OpenAI has launched the GPTs function, greatly reducing the development threshold for AI applications. Global AI applications are about to enter a new stage of qualitative change caused by quantitative change. With the continuous emergence of popular applications, the demand for computing power will surely grow rapidly. The ban on sales of Avida in China has led to an increase in the shortage of computing power. Domestic computing power is expected to replace Avida's market share. Domestic AI chips, AI servers and AI all-in-one machines will usher in development.

(1) What is an AI big model all-in-one machine?

The essence of AI big model all-in-one machine can be understood as AI server superposition big model private deployment. The big model all-in-one machine is a combination of software and hardware, which integrates AI chips, AI platform software and various model algorithms. Generally speaking, the big model training process needs to transmit data to the central server or the cloud for calculation, and then return the results, which involves a large amount of data transmission and network latency. The big model all-in-one machine presets the model in the server, which can realize the privatization deployment of the big model.

The training and push all-in-one machine generally presets the AI big model engine, and also integrates users' requirements such as knowledge base Q & A, Knowledge Graph, AI raw graph and other functions, so that the corresponding chip has higher computing power, it is easier for users to develop AI applications in their own environments. Therefore, the requirement of becoming a supplier of all-in-one training and pushing machine is higher.

More than 10 manufacturers, such as xunfei and Yuncong, have released a solution for the all-in-one machine of the upstart AI model, covering many vertical fields such as government affairs, education, medical care and security.

(2) Why do I need an all-in-one AI machine?

Cause 1: Local deployment, data is not disconnected from the internal network

compared with Internet customers, customers in government, finance, telecommunications and other industries have a large amount of sensitive data and are considering data security and privacy security, and upload data to the cloud for training. In fact, some large foreign enterprises have restricted the use of AI. According to Insider reports, some large companies are restricting their employees' access to ChatGPT. Amazon and Apple, for example, have expressed concern about the possible risk of data leakage caused by artificial intelligence.

Reason 2: combination of software and hardware, out of the box

in the whole process of big model development, enterprises often face complex problems such as switching hardware clusters, adapting underlying hardware, fine-tuning and optimization. The AI all-in-one machine can seamlessly switch the training and reasoning processes of large models in the same cluster, thus providing customers with efficient large model training and reasoning deployment processes. At the same time, the solution has built in a large model inference engine, tuning tools, and operator acceleration library to support customers to use the model out of the box without secondary hardware adaptation, therefore, the technical threshold is lowered. In addition, the hardware performance can be fully released through algorithm optimization at the chip level, thus improving the efficiency of model training and reasoning.

According to the data of the Institute of Information Technology, the industry with the largest proportion of AI computing power applications in China is the internet, followed by government, service industry, telecommunications, finance, manufacturing, etc. Among them, the proportion of computing power application in key industries such as government, telecommunications, finance, education and transportation is close to 40%.

(3) What is the market space for AI all-in-one machines?

According to the data of the National Bureau of Statistics, there were more than ten thousand large enterprises, nearly 30000 medium-sized enterprises and nearly 1.24 million small enterprises in the country in 2022; According to the data of the National Bureau of Statistics, there were 37000 hospitals in the country in 2022;

according to the 2022 National Educational Development Statistical bulletin issued by the Ministry of Education, there are 518500 schools of all levels and types in the country;

according to prospective network, as of August 2023, there were 40 intelligent computing centers in China, and the compound growth rate of intelligent computing power scale in China reached 47.58% from 2022 to 2026.

According to the announcement of Hengwei technology on November 16th, the company won the bid of 480 cabinets of AI computing power all-in-one machine, with a total bid amount of 2.112 billion yuan. The average unit price is 4.4 million yuan/set.

At present, a number of companies have released the solution of the all-in-one machine, involving many vertical industries such as government affairs, medical treatment, remote sensing, etc. We believe that AI big model all-in-one machine has become a consensus in the industry and is expected to become the mainstream form of domestic computing power.

Therefore, the AI all-in-one machine market is close to hundreds of billions only in the enterprise market (each medium and large enterprise purchases an AI all-in-one machine). If the market of schools, hospitals and data centers is calculated again, the market will be even larger. If Avida completely prohibits the export of H800 chips to China, the current AI all-in-one machine market may become the upper market space for domestic AI all-in-one machines such as Huawei AI all-in-one machines.

Source: architect Technology Alliance