The increasing popularity of ChatBOTs, generative AI, and other applications is driving major CSPs like Microsoft, Google, and AWS to increase their expenditures in AI, which in turn is driving up demand for AI servers, according to TrendForce. Shipments of AI servers (including those with GPUs, FPGAs, and ASICs) are expected to reach over 1.2 million units in 2023, representing a YoY growth of 37.7% and accounting for approximately 9% of all server shipments.
In 2024, the shipments of AI servers is projected to increase by 38%, with AI accounting for 12% of the growth.
Apart from the GPU solutions offered by NVIDIA and AMD, prominent CSPs have been inclined to create their own ASIC processors. Since 2H23, Google, for example, has been speeding up the integration of its unique TPU into AI servers, with an annual growth rate that surpasses 70%.
In 2024, AWS plans to increase the number of custom ASICs it uses, and shipping volume is anticipated to double. Others, such as Meta and Microsoft, want to increase their in-house ASIC solutions as well, which might limit the future expansion of GPUs.
In conclusion, strong investments by CSPs are projected to be the primary driver of the growth in demand for AI servers in 2023–2024. After 2024, it’s anticipated that more application-focused businesses will focus on creating specialized AI models and software servers, which will encourage the expansion of edge AI servers outfitted with FPGAs or mid-to-low-end GPUs (like the L40S series).
Shipments of edge AI servers are predicted to expand at an average yearly pace of more than 20% between 2023 and 2026.
HBM3e Boosts HBM Revenue 172%
As more AI servers are built, there is an increasing need for AI accelerator chips. Among the DRAM products that are essential for these accelerator chips is HBM.
In terms of specifications, this year’s mass manufacturing of NVIDIA’s H100/H800 and AMD’s MI300 series has raised the share of demand designated for HBM3, in addition to the existing standard HBM2e. The three main memory providers anticipate releasing HBM3e in 2024, increasing throughput to 8 Gbps, and guaranteeing even greater performance for AI accelerator processors in 2024–2025.
Beyond top server GPU producers like NVIDIA and AMD, CSPs are also stepping up the development of proprietary AI chips in the market for accelerator chips; these chips share HBM integration, according to TrendForce. The need for HBM is predicted to rise dramatically as training models and applications get more complicated. With an average unit price many times higher than other DRAM products and an annual growth rate of 172%, it is anticipated that HBM will contribute significantly to memory vendors’ sales by 2024.
