The weather is getting cooler and summer is coming to an end. If someone asks, what are cloud computing vendors doing in this hot summer? The answer is, they do three things: big models, big models, and big models.
In July, during the Huawei Developer Conference 2023 (Cloud), Huawei Cloud announced Pangu Model 3.0, and subsequently carried out ecological and other upgrades. Earlier, Alibaba Cloud announced the Tongyi Qianwen large model and announced on August 3 that it was officially open source.
Come September, cloud computing vendors are releasing large models more intensively. On September 5, Baidu Smart Cloud released Qianfan Large Model Platform 2.0 at the 2023 Baidu Cloud Intelligence Conference, further integrating Baidu’s Wenxin series of large models. Just two days later, Tencent released Hunyuan at the 2023 Tencent Global Digital Ecology Conference For large models, its main export is Tencent Cloud.
At this point, China's major cloud computing vendors can be said to have gathered heavily in the field of large models. Compared with AI algorithm companies and research institutions, the large models produced by cloud computing vendors are closer to the front lines of industry and application, and the large models have a relatively complex relationship with the cloud vendors' original business systems and revenue modules. Therefore, the big model battle between cloud vendors is by no means a simple technical comparison between models.
So, why do major cloud computing companies make big models? Where is the decisive point in the cloud + large model battle?
In this article, we hope to answer this question with you through multiple levels of comparison and combing.
First, two premises must be clarified: First, with the reduction of IT investment in various industries, cloud computing manufacturers are generally facing a slowdown in growth and unfavorable revenue. According to the IDC report, the average growth rate of China's cloud computing industry will be reduced by about 10% in the next five years, and cloud computing manufacturers are generally lowering their business expectations. Therefore, the sudden explosion of large models is a rare window for the cloud computing industry. This opportunity may not be as profitable as imagined, but it has to be done now.
Another situation is that the requirements for cloud + large models are very complex from the user interface point of view. Some users need to directly access the model, some users need AI computing power to train the model themselves, and some users need a lot of model customization and solution integration. Therefore, cloud vendors' expansion of models is not the hand-to-hand battle that many people imagine. Businesses related to large models need to target multiple markets and multiple business models. This event is more like a positional battle, and there must be no gaps at any strategic node.
At every level of this fierce model battle, there is an open line and a hidden line leading the development of the situation.
IaaS layer: open-line stacking
Secret line launches domestic AI computing power
The most direct change that large models have brought to cloud computing manufacturers actually does not come from the models themselves, but from the fact that after large models became popular, the large model craze that emerged required huge AI computing power. The scale of large model data is huge, and training models requires dedicated AI computing power, which has brought about a short-term upgrade in cloud usage for cloud computing service providers. With the overall cloud computing IaaS layer market relatively sluggish, the emergence of large models can be regarded as a shot in the arm.
At this level, the competitive nodes of cloud computing vendors can be summed up in one word as "stacking." Whoever can provide sufficient AI computing power with less queues and as cheaply as possible will win. The main source of AI computing power is NVIDIA's GPUs, so there has been a phenomenon of cloud computing manufacturers short-selling GPUs on the market, and the saying that "cloud manufacturers all work for NVIDIA."
But no matter what, the competition among cloud computing vendors at the IaaS layer will not end. A large number of users will still use computing cost and computing efficiency as the basis for choosing public cloud AI computing power. At this level, cloud computing manufacturers need to improve their capabilities in computing clustering and computing compatibility, and try to maximize the value of each GPU.
For example, Baidu Intelligent Cloud promises to achieve more than 95% of the effective training time when training large models on the Wanka cluster, thereby reducing users' training time costs and achieving better computing acceleration. Alibaba Cloud has proposed a single cluster scale that can support up to 100,000 GPUs, allowing multiple large models with trillions of parameters to be trained online at the same time.
In addition to the bright line of stacking cards, the competition between cloud + large models at the infrastructure layer actually has a hidden line, which is the localization of AI computing power.
Not long ago, there was a lot of news about the unstable supply chain of Nvidia's high-end GPUs to the Chinese market. Later, it was reported that the "China Special Edition GPU" was expensive and had low performance. These phenomena have increasingly made people from all walks of life see the inevitability of AI computing power becoming autonomous and controllable.
In this round of large-scale model craze, the localization and cloud acquisition of AI computing power have turned from a trend into a reality. There are two main ways. One is that cloud computing manufacturers are compatible with more domestic chips and provide diversified AI computing power. For example, manufacturers such as Tencent Cloud and Baidu Smart Cloud are strengthening the compatibility of their domestic software and hardware and building an AI computing-related ecosystem.
At this point, Huawei Cloud has a natural advantage. In the years after being sanctioned, Huawei has gradually made its independent AI computing ecosystem bigger and stronger, and has become a relatively mature branch of domestic AI computing. Along with the upgrade of Pangu's large model, Huawei Cloud also announced that it will provide independent AI cloud services to provide a computing power base for large model training. This means that Huawei's independent AI computing power has officially moved from offline to the cloud.
Emphasizing the autonomy and controllability of large models and AI frameworks has become a general trend. Next, there is reason to believe that the model of public cloud + domestic AI computing power will continue to rise driven by macro trends. Eventually it became the key variable influencing the IaaS market.
Model layer: MaaS implementation
Hidden lines reduce customization costs
From an industrial perspective, what are the biggest changes that large models have brought to cloud computing? Judging from the current situation, the new business model of MaaS is the most important point for cloud vendors. The so-called MaaS refers to cloud vendors directly delivering large AI models to users, thereby realizing model as a service.
At least in the current initial stage, various cloud vendors have high hopes for the new model of MaaS. Some vendors have even directly replaced the previous SaaS with MaaS. It seems that after a long period of success, the SaaS model will finally be abandoned by the cloud industry. No wonder. After all, the unit price of SaaS in the Chinese market is low, but it requires a lot of customization work and consumes post-service services. Judging from the comprehensive costs of large manufacturers, SaaS has always been difficult to get rid of the useless positioning of being tasteless and a pity to abandon.
In this case, it is better to turn around and embrace the hotter big model. Therefore, we can see that various cloud vendors will collectively develop the new business model of MaaS in 2023, and have made a series of efforts to this end.
In the first stage of entering MaaS, cloud vendors mainly focus on three aspects:
1. The basic models should be numerous and precise , able to meet the diverse needs of users in several general directions such as NLP, CV, and multimodality. At the same time, the basic model is still the large model facade of cloud vendors. The experience of the basic model determines the first impression of users and developers on the large model capabilities of cloud vendors. For example, the popularity of Wenxin Yiyan has brought obvious brand blessing effects to the Wenxin series of large models and Baidu Smart Cloud's MaaS services.
2. Key areas should be covered . In industry categories and application categories that may be called frequently, manufacturers should try their best to make mature large models and applications developed based on large models, and try to achieve low-threshold integration and out-of-the-box integration. use. For example, Tencent Cloud's industry model selection store not only provides Hunyuan models, but also lists industry models in more than 20 fields such as finance, cultural tourism, and retail. Industry large models have become the backbone of the MaaS model.
3. Provide sufficient tool capabilities. In addition to basic models and high-frequency customized models, there are also massive demands for large models. This requires users and developers to have certain development capabilities, and cloud vendors need to provide tool chains for model fine-tuning and application development. For example, Baidu Intelligent Cloud's Qianfan platform provides pre-made data sets, application paradigms, and other tools to help enterprises apply large models in addition to models.
But the MaaS model that seems to be in full swing actually hides a hidden problem: SaaS doesn't make money, so it turned around and embraced MaaS, but who said MaaS makes money?
Cloud vendors have been doing AI for many years. But the result is often good results and poor profits. The key point here is still the key reason why SaaS stalls: the cost of customization.
Whether it is enterprise application of AI or current application of large models, AI as a software is inherently uncertain. The needs of different enterprises vary greatly, and AI has extremely high costs in terms of computing power, data and talent. Any customization will have a series of chain reactions. It sounds nice to have experts working in factories, but in the end, factories cannot afford the salaries of experts. But if customization is not provided, there will be the embarrassment that most AI needs cannot be met.
Therefore, how to reduce the customization costs that may arise from the MaaS model from the top-level design stage has become the core competitive secret in the battle between large models.
In this regard, each manufacturer's exploration scope and implementation methods vary. For example, Baidu Smart Cloud prefers to use preset models and preset applications to reduce customization. In addition to the model platform, Baidu Intelligent Cloud also released the "AI native application Family", which allows enterprises to meet their own needs through the model + large model application model.
In this regard, Huawei Cloud is currently exploring the most. Huawei Cloud's Pangu Model 3.0 builds a "5+N+X" three-layer architecture from the framework design. This architecture divides the large model into a three-layer system of L0, L1, and L2. The L0 layer includes five basic large models of NLP, CV, multi-modality, prediction, and scientific computing; L1 is a large model for N industries, such as government affairs, mining, finance, etc.; L2 is a detailed scenario model for various industries. Such as lead drug screening, conveyor belt foreign body detection, etc.
The design concept of this framework is that users, partners, and developers can call different levels of models to combine and assemble them according to their own needs. They can either directly call the integrated model, or perform fine-tuning based on the model, or they can also be trained by different developers. specific scene model. The design form of this idea has the characteristics of modularity and componentization in the industrial revolution, but the specific effect has yet to be tested in the industry.
Overall, what cloud vendors hate most about MaaS is the need for high customization, low reusability, and heavy workload for follow-up services. This will over-disperse the limited resources of the original factory, and the final return will be difficult to achieve. This is the "small workshop-style AI development" often discussed in the cloud computing industry.
At the current stage, Cloud Factory can only invest in MaaS without output. But in the long run, whether the change from small workshops to assembly lines can be realized will be the last battle that determines the life and death of MaaS.
Ecological layer: bright lines gather everyone’s strength
The battle over hidden open source
We need to continue talking about the problem of large models. The cost of customization is too high. Because of the high cost of customization and limited comprehensive resources of a single service provider, cloud vendors must subcontract a large amount of work. It is up to partners to complete consulting, services, distribution and other work. Otherwise, the original factory will be repeatedly tortured by lengthy processes and huge service costs. This also leads to the fact that when moving towards the MaaS model, cloud computing companies need to build a partner ecosystem more than ever.
On the other hand, cloud vendors currently not only need partners, but also need to gather as many application developers as possible. AI large models are a new thing, and the new application models they can create are very imaginative. Just as the iPhone moment relies on a large number of APP developers, the "new iPhone moment" of large models also relies on a large number of AI developers with breakthrough capabilities. For this reason, major Internet companies are making their own applications on the one hand, and they also need to recruit more application developers on the other. Because any application that becomes popular will bring a series of knock-on effects to cloud vendors that provide basic models and computing power.
The combination of these factors makes the cloud + large model battle become an ecological battle from the very beginning. It can be said that various manufacturers are setting up their positions and doing everything possible to attract partners and developers to join their own ecosystem. Among them, the basic measures are to provide the technologies and capabilities required by developers and partners, and the advanced model is to promote large-model-based skills training, application development competitions, free resources, and joint entrepreneurship plans to empower partners and developers at the commercial level. . There are also cloud vendors that advocate full-scale collaboration and joint innovation with partners, and implement large models mainly with partners in market segments and segmented scenarios.
In the ecological battle over how to attract developers and partners, hidden hidden threads are a proposition very characteristic of Internet thinking: Can we simply make the model open source and free, and use extreme cost reduction methods to attract partners to join?
This is also one of the sources of the "big model open source and closed source dispute" that has been hotly discussed recently.
Those who support open source believe that free is the best ecological aggregation strategy, and the strategy of relying on free traffic has been tried and tested in the Internet era. Coupled with the large AI model as the basic software, it is probably only a matter of time before it becomes open source.
Opponents argue that large models are still in their infancy. Manufacturers need to continue to invest a lot of R&D costs for upgrading. Blindly free of charge will cause the development of large models to stagnate and disrupt the order of scientific and technological progress. In addition, open source of large models can not only reduce the costs of partners and developers, but also lead to lower final market pricing, dilute partners' profits, and ultimately lead to enterprises being unwilling to invest in R&D and innovation.
In any case, the confrontation between open source and closed source for large models has changed from an industrial discussion to a reality in the cloud computing market. On August 3, Alibaba Cloud announced that Tongyi Qianwen is open source, becoming the first Chinese Internet cloud vendor to announce that a large model is open source. ModelScope, the AI model community created by Alibaba Cloud, is also open source, free, and commercially available as its main selling points. Provide various open source large models at home and abroad.
Since then, will more and more cloud vendors move towards open source under the catfish effect, or will they still maintain the confrontation between high technology and low cost? let us wait and see.
From my personal point of view, the development potential of large models is still great, and there is a lot of room for exploration. In fact, developing technologies are not suitable for quickly moving towards open source. Therefore, it is a more likely event that a considerable number of large models remain closed source business models.
Solution layer: Mingxian’s first batch of major customers
New window for hidden government and enterprise cloud migration
No matter how exciting it is, in fact, enterprises directly call large models through API interfaces, and the profit of this MaaS service model is not high. It can even be said that in the context of the current "Battle of Hundreds of Models", simply calling large models has become a rather cheap thing.
The cloud computing industry, which continues to invest, is obviously not satisfied with this business model, so it must strive for projects with high unit prices and large profits. The choice of large model-based digital solutions by large government and enterprise customers has become a new window of opportunity in the eyes of cloud computing vendors.
Among the many types of large government and enterprise customers, there are only two types that have money, digital capabilities, and are willing to explore the possibilities of large models as soon as possible: smart cities and finance. Other industries such as manufacturing, energy, and transportation that are more physical-oriented are relatively more cautious and are still waiting to understand the big models.
Therefore, competing for orders from large government and financial customers has become a standard move for cloud computing manufacturers after entering the large-model track. For example, we can see that at the 2023 Baidu Cloud Intelligence Conference, Baidu Intelligent Cloud released Jiuzhou, a large-model-based digital government solution. Huawei Cloud is strengthening and promoting solutions based on the Pangu model in fields such as finance and smart cities.
In the foreseeable future, cloud computing vendors will not only compete in model capabilities and model platforms, but also compete in key areas represented by cities and finance. Answering why large models are needed in these fields and what different values large models can bring are the first questions that cloud vendors must answer.
There is also a hidden line hidden in building solutions for large customers. This clue is directly related to the anxiety of cloud computing vendors: large government and enterprise customers are less willing to go to the cloud.
A few years ago, it was a general trend for large-scale government and enterprises to move to the cloud. It can be said that they can move to the cloud as much as they can, and move to the cloud as quickly as possible. However, at the current stage, government and enterprise customers place more emphasis on the security and controllability of data and digital systems, and blindly moving to the cloud has been criticized. considered undesirable. In particular, whether a large amount of critical data can be put on the Internet cloud platform has gradually become a question. At this stage, the terms state-owned cloud and national cloud are constantly being strengthened. Even if they go to the cloud, large government and enterprises will give priority to self-centered, multi-cloud procurement strategies, which invisibly disperses the profit margins of cloud vendors.
In this context, an implicit expectation of cloud vendors for large models is that they can become a technological opportunity to promote large-scale government enterprises to continue to move to the cloud. After all, obtaining large models on the cloud has natural cost and operability advantages.
Therefore, whether the cloud can better demonstrate the importance of deploying large model solutions for large government enterprises and key real industries, and at the same time dispel the doubts of governments and enterprises in the fields of data security, independent controllability, continuous service, and brand trust, has become the cloud Another must-answer question for manufacturers competing for large models.
In fact, there are still many points to discuss about the combination of cloud computing and large models. For example, the combination of large models and PaaS; the combination of large models by cloud vendors in toB applications such as offices and network disks, etc. Overall, computing power, MaaS, ecology, and large-scale government and enterprise solutions constitute the four levels of whether cloud computing vendors can gain competitiveness in the battle of large models.
No matter which public cloud vendor, the goal of this competition is the same: to make models useful, reduce costs, and make AI a starting point for profitability.
Will the big model eventually become another piece of cake, or is it the door to a new era? The road is long and long, and cloud computing still needs to be explored.
