Omdia: Nvidia Q3 sold nearly 500000 A100 and H100 AI GPU,Meta and Microsoft are the biggest buyers

CTOnews.com, Nov. 28 (Xinhua) in the third quarter of this fiscal year, Nvidia made $14.5 billion on data center hardware thanks to strong demand for AI GPU (CTOnews.com Note: currently about 103.82 billion yuan).

Nvidia Q3 sold nearly 500000 A100 and H100 GPU, and delivery cycles for H100-based servers have been extended to 36 to 52 weeks, according to research firm Omdia.

Omdia believes that Meta and Microsoft are the biggest buyers of the Nvidia H100 GPU. They all purchased as many as 150000 H100 GPU, far more than Google, Amazon, Oracle and Tencent (50,000).

It is worth noting that most of the servers GPU are provided to very large-scale cloud service providers. Omdia claims that server OEMs (Dell, Lenovo, HPE) have yet to get enough AI and HPC GPU to complete server orders.

The analyst firm believes that Nvidia H100 and A100 calculated GPU sales will also exceed 500000 in the fourth quarter. At the same time, the demand for H100 and A100 is very strong, and the delivery cycle of GPU servers is as long as 52 weeks. Meanwhile, Omdia says server revenue grew between + 6 per cent and + 8 per cent year-on-year in 2023.

Almost all companies that buy large quantities of Nvidia H100 GPU are developing custom chips for AI, HPC and video workloads, the report said. As a result, as companies turn to their own chips, their purchases of Nvidia hardware are likely to decline over time.

The report looks forward to 2027, and the value of the server market is expected to reach a staggering $195.6 billion (currently about 1.4 trillion yuan). Companies are starting to shift to customizing servers for specific applications with different coprocessor arrays, the report said. For example, Amazon's server dedicated to AI reasoning has 16 Inferentia2 coprocessors, and Google's video transcoding server is equipped with 20 custom VCU. Meta followed suit, equipping the server with 12 custom processors for video processing.