Computing Power Services: A New Business Model in the Era of Large Models

With the rapid development of large models in artificial intelligence, Computing Power is becoming an emerging business model. Although the current trend of "refining elixirs" with large models is still ongoing, Computing Power service providers need to be prepared for future market changes.

Recently, a young man who graduated from Tsinghua University three years ago trained a meteorological large model with a parameter count in the hundreds of millions. This project utilized 40 years of global weather data and employed 200 GPU cards for pre-training, taking about 2 months. Based on current GPU prices, the training cost alone could exceed 2 million yuan. If it were to train a general large model, the cost could be multiplied by a hundred.

Currently, China has more than 100 large models with a scale of 1 billion parameters. However, the industry generally faces the problem of a shortage of high-end GPUs. The cost of Computing Power remains high, and the lack of Computing Power and funding has become a real challenge for many enterprises.

In a situation of high demand and low supply, the prices of high-end GPUs have been driven up to high levels. The price of an NVIDIA A100 reached as high as 200,000 to 300,000 yuan at its peak, and the monthly rental for a single A100 server also soared to 50,000 to 70,000 yuan. Nevertheless, many enterprises still find it difficult to obtain the necessary chip resources.

Faced with this dilemma, companies are actively seeking countermeasures. Some choose to use higher quality data to improve training efficiency; some are committed to enhancing infrastructure capabilities to achieve stable operation of large-scale GPU clusters; while others turn to using domestic platforms for large model training and inference.

As the market gradually becomes more rational, companies are also adjusting their strategies to control costs. At the same time, Computing Power services are becoming a new business model. Computing Power service providers deliver Computing Power to users in the form of APIs by integrating resources such as Computing Power, storage, and networks. This model allows users to obtain the required Computing Power support simply by stating their needs, without having to build their own infrastructure.

In the computing power industry chain, upstream companies are responsible for providing basic computing power resources, midstream companies are responsible for computing power production and scheduling, and downstream are users from various industries. With the development of computing power services, new billing models and service forms are also emerging continuously, such as pay-as-you-go, annual and monthly subscriptions, and integrated computing and networking.

Although the current shortage of high-end GPUs still exists, in the long run, the computing power service market will inevitably transition from a seller's market to a buyer's market. In the face of this trend, computing power service providers need to be well-prepared to adjust their strategies promptly when the market shifts and seize new development opportunities.