Text: Wu Zhanguo, Editor: He Yueyang, Producer: Digital Intelligence Circle.
This year, the wind of large models has been blowing wave after wave in China.
Whether it is the Internet conference held in Wuzhen, or the major Internet forums, the big model has been praised by the bigwigs again and again, and now the smoke of the 100 model war has not stopped, and it will eventually be like the "100 regiment war, 100 cloud war", the winner takes all, it is still unknown.
At least for now, most of them belong to the R&D stage, and there are only a handful of them that can achieve profitability and monetization. However, the battle of cloud vendors for large model training has begun, and major cloud vendors have released their own service plans for large models.
Up to now, domestic and foreign cloud vendors, including Alibaba Cloud, Huawei Cloud, Tencent Cloud, Cloud, JD Cloud, Microsoft Cloud Azure and other cloud computing manufacturers, have launched MaaS services.
What is the impact of the launch of MaaS services on cloud vendors?How can cloud vendors monetize with MaaS?Can MaaS become a new business growth pole for cloud vendors?This article will answer these questions.
MaaS is a model as a service, which is a new concept in addition to IaaS (infrastructure as a service), PaaS (platform as a service), and SaaS (software as a service), and is also a new business for cloud vendors.
The National Institute of Standards and Technology (NIST), which has developed technical standards for IaaS, PaaS and SaaS, does not currently provide technical standards for MaaS, and there are no other authoritative certifications in the world.
According to the description of cloud vendors, the core value of the MaaS model can be summarized as follows: reducing the development technology and use cost threshold on the demand side of the algorithm, users can directly call the basic large model through APIs to build, train and deploy exclusive models for different business scenarios.
There are two reasons for mainstream cloud vendors to focus on MaaS services: active and passive.
First, the growth of traditional businesses of mainstream cloud vendors has slowed down.
In general, most of the traditional business of cloud vendors is based on the IaaS layer, while the business scale of the PaaS and SaaS layers is relatively small. The salient feature of providing services at the IaaS layer is that it is heavy on assets, heavy investment, and relies on scale, but the homogeneity is serious, so the profit margin is low.
Coupled with the addition of operators, the competition is more intense, and the first war is frequently staged.
On the other hand, cloud vendors such as Alibaba Cloud and Tencent Cloud have adjusted their previous routes to scale and actively abandon some low-profit projects to maintain profit margins. As expected by Internet cloud vendors, the profit margins of cloud vendors have improved in the past two years, but the growth has fallen below 10%. For example, Alibaba Cloud, which ranks first, will have a growth rate close to zero in 2023, which has dropped significantly compared with the 20%-100% growth rate in the previous two years.
In the PaaS and SaaS layers, which have higher profit margins, due to the lack of demand for basic development software and SaaS business, the business volume of these two layers in China is low, and the proportion of the entire public cloud market is much lower than that of foreign countries.
Second, there is a gap in computing power and service for large model training and operation.
As of October, there are 254 manufacturers and universities in China with large models above 1 billion parameters, and 238 large models have been released in China, compared with 79 in June, an increase of three times in four months.
The training and operation of so many large models require huge computing power. For example, OpenAI trains a large model, and the training of GPT-3 in the early stage requires the purchase of 49 servers at a time, at a cost of $1.4 million, and the daily operation cost is higher, with 25 million daily visits in the early days, 3798 servers need to be purchased, and the cost is 7$5.9 billion.
Huawei**, the total amount of general-purpose computing in 2030 will increase by 10 times compared to 2020, to 33zflops；The total amount of AI computing will grow 500 times to 105 ZFLOPS.
Based on this expectation, cloud vendors have launched their own MaaS services.
In March this year, Robin Li proposed at the Wenxin Yiyan press conference that MaaS will replace IaaS in the era of large models and become the mainstream. In April, Alibaba released the Tongyi Qianwen model, and at the Alibaba Cloud Summit, Daniel Zhang said that Alibaba Cloud has formed a three-tier architecture of model as a service (MaaS), platform as a service (PaaS) and infrastructure as a service (IaaS). From this statement, it can also be seen that Alibaba Cloud attaches great importance to MaaS.
At the beginning of July this year, HUAWEI CLOUD unveiled Pangu Model 30 and Ascend AI Cloud Computing Service, Pangu 30 can provide four series of basic large models from 10 billion to 100 billion parameters, and Ascend AI Cloud can provide computing services such as 2000p FLOPS in a single cluster.
At the beginning of September, Tencent released a self-developed hybrid model, and domestic enterprises can access hybrid elements through Tencent's public cloud platform and fine-tune them according to specific needs.
At present, there are different opinions on the definition of MaaS, including the relationship between MaaS and IaaS, PaaS, and SaaS, and whether MaaS will redefine IaaS, PaaS, and SaaS in the future.
At present, cloud vendors have launched MaaS-related services, mainly including IaaS-based AI computing services, as well as API call services through self-developed large models or open-source large models.
First, API call service is the core monetization method of MaaS.
API call service is the ability of the cloud computing platform to encapsulate the machine learning Xi model into a cloud service that can be called, and the user can call the model through API interfaces or other means.
In this process, cloud vendors can charge based on usage or time.
For example, OpenAI launched GPTS and corresponding natural language development tools. OpenAI has formulated a total of four charging models, namely ChatGPT Plus subscription fees, API (in addition to GPT model interfaces, including model fine-tuning interfaces and embedded interfaces) call volume charges, Wensheng diagram fees by generation volume, transcription text by minute charges, model instance rental fees.
Among them, GPT-3The 5 model is billed according to the number of tokens (decomposition units, roughly equivalent to one word in Chinese), charging 4 cents per 100,000 tokens. The top apps in the U.S. app market, Jasper for service marketing copywriting and chatbot Chat with Ask AI are both applications developed based on OpenAI's model, and their core cost is also the API call fee for OpenAI.
Second, it provides AI computing services for training and running large models.
Computing power service has always been the most basic service of cloud computing, that is, the service of the IaaS layer. IaaS was originally designed to provide enterprises with underlying technical services such as centralized servers and data storage, and later developed PaaS and SaaS on this basis.
The demand for AI computing power has reconstructed the servers, networks, and storage of the IaaS layer of cloud vendors. For example, in the server that provides computing power, it is necessary to purchase a large number of GPU servers equipped with NVIDIA and rebuild the system and network services built on top of it.
In the future, large models may also reconstruct the PaaS layer and SaaS layer.
On top of that,Cloud vendors can also explore new payment modelsFor example, based on the open source model, a developer community is formed, AI PaaS services are implemented, and developers are provided with services other than computing power and models, such as databases, middleware, and other services that developers need to train large models.
In addition, at the SaaS layer, more enterprise SaaS products are beginning to be based on AI (AI-based-SaaS), and AI will be explored at the application layer from SaaS auxiliary tools to AI-native SaaS products (SaaS products based on specific large models) and then to AI agent SaaS (agent-as-a-service).
First of all, we need to clarify who the cloud vendor wants to make when it provides MaaS services
The money of model manufacturers, including large models and industry models, such as Baichuan Intelligence, is running on Alibaba Cloud, and industry models are a trend. In the end, it is the money of B-end enterprises, and in response to the problem of AI application monetization on the C-side, we wrote in our previous article "Is the large model too rolly, is the AI application easy to do?".discussed in .
Cloud vendors can also do their own industry models, but each industry comes again, the investment is high, and the cycle is relatively long, which increases the difficulty of making profits.
When the large model investment fever returns to rationality, the ideal cycle is that enterprise customers can apply the industry model and the tools it provides to their own operations, production, financial management and other businesses, and after the application is improved, they will be willing to continue to pay, so that the model vendor makes money, and the cloud vendor also makes money.
The process of technology implementation is from model to tool to scenario, but commercialization starts from application scenarios.
China's low share of the public cloud market is partly due to weak profitability, so after evaluating the benefits and costs, there is not a high willingness to pay for SaaS.
Therefore, cloud vendors may encounter similar problems as SaaS through MaaS services, that is, from the supply side to the demand side, they have not formed a mature management model and market environment like foreign countries, nor have they formed standardized products, and cannot be replicated at low cost. (For details, see "The Death of China's SaaS: How the Gap Is Widening Step by Step").
Of course, MaaS is also different from SaaS, because with the assistance of large models, MaaS can have better adaptation to the needs of user customization, so the dilemma between standardization and customization will be solved to a certain extent.
Another important factor is the cost.
Providing MaaS services requires extremely high AI computing power, and cloud vendors need to purchase a large number of GPU chips to build new services to meet the growing demand for AI computing power. This year, Microsoft and Meta each bought 150,000 H100GPUs from Nvidia, and Ali and Byte bought 30,000 and 250,000 and 20,000 pieces. These chips will be deployed on new servers that are suitable for training and running large models.
In addition to external procurement, domestic cloud vendors are also stepping up their efforts to deploy self-developed AI chips or expand other purchase channels.
On October 23 this year, the United States began to implement new chip export controls, and Nvidia's high-performance AI chips - A800, H800, L40S, etc., were banned from exporting. As a result, domestic cloud vendors are unable to buy foreign high-performance chips, which has become a constraint for domestic cloud vendors to provide AI computing power.
According to a research report, a server equipped with Nvidia A100 chips costs $200,000, a single server is equipped with 7 A100 chips, and a single chip ** is 1Around $50,000.
The computing resources purchased by cloud vendors require a large enough amount of usage and user scale to generate benefits. And as the large model fever fades, no one can say whether there will be idle resources.