The picture comes from Canva can draw the same internal and external atmosphere, the same economic momentum is still in the accumulation period, and the shareholders with similar state-owned backgrounds began to build positions in the secondary market, survived the winter, and saw 5 in the summer of 199919**；I wonder if there will be a similar result in 2024.
As a market with financing as the primary function, the biggest mission of AH shares is to be led by state-owned assets to support the development of the real economy.
If a company buys back almost all the 200 million profits earned in the first half of the year, what will be the purpose? "If no one buys themselves, then buy them back themselves". At this time, you stand with the interests of listed companies;
If the leading stocks in an industry have ushered in less than 5% of the shares of social security, insurance capital, and pension but are among the top ten, what is the meaning behind it? At this time, you have stood with the state-owned assets.
As long as it is the capital that enters the secondary market, it will not be difficult to compare with the potential "wind report". That's right, there are some state-owned assets that favor high-dividend sectors as Caijing ** said, but in the 2023 interim report, the funds represented by social security have already cashed out from some AIGC varieties that have risen more than 4 times, and returned to the low-level branches of AI in the third quarterly report, including but not limited to servers, industrial software, switches, and games.
"State-owned assets are in this position, and the purpose is to let social security and insurance funds buy enough chips." Regardless of whether this view is extreme, and regardless of whether there is a basis and information asymmetry for the increase of state-owned assets and the selection of targets, for leeks who have been mixed in A shares and H shares for many years, the only profit model is to stand with the funds with pricing power, buy low and sell high.
We just want to stand with the state-owned assets, and it's better not to have anyone else." There are not many A-share AI companies that can only have state-owned assets in institutional holdings, which is why the AI sector of A-shares only has a public offering at the end of the year** and a dense upper shadow line. Is it possible that there are some AI companies that are still heavily held by a small number of **institutions** and state-owned background shareholders?
Although it is not much, it is not nothing, the @Bairongyun-w (6608.) of Hong Kong stocksHK) is a representative, the list of institutional holdings in the F10 can be viewed by yourself, and the encyclopedia also has the background of the controlling shareholders of these holding institutions.
The company's business model is very simple, with two delivery methods, MaaS (Model as a Service) and BaaS (Business as a Service), the former allowing B-end customers to call AI models, and the latter using AI technology to help enterprises achieve KPIs. However, for enterprise customers, there is a problem that is very difficult in the industry but no one has raised in the secondary market: when deploying large model computing power, many enterprises still have a lot of work to do after getting the server, and it may take up to 30 days or even longer to build the environment, framework, and model, and then finetune needs to complete the privatization.
So what are the consequences of this? It is the high technical threshold and team cost of AI model development, and when enterprise customers finally deploy AI models, peer competitors have already completed product iteration.
Therefore, at this time, a technology is needed, which is also one of the key points of Bairong Cloud in the research and development of AI large models: MOE (hybrid expert model). What does that mean? To put it simply, it's an effort: training a larger model with fewer training steps with a limited budget of computing resources, as this tends to work better than training a smaller model with more steps.
A significant advantage of MOE is the ability to perform effective pre-training with far less computational resources than the dense model requirement. This means that the size of a model or dataset can be significantly increased with the same compute budget. Especially in the pre-training phase, MOEs are often able to reach the same level of quality more quickly than dense models.
If you look up the original scientific literature, you will see a lot of words such as "sparsity", "expert capacity", "token load balancing", which seem to be professional but very abstract, but in fact do not tell you what effect this thing can bring to customers. With Mandarin translation, customers can train larger private AI models with fewer training steps through MOE, so that the cost of computing power and localization are solved, and that's it.
MOE can only be based on Transformer and is naturally compatible with large AI models. The moe is made up of two parts:
(1) Sparse MOE layers: These layers replace the feedforward network (FFN) layer in the traditional Transformer model. The MOE layer contains several "experts", each of which is an independent neural network in its own right. In practice, these experts are usually feedforward networks (FFNs), but they can also be more complex network structures, or even the MOE layer itself, thus forming a hierarchical MOE structure.
In other words, when Bairong Cloud customers call MaaS services, for example, they call 10 models (i.e., services), which are all neural networks that run independently, so that customers can experience a more accurate independent operation mode while calling more AI models and services.
(2) Gated network or routing: This part is used to decide which tokens are sent to which "experts", for example, the tokens entered by B-end customers are "more" or "parameters", and these two tokens will be sent to different experts. Sometimes, a token can even be sent to multiple experts. The way tokens are routed is a key point in the use of MOEs, as routers are made up of learned parameters and are pre-trained with the rest of the network.
In the BaaS business model, customers of Bairong Cloud can use this principle to find the most accurate model stratification and "experts" involved in Bairong Cloud's BaaS according to the different keywords involved in their own KPIs, that is, tokens or natural language.
Some of these institutional customers of Bairong Cloud are state-owned assets, which also endorses Bairong Cloud's business model and AI model capabilities.
In the same way, if some shareholders also have a state-owned background, can they also endorse Bairongyun's stock price? There is no observation of each adjustment to 12Around 5 yuan, there is a force pulling up its stock price? Moreover, it is in the case of Hong Kong stocks.