Michelle ThompsonAccording to the most recent server report covering the AI server market, DIGITIMES Research believes that limited CoWoS capacity, which is required for packaging HBM, will cause the supply of high-end AI servers to fall more than 35% short of demand in 2023. Despite this, high-end AI server shipments worldwide are still anticipated to increase fivefold from a year ago.
Major cloud service providers and top server manufacturers are stepping up their efforts toward AI servers, with a focus on ramping up their procurement of high-end AI servers featuring accelerators that are integrated with high bandwidth memory (HBM), in response to the explosive growth in generative AI and large language model (LLM) demand.
In 2023, according to a study by DIGITIMES Research, the top five cloud service providers and server manufacturers would get over 80% of the world’s high-end AI server shipments directly from the factories. The report’s numbers indicate that among these companies, Microsoft will acquire the highest share of volumes, with all shipments obtained directly through its ODM partners. Microsoft invested in OpenAI’s ChatGPT and heavily invested in generative AI cloud services.
The second-largest share of shipments will go to Google, which has long been devoted to generative AI research and development and secures crucial technology. Google is primarily pushing into the market with high-end AI servers that are driven by its in-house built Tensor processing units (TPUs).
The third-largest amount will go to Meta, which is most interested in deploying open-source LLMs, and the fourth-largest portion will go to Amazon, which entered the AI area later but has been eager to create collaborations with other businesses to catch up. US-based server manufacturers Supermicro, Dell, Nvidia, and HPE are well behind the four main cloud service providers in terms of market share.
A significant part of models equipped with Nvidia’s HGX series accelerators make up the high-end AI server shipments in 2023 when broken down by the type of accelerators they contain. Shipments of high-end AI servers equipped with the H100 HGX will somewhat outpace those of servers equipped with the A100 HGX. Over 10% of all shipments will be made up of those featuring TPU, which Google frequently uses. Nvidia’s PCIe-based 100 series GPU will make up a significant chunk of the remaining GPUs, while the GH200 and high-end GPUs from AMD and Intel will make up the remainder.
Inventec will hold the biggest market share in the L6 manufacturing sector for high-end AI server shipments in 2023, while US-based ZT Systems will hold the top spot in the L10–L12 manufacturing sector. In 2023, Inventec will manufacture more than half of the L6 server market thanks to orders it has received from numerous significant cloud service providers and server manufacturers.
ZT Systems has advantages when integrating big GPU clusters in server systems thanks to its long-term agreements with leading cloud service providers and high-end server technology manufacturers. It is anticipated to hold a sizable portion of the L10-L12 server production market. Quanta is the Taiwan-based ODM that pursues L10-L12 manufacturing with the most vigor.
