BIT Capital AI Whitepaper "Infrastructure of the AI Era"
The AI revolution will not be decided in the application layer alone – but in the infrastructure behind it. The second whitepaper shows why generative AI triggers massive demand for computing power and which companies along the technological value chain stand to benefit: from GPUs and high-bandwidth memory to networking technologies and specialised AI data centres.
- GPT-4 as an infrastructure benchmark: according to the paper, training required around 25,000 Nvidia A100 GPUs over 100 days – with implicit training costs of approximately USD 60 million.
- ChatGPT operations as a cost factor: with 132 million users at the time, daily operating costs were estimated at around USD 700,000.
- AI server market poised for strong growth: while traditional servers are expected to grow at around 7 percent p.a., industry experts forecast 30–50 percent p.a. for AI servers – up to USD 150 billion in annual revenue in 2028.
- Nvidia as a platform winner: through vertical integration of GPUs, networking technology, CUDA software, and AI applications, Nvidia is positioning itself as the dominant platform provider of the AI era.
- High-bandwidth memory becoming a bottleneck: the HBM market could grow 2.5-fold by 2027, reaching an annual volume of USD 5.2 billion.
- Data centres becoming a bottleneck: AI-related CapEx for data centres is expected to double within four years to USD 556 billion by 2026; GPU servers require 60–80 kW per rack, a multiple of classic CPU servers.
Note: The full whitepaper is available in German only.
