ChatGPT and AI Could be the Bane of GPU Supply
It’s important to note Microsoft isn’t hoarding RTX 40 series GPUs for ChatGPT development. Rather, AI applications generally rely on data center GPUs. In the case ChatGPT’s beta version, developers Open.AI reportedly used 10,000 Nvidia A100 GPUs.
- ChatGPT’s next iteration is reportedly being trained on 25,000 GPUs.
- Mass deployment of ChatGPT – let alone its competitors – would require millions of GPUs.
- The H100’s 814 mm² die dwarfs the huge (by gaming standards) 608 mm² die of the RTX 4090.
- The older A100 data center GPU’s 826 mm² die is about twice as big as the mainstream RTX 3060.
From Nvidia’s perspective, data center – not gaming – GPUs are the future. The company has already heavily pivoted toward non-gaming customers like Microsoft and ChatGPT in recent years, and it’s paying off. Gaming revenue peaked for Nvidia in Q1 FY 2023 at $3.6 billion (46% of revenue); that same quarter, data center revenue overtook it ($3.8 billion). When asked about ChatGPT during a recent Q&A, CEO Jensen Huang spoke in glowing terms about AI's potential.
- Data Center revenue was more than double gaming revenue for Q3 2023, its most recent quarter.
This has major implications for gamers. For example, Nvidia shifted some of its TSMC RTX 4090 orders to H100 orders last year, likely in an attempt to sell as much before US chip restrictions for China kicked in.
An Opening for Intel and AMD
It’s not all bad news for gamers. Nvidia’s pivot toward non-gaming products is driven by its substantial technological edge over competitors. Any demand for gaming GPUs that Nvidia leaves unmet presents a golden opportunity for Intel and AMD to grow their market share while playing catchup in the data center segment.
- TSMC also currently has excess capacity as the likes of Apple and Nvidia reduce their orders during the economic downturn.
- US restrictions on China’s import of chips also limits the market for data center GPUs.
Featured imaged courtesy of Nvidia