During the weekend, the domestic large model DeepSeek-v3 made a strong "outreach" as computing power is defined as national strength.

Today, the computing power sector is performing actively and leading the market.

As of the time of publication, Shanghai Anoky Group is at its daily limit, Zhongchen Co., Ltd. has risen over 16%, Guangdong Qunxing Toys Joint-stock, INESA Intelligent Tech Inc., and Dianguang Explosion-proof Technology have hit their daily limit, Guangdong Aofei Data Technology is up over 8%, Netac Technology and Parallel Technology have increased over 7%, Sharetronic Data Technology and Lianhua Holdings have risen over 6%, and Beijing Topnew Info&Tech has climbed over 5%.

Will the demand for computing power decline.

Over the weekend, the Algo giant Huanfang’s self-developed large model DeepSeek-v3 has gone viral both domestically and internationally, sparking a big discussion in the industry.

On the one hand, DeepSeek-v3 has a very low cost, using only 2048 GPUs for training over 2 months during the pre-training phase, costing just 5.576 million dollars.

On the other hand, in terms of performance, DeepSeek-v3 outperformed other open-source models such as Qwen2.5-72B and Llama-3.1-405B in multiple evaluations, and is on par with the world-class closed-source models GPT-4o and Claude-3.5-Sonnet.

However, it should be noted that the 2048 H800s do not account for all the training computational power invested.

This does not include the computational power consumed by the high-quality data generated from the DeepSeek R1 model (benchmarking OpenAI o1), as well as the computational power spent during the exploration and debugging phase of the model architecture.

It is reported that the reason DeepSeek was able to develop a powerful large model with relatively low computational power is primarily due to distilling the inference capability from a model called R1 onto the V3 model during post-training, which made the post-training phase virtually costless.

The success of DeepSeek-v3 has sparked a great discussion regarding computational power and the training methods of large models, with some investors concerned that the industry's demand for computational power may drastically decline.

However, analysis suggests that even if large models like DeepSeek-v3 are successful, investments in computational power by tech giants such as ByteDance, Xiaomi, and Li Auto will not decrease; rather, these success stories reveal industry potential and are likely to encourage further investment.

Moreover, the success of DeepSeek is expected to inspire other domestic companies on how to utilize computational resources more efficiently, which may encourage more small and medium-sized enterprises to enter the market.

In the future, as development costs decrease, AI applications and software are likely to usher in significant opportunities.

The official stance is: Computing power equals national power.

On December 28, a large series documentary titled "Cornerstone of a Great Power" produced jointly by the Central Broadcasting - TV Station and the State-owned Assets Supervision and Administration Commission launched its third episode "Computing Power Engine," which clearly states that computing power equals national power, as it is a new quality productive force in the era of digital economy.

As of 2023, China ranks second in the world in terms of total computing power, with a total of 14 national-level supercomputing centers established, and there are 633 ultra-large and large datacenters currently in use across the country, along with 60 intelligent computing centers.

In investment, CITIC SEC believes that the release of the new generation model DeepSeek signifies that the application of AI large models will gradually move towards universal accessibility, facilitating the widespread implementation of AI applications; at the same time, the significant improvement in training efficiency will also drive a high increase in demand for inference computing power. The exponential increase in computing power demand, combined with the reduction in model training and inference costs, will form an industrial flywheel, driving the expansion of the entire Industry Chain. As AI further permeates daily life and industries, the demand for inference computing power will continue to drive sustainable growth in the AI computing power Industry Chain.

China Securities Co.,Ltd. points out that currently, with the increase in demand for AI computing power, particularly positive changes occurring on both the domestic supply and demand sides, the demand for datacenters will also rise accordingly, especially for high-power ultra-large datacenter rooms. This also imposes higher requirements on aspects such as electrical utilities and temperature control, and it is recommended to focus on the IDC Industry Chain, including IDC service providers, electrical equipment, liquid cooling temperature control, and other related areas.