To make money, to land, Huawei large model is not and Baidu Aligang | Focus analysis

Author | Yuan Silai

Edit | Su Jianxun

Big companies are flocking into the big language model, and Huawei seems a little out of place.

On July 7, Huawei released version 3.0 of Pangu, an AI large model. For ordinary users, this launch will be a bit boring. Although the big government affairs model has AI responses, they are all policy explanations, which is not interesting. Zhang Pingan, CEO of HUAWEI CLOUD, for the first time, clarified the positioning of his big model, that is, "born for the industry", focusing on B-end industry customers. He said with some pride that Pangu's big models are solving the pain points of the industry, "no time to write poetry" and "no time to chat".

This statement is not surprising. Three months ago, Tian Qi, chief scientist of HUAWEI CLOUD artificial intelligence, made a preview. In his speech, Tian Qi talked about the GPT natural language processing (NLP) big model, and spent a lot of time introducing the big model developed for industry customers.

Pangu Large Model 3.0 has more multimodal large models than previous releases. But Huawei doesn't have much to say about the significance of this big model. The shadow of the large natural language processing model appears in intelligent assistant and government response scenarios.

Huawei's style of entering the AI industry is not sexy enough, but practical enough. It is very smart, not-for-tat with Internet companies, and the anchor is concentrated on the B side where it is best and easiest to land.

This is also the best choice Huawei can do. Huawei's 2022 annual report shows that its sales revenue growth is close to stagnation, with net profit falling by 68.7% and operating cash flow falling by 70.2%. Although Huawei is working research and development, in fact, it does not have much time and money to test the direction it is not good at. Selling HUAWEI CLOUD products and earning cash from industry customers as soon as possible is the most urgent thing for Huawei to do AI.

Only for the B-side

Huawei Pangu large model has not had a new story for a long time. In April 2021, Huawei officially released the Pangu Big Model, which already includes natural language processing big models, machine vision big models, and scientific computing big models.

Since then, Pangu large models have occasionally appeared in industry customer cooperation cases, and corresponding drug research and development, mining, meteorology, waves, financial OCR (text recognition) large models have been launched.

As for Huawei's natural language processing model, what can be determined is that its parameters and data are large enough, the amount of parameters reaches 100 billion, and the data is claimed to be 40TB, slightly smaller than GPT-4, which is ten times that of Wen Xin's words. At the same time, there are applications such as document retrieval, ERP, and small languages.

However, this large natural language processing model has never revealed its true identity. No live demos, no application demonstrations, and lack of GPT-like flashpoints.

At this conference, the Pangu natural language processing large model is still very "boring". Huawei has updated some information, including more than 100 billion, 38 billion, and 10 billion parameter NLP models that can be provided to customers to meet different requirements of different industries. In the big financial model, AI answers policies and analyzes data for financial analysts. It can also help government staff analyze traffic pictures in a responsive way.

These applications are far from C-end users, but can really improve productivity and attract industry customers. "We're not on the same track as ChatGPT, we're looking at how many problems are being solved in the industry." At the media conference, Zhang Pingan said.

The high unit price and stability of industry customers make all manufacturers red. Microsoft, Tencent, Baidu, and Ali have spent a lot of money to make a big model, and eventually they have to sell with the government and enterprises.

Each large factory has its own areas of expertise, such as e-commerce and office, while Huawei has a good reputation in the industrial intelligent manufacturing, government affairs, and financial industries. They have dealt with these companies for decades and have access to a wealth of authorization data from industry customers.

Therefore, the Pangu 3.0 conference is basically a collection of government and enterprise customer cases of Pangu's large model. These include boring policy answers, financial data analysis, and supply chain component allocation that is completely incomprehensible to laymen. The purpose of these products is to improve the efficiency of industry workers, not to please C-end users. Zhang Pingan said that for a period of time, Huawei's Pangu large model will not be open to personal use.

At the press conference of Baidu, Alibaba, and iFLYTEK, the keynote speaker and AI have become a must-have column. In comparison, the release of Huawei's Pangu large model will be short and interesting. But for Huawei, fun doesn't matter at this moment. As Zhang Pingan, Executive Director of Huawei and CEO of HUAWEI CLOUD, made clear at the press conference, Huawei's large model "does not write poetry". Satisfying factories, state-owned enterprises, and government customers is Huawei's concern.

Hashrate sellers

Huawei is a gold digger and a shovel seller.

Huawei is a set of self-developed systems from the underlying chip, database, development framework to development platform. The advantage of full-stack technology is that it is high enough to build a high barrier and it is easier to optimize deeply, and the opposite is poor compatibility.

Huawei's self-developed Ascend 910 chip performance and NVIDIA, Huawei's self-developed framework MindSpore can stimulate maximum performance. But even in China, even in China, AI developers are still using TensorFlow and PyTorch, both of which are used by more than 30 percent, while MindSpore is used at 11 percent.

This also means that Huawei has to do in-depth customization for customers to make different models run smoothly in Huawei's Ascend native system.

This is a project that is difficult to generate marginal benefits, and it is more cost-effective to package B-end customers by industry and make vertical professional solutions.

Industry customers are naturally adapted to Huawei's closed AI model. In particular, large state-owned enterprises and financial institutions need to establish a complete set of customized solutions from the bottom layer to the application side, and these customers require security, reliability, localization, and relative closure, and Huawei Pangu can meet the requirements.

Heavy vertical industry customers have always been Huawei's strength, and it is difficult for other Internet companies to dig deeper. Industry products and C-end products have completely different operating logics, but Huawei has accumulated a complete set of ways to deal with government and enterprise customers in the ICT era, and HUAWEI CLOUD does not lack government and enterprise resources, which are not barriers that Internet upstarts can obtain by throwing money.

More importantly, compared with Internet companies, Huawei has solved the chip problem that cannot be avoided. At present, Ali and Baidu purchase mostly NVIDIA chips in China, and Ali has the most NVIDIA A100 GPUs, with obvious advantages in computing power. However, it is not yet known whether this supply will be stable. The vast majority of domestic chip manufacturers are not yet able to supply chips for training complex large models.

Although Huawei's Ascend chip is slightly less adaptable, it has strong computing power and is well optimized in its own system. As long as there are no manufacturing problems and no supply interruptions, Huawei can always update and upgrade. In the artificial intelligence computing power competition, it will not be too far away from NVIDIA's latest chip. This is enough to reassure government and enterprise customers.

Self-contained computing power has become Huawei's biggest competitiveness and will also be another source of revenue for Huawei. Huawei launched Ascend Cloud for the AI industry and built two AI data centers. Customers who pay more and more attention to domestic substitution will not let Ascend Cloud lack orders.

OpenAI has once again ignited the AI industry after many years. But this road is still not clear in the direction, Huawei's giant ship will sail along the most secure route at present. In China's huge market, Huawei has always been in a unique position, and its advantages on the B-side for more than 30 years will continue into the AI era.