Nvidia now plays kingmaker in the server court, says Omdia
AI's appetite for compute means general refreshes stalled as customer used budgets elsewhere
The server market is bouncing back in value terms, as customers demand beefier systems to train or run AI models – and the success of server makers is now largely dependent on getting GPUs from Nvidia.
Analyst Omdia says in its latest Cloud and Datacenter Market Snapshot that global server revenue grew to $31 billion in calendar Q4, up 12.7 percent year-on-year and up 21.5 percent on the prior quarter.
The trend of customers buying more richly configured servers to meet the requirements of AI processing was noted by Omdia last year, and is gaining momentum driven by the surging interest in generative AI models sparked by ChatGPT.
The upshot is that the volume of servers shipped in 2023 actually dropped below the forecast as buyers funneled their investment into AI heavyweight hardware. Omdia says that preliminary data indicates the number of servers shipped in the quarter was between 2.8 and 2.9 million units, or half a million fewer than estimated, and the lowest since 2017.
"This reaffirms our thesis that end users are prioritising investment in highly configured server clusters for AI to the detriment of other projects, including delaying the refresh of older server fleets," the report states.
Unsurprisingly, the big winner here is Nvidia, whose chips such as the current H100 and A100 are in high demand for accelerating the training of AI models in particular. Nvidia GPUs now account for about 44 percent of the server bill of materials (BOM) on average, Omdia estimates.
With each H100 carrying an eye-popping price tag of approximately $21,000, it's little wonder that the price of servers has increased, and Omdia put Nvidia's revenue across its datacenter GPU portfolio at $13.7 billion for Q4 of 2023.
Nvidia's importance to the server market is now so great that Omdia says the GPU builder is effectively a kingmaker, and a close working relationship with the Santa Clara-based biz will be key for any server maker wanting to grow its share of the market in 2024.
- HPE blames GPU shortage for contributing to unexpected sales slide
- Nvidia lures senior autonomous car developer from China's Baidu
- Nvidia talks up local AI with RTX 500, 1000 Ada mobile GPUs
- Nvidia revenue grows 265 percent with more to come as new GPUs and Ethernet near
To support this, Omdia highlights hyperscale cloud supplier Supermicro, which grew its share of server market revenue from five to 10 percent over the course of 2023 based on AI server sales, overtaking HPE in Q4.
HP was itself hit by GPU shortages that dented revenue growth, as outlined in financial results disclosed last week. Just days ago, Dell also talked of a $2.9 billion AI server backlog, which served to drive its share price to record highs not seen since the company went public again in 2018.
Another factor in falling general server shipment might be companies extending the life of their systems. The Register reported last month that Amazon had extended the working life of its kit from five to six years, a move it expects will save hundreds of millions of dollars in Q1 of this year alone. Microsoft and Google have also made similar moves.
Omdia estimates that the average life of servers located in enterprise datacenters or colocation has now increased to 7.6 years, while the hyperscale providers raised the average lifespan of their kit to 6.6 years during 2023.
Delaying the refresh of existing servers has allowed companies to free up more funds for AI clusters, Omdia claims, saying that if server life spans for these hyperscalers were to reach seven years, formal disclosures will likely be made. ®