DIGITIMES Research believes tight CoWoS capacity, which is needed for packaging HBM, will result in the supply of high-end AI servers falling short of demand by more than 35% in 2023, but high-end AI server shipments worldwide are still expected to grow fivefold from a year ago, according to the latest server report covering the AI server market.
In response to explosive generative AI and large language model (LLM) demand, major cloud service providers and leading server brands are stepping up 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).
DIGITIMES Research projects that over 80% of global high-end AI server shipments will be delivered to the global top-5 cloud service providers and server brands right out of the factories in 2023. Among these firms, Microsoft, which invested in OpenAI's ChatGPT and devoted heavy resources to generative AI cloud services, will obtain the largest portion of the volumes with all the shipments procured directly from its ODM partners, the report's figures show.
The second largest portion of shipments will go to Google, which has long been committed to generative AI R&D and secures key technologies, and is mainly pushing into the sector with high-end AI servers that are powered with its in-house developed Tensor processing units (TPUs).
Meta, which is the most keen on deploying open-source LLMs, will have the third largest portion, followed by Amazon, which was late in entering the AI field but has been keen on forming partnerships with third-party companies to catch up, at the fourth. US-based server brands including Supermicro, Dell, Nvidia, and HPE are trailing closely behind the four major cloud service providers
Breaking down 2023 high-end AI server shipments by the type of accelerators they equip, a major portion is models installed with Nvidia's HGX series. The shipments of high-end AI servers with the H100 HGX will slightly exceed the shipments of those with the A100 HGX. The shipments of those with TPU, widely used by Google, will account for over 10%. Among the rest, a big share will feature Nvidia's PCIe-based 100 series GPU and a minor portion Nvidia's GH200 as well as AMD's and Intel's high-end GPUs.
Breaking down 2023 high-end AI server shipments by ODMs/OEMs, Inventec will have the largest share in the L6 manufacturing sector and US-based ZT Systems in the L10-L12 manufacturing one. Inventec, receiving orders from multiple major cloud service providers and server brands, will account for over half of the L6 server manufacturing sector in 2023.
ZT Systems, having engaged in long-term partnerships with major cloud service providers and high-end server technology suppliers, secures advantages in integrating large GPU clusters in server systems. It is expected to account for a major share of the L10-L12 server manufacturing sector. Among Taiwan-based ODMs, Quanta makes the most aggressive efforts toward the L10-L12 manufacturing.