge sizes pose routing challenges. In contrast, the active electrical cables (AEC) can break through the above limitations. Its conductors are thinner, its range of action is longer, and its reliability is stronger. At the same time, it has lower power consumption and cost. It is expected that the incremental part of the AEC market is expected to grow at a compound annual growth rate of more than 100%.

Figure 5: Direct Attach Copper (DAC) Cable Performance