Covers the data center and edge server markets, including vendor market share; enterprise, cloud, and communications service provider investment; and semiconductor-level transitions.

Omdia view

The ascendancy of GPU servers for AI: a new epoch in demand

Highlights of the quarter:

  • The server market recorded a revenue growth of 22% QoQ and a 2% decline YoY in 2Q23.
  • The server unit shipments were down to 2.6 million, declining 7% QoQ and 25% YoY.
  • In 2Q23, NVIDIA shipped more than 300,000 H100 GPUs for use in OEM and ODM servers. Each of these GPUs costs about $21,000, and the DGX/HGX systems they go into cost more than $240,000, pushing the server ASPs up 30% QoQ and YoY.
  • Few vendors benefitted significantly from the top 10 cloud providers’ appetite for GPU servers. In the ODM direct market segment, ZT Systems and QCT shipped the most GPU servers. Among OEMs, Supermicro faced significant tailwinds from GPU server shipments.
  • The mainstream server shipments (non-GPU dual/single socket) were down significantly as cloud service providers prioritized their investments in building AI compute capacity. For example, Meta used to deploy more than 250,000 Yosemite V3 servers per quarter until 4Q22, but it looks like the company has considerably slowed down its deployment of Yosemite servers and is probably moving away from that design altogether.
  • Meta would replace the Yosemite V3 designs with an AMD Bergamo CPU-based design, which offers higher core density. A 12-node Yosemite V3 server chassis could be replaced with a single Bergamo-based server, driving consolidation in server units.

A subscription is required to view this content.

Already subscribed? Continue Continue