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

The demand for Computing Power in large model training is driving Computing Power to become a new business model. Although the current craze for large model "alchemy" may pass, Computing Power service providers need to be proactive and prepare for changes in market demand.

Recently, researchers utilized 40 years of global weather data and conducted a pre-training for about 2 months using 200 GPU cards, ultimately training a meteorological large model with a parameter count reaching hundreds of millions. Based on a GPU usage cost of 7.8 yuan per hour, the training cost of this vertical domain large model could exceed 2 million yuan. If it were for training a general large model, the cost could increase a hundredfold.

Currently, China has more than a hundred large models with a scale of 1 billion parameters. However, the industry is rushing to develop large models while facing the dilemma of a shortage of high-end GPUs. The high cost of Computing Power and the lack of Computing Power and funds have become the most direct issues confronting the industry.

The shortage of high-end GPUs is difficult to solve in the short term. The explosion of large models has led to a rapid increase in market demand for Computing Power, but the supply growth is far behind. Although in the long run, the supply of Computing Power will certainly transition from a seller's market to a buyer's market, the duration of this process remains unknown.

In the face of this situation, it is widely believed in the industry that as competition in the large model market intensifies, the market will gradually return from frenzy to rationality, and companies will control costs and adjust strategies based on changes in expectations.

To address the shortage of Computing Power, enterprises have adopted various methods:

1. Train with higher quality data to improve training efficiency.

2. Enhance infrastructure capabilities to achieve stable operation of large-scale GPU clusters.

3. Optimize Computing Power card resource scheduling to improve resource utilization.

4. Improve network performance and enhance the efficiency of large-scale distributed training.

5. Transition from cloud computing architecture to supercomputing architecture to reduce costs.

6. Use domestic platforms for large model training and inference, replacing NVIDIA GPUs.

Computing Power has gradually formed a new service model in the context of market demand and technological iteration. Computing Power services are based on diversified computing power, linked by computing power networks, and aimed at effectively supplying computing power. It not only includes computing power but also unifies and packages resources such as storage and networks, delivering computing power in the form of services like API(.

In the computing power industry chain, upstream companies mainly supply basic computing power resources, midstream companies are responsible for computing power production and supply, while downstream companies rely on computing power services to provide value-added services. This model has more advantages in terms of cost and technology compared to companies building their own computing power environments.

With the normalization of high-performance computing demands for large models, Computing Power services are rapidly forming a unique industrial chain and business model. Although issues such as the current shortage of high-end GPUs and high Computing Power costs still exist, in the long run, the trend of Computing Power as a service has already been established. Computing Power service providers need to be prepared at all times to adjust their strategies promptly when market demand changes.