From Large Model "Alchemy" to New Computing Power Model

Although the hype around large model training is strong, the shortage of high-end GPUs and the high cost of Computing Power have become industry challenges. Enterprises are seeking innovative methods to cope with these challenges, while Computing Power services are also becoming a new business model.

Training large-scale AI models requires substantial Computing Power support. For example, using a certain meteorological large model, just utilizing 200 GPU cards for two months of training could cost over 2 million yuan. For general large models, training costs could reach several billion yuan. This makes it difficult for many small and medium-sized enterprises to bear.

Faced with the situation where high-end GPU cards are hard to come by, companies have adopted various response strategies:

1. Improve data quality and enhance training efficiency
2. Optimize the infrastructure to achieve stable operation of large-scale clusters
3. Improve Computing Power resource scheduling to increase utilization.
4. Replace cloud computing architecture with supercomputing architecture
5. Use domestic GPU platforms to replace NVIDIA products

At the same time, Computing Power services are forming new industrial chains and business models. The upstream provides basic Computing Power resources, the midstream is responsible for Computing Power production and scheduling, and the downstream consists of industry users. Cloud service providers and professional Computing Power service providers are becoming important midstream players.

Computing Power services mainly adopt two models: pay-per-use and annual or monthly subscription. Users can choose different forms such as GPU instances or MaaS platforms. In the future, we will promote "integrated computing and networking" to achieve flexible scheduling across architectures and regions.

Although the industry is currently eager to seize high-end GPU resources, in the long run, the service-oriented computing power is an inevitable trend. Computing power service providers need to prepare in advance for the transformation after the market returns to rationality.