The heat of the large model continues to ferment. The arrival of MaaS has not only changed the competitive landscape of cloud vendors, but the SaaS industry will also usher in a "singularity" moment. In the next ten years, based on the MaaS base, domestic SaaS may even have a giant like Salesforce.
Author | Sihang
Editor | Pi Ye
Produced | Industrialist
The popularity of large models continues to ferment.
In this regard, although many people in China are keen to pursue it, there are also those who pour cold water on it. In fact, the power of the big model does not lie in its technological innovation, but in the "emergence" after the accumulation of data and parameters to a certain extent. This also well explains the seemingly contradictory views of many well-known domestic experts.
Among them, the most enthusiastic voices are Dr. Lu Qi, Dean of YC Global Research Institute, and Academician Zhang Yaqin, Dean of Intelligent Industry Research Institute of Tsinghua University.
"Anything that changes society and industry will always be a structural change. This structural change is often a type of large-scale cost, from marginal cost to fixed cost." In his speech some time ago, Lu Qi directly poked "AI Big Model" "The nature of the boom. OpenAI does the same thing as Google. In 1998, the birth of Google turned the marginal cost of obtaining information into a fixed cost. Today, OpenAI also makes the cost of the model move from marginal to fixed.
Academician Zhang Yaqin, dean of the Intelligent Industry Research Institute of Tsinghua University, also bluntly stated that in the next ten years, large models may become the "operating system" of the next AI era. From Windows in the PC era to iOS/Android in the mobile Internet era, every change in the industry platform will produce new models and applications. "Industrial opportunities in the mobile Internet era are at least 10 times larger than in the PC era, at least 100 times larger in the artificial intelligence era than in the PC era, and 10 times or more larger than in the mobile Internet era."
Indeed, the popularity of AI large models does not fade, in the final analysis, it has brought subversive changes to the industry and society. This change is not temporary, but will lead all enterprises into the next era of AI-powered autonomy.
But these are all due to the emergence of GPT, allowing people to see the opportunities that may arise in the next ten or even twenty years. After all, in essence, its astonishment is not in technological innovation, but in the emergence after reaching a certain amount. From the current point of view, the upsurge of large-scale entrepreneurship caused by GPT-4 is a bit too much. It is true that as computer scientist and natural language model expert Wu Jun said, there are not many entrepreneurial opportunities for large models because they consume too much resources.
In the era of mobile Internet, the birth of cloud computing has transformed software from an OP model (On-Premise) to a rentable and portable cloud SaaS model, changing the software delivery method and customer usage habits.
So, in today's era of big models, will there also be new systems or models to subvert SaaS? How will the big model affect the delivery mode and development mode of SaaS? In the next ten or even twenty years, will the large model really become the "operating system" for the digital transformation of enterprises as predicted by Academician Zhang Yaqin?
In the era of large models, the singularity moment of the SaaS industry is approaching.
1. Will MaaS completely subvert SaaS?
In the past two months, major manufacturers have gathered together to release large-scale model products. Last month was even dubbed the "large-scale model release month", and its update speed for a whole month can be recorded in history. Among them, the most disruptive "New Things" is none other than MaaS. Robin Li, founder of Baidu, proposed the MaaS (Model as a Service) model and services at the "Wen Xin Yi Yan" press conference, and said that in the era of large models, new types of cloud computing companies will emerge, and their mainstream business models will also change from IaaS to for MaaS.
Is MaaS really of such great value? Although this term was publicly discussed by the media in China for the first time, as early as 2012, Professor Zou Guobing of the School of Computer Engineering and Science of Shanghai University proposed the concept of "MaaS". In his thesis, he analyzed in detail what is model as a service.
MaaS consists of three parts, including the base layer, the intermediate core layer and the underlying extension layer. The figure below is the MaaS theoretical model proposed by Zou Guobing in 2012.
MaaS Theoretical Model (Zou, 2012)
The basic layer covers the user's identity information, such as user name and occupation and other basic information; the middle core layer describes important user characteristics, such as user interests, preferences, goals, etc.; the bottom extension layer contains personalized knowledge with user characteristics, Including user interests, preferences and personalized model networks obtained through semantic relationship analysis and reasoning.
Finally, a MaaS platform is formed through the base layer to the middle core layer, and then to the bottom extension layer. This platform can be inserted into cloud computing as an independent service platform, and the specific location is between the PaaS layer and the SaaS layer.
From the introduction of the MaaS theory in 2012 to the final implementation of MaaS, the algorithm and computing power have undergone earth-shaking changes in the past eleven years. At the Wenxin Yiyan press conference in March this year, Li Yanhong even boldly predicted that the mainstream business model of cloud computing companies will change, and MaaS will completely subvert SaaS and become the mainstream business model.
As for why MaaS can "subvert" SaaS, the answer was given as early as in Professor Zou Guobing's thesis. He believes that "MaaS is a 'everywhere' model". From the basic composition of MaaS, it can be seen that it can obtain a personalized model network through the personal information collected by customers and cloud vendors at the IaaS layer.
The personalization of MaaS can just make up for the shortcomings of current SaaS. The current situation is that for any two different end users, there is no difference in the SaaS services they get. However, SaaS itself is used to solve the marginalized needs of customers. At present, domestic SaaS is mostly standardized. Even if it is personalized, it can only be delivered in medium and large enterprises. In other words, SaaS cannot solve personalization very well . The problem. But if there is a MaaS platform that can provide different information for different users, then users can get more personalized services.
In the past, if SaaS customers had personalized needs, they needed to use the low-code platform to build personalized functions and services. Moreover, it must be based on the premise that the SaaS company used has its own PaaS, but if MaaS can play a role in it in the future, perhaps when customers use SaaS software, they can directly use their own accumulation on the MaaS platform, that is, the data layer. Personalized configuration.
And how will MaaS achieve the subversion of SaaS? Specifically, we must start with the impact of MaaS on IaaS and PaaS.
First of all, the role of the IaaS layer is to provide computing power services and large-scale data storage centers for the upper-layer PaaS and SaaS. Although MaaS stays between the PaaS and SaaS layers, it will have an impact on the business model of IaaS and even the competitive landscape of cloud vendors.
At present, in China, cloud vendors are highly homogeneous, and price wars are fierce. As a basic server, IaaS rarely provides external services in China, and most of its business models rely on the delivery of IaaS+PaaS and IaaS+SaaS. Compared with SaaS with a gross profit rate of more than 50%, domestic IaaS is only 10-15%. The emergence of large models has opened up a new competitive landscape for cloud vendors.
In the future, a new business model will be "IaaS+MaaS". All MaaS configuration file templates designed by MaaS developers are stored in the IaaS layer, which means that the database of the IaaS layer is used to feed MaaS. This means that the quality of MaaS also depends on the quality of the IaaS layer database, including relational databases, NewSQL, data warehouses and data lakes, etc. for data processing.
At present, not only major Internet companies such as Baidu, Tencent, and Ali have deployed MaaS to seek new incremental markets; some start-up companies are also constantly optimizing the data layer and launching new products to prepare for the arrival of the era of large models.
In the year when Zou Guobing proposed the MaaS theoretical model, the country was still far away from the big model. But now, big manufacturers have entered the market one after another to make large-scale models. The challenge of landing the MaaS layer has been overcome. What remains is how MaaS will be passed to the SaaS layer in the future, what will happen to the landing situation, and whether it will completely subvert SaaS, or subvert to to what extent? None of this is known yet.
In addition to changing the IaaS business model, MaaS has a deeper impact on PaaS.
Domestic leading SaaS companies have developed their own PaaS platforms. But from an objective point of view, it is extremely difficult for a SaaS vendor to do PaaS. Back then, the 10-year-old Beisen started to explore PaaS in 2012, and finally completed the construction of the PaaS platform in 2019. It was officially opened to customers in 2020 to conduct business development on customers' personalized products.
The average development cycle of PaaS is very long. Now the implementation of MaaS will not only change the business model of PaaS+SaaS, but also change the development model of PaaS. On the one hand, SaaS companies that have developed for several years will not be satisfied with small and medium-sized customers, but developing large customers without a PaaS platform is tantamount to a blind man trying to figure out an elephant.
Because big customers need personalized customization, at present, only SaaS on the PaaS platform can meet the personalized needs of big customers. According to the MaaS theoretical model proposed by Zou Guobing, SaaS+MaaS can provide each end user with personalized services for their own business.
On the other hand, MaaS, as the fourth cloud computing architecture, currently only has its imagination in the middle layer between PaaS and SaaS. In other words, the upper layer architecture of PaaS has changed from SaaS to MaaS. Then the development model of PaaS will be affected, and the development cycle will be shortened.
At this stage, major manufacturers are still at the conceptual level for MaaS, and it has not yet been implemented in practice to produce results. But from a theoretical point of view, the changes of MaaS to the business model and development model of IaaS and PaaS will affect all aspects of upper-level SaaS companies, such as the delivery method and development model of SaaS.
In the development mode, some companies will choose to develop SaaS software with a graphical interface on the PaaS platform. But with MaaS, customers can pass requirements directly to the system, and it will automatically invoke functions and display results. The difference is that the SaaS software built through the MaaS platform will display more personalized tools, and its effect may be better than that of the PaaS+SaaS model. Finally, the subscription model of SaaS will also change accordingly.
The disruption brought by MaaS to SaaS goes far beyond this. With the advent of the era of large models, the future will not only be dominated by To B, but also the golden decade of SaaS will come. In the next five years, more unicorns will appear in Chinese SaaS, and in the next ten years, based on the MaaS base, domestic SaaS may even have giants like Salesforce.
2. AI-based SaaS or SaaS-based AI?
Before understanding the impact of AI on SaaS, it is necessary to know whether it is AI-based SaaS or SaaS-based AI. The difference between the two is that the value ratio is different.
If it is the combination of AI and SaaS, AI is used as a technical tool to assist SaaS products, and the final service form is still SaaS, which is AI-based SaaS; the latter is different, SaaS-based AI means that the attributes of SaaS will be very weak, only As a microservice hidden in the product, the value is not obvious. This scenario is not impossible in the future.
In the era of big models, everything will change from the interactive form of SaaS to the development efficiency of engineers, from the gross profit margin of SaaS to the development model, and finally to the delivery method of SaaS.
1. Interactive form
In SaaS in the next five years, natural language interaction will be ubiquitous.
"Display the revenue and net profit of the past 5 years by product and year, generate a chart, and give a summary below the table for special transactions that have had a significant impact." This is the function that financial personnel most want ERP to achieve.
There will be many natural language application programming interfaces in future SaaS products. Users can directly describe their needs in language, and then generate dashboards, export reports, and even AI analysis. For example, in a CRM system, AI can assist in the processing of unstructured information and sort out customer data in chat messages. Let CRM change from a tool for controlling sales to a tool for assisting sales.
With the help of the natural language model, engineers do not need to write custom code, and the application can directly open the data to the user, allowing the user to customize the required functions. Through the API interface, the large language model is connected to the SaaS product, so as to bring a better experience to the user and reduce the custom request of the developer.
Ultimately, users can interact with the software through simple language, which will shorten the learning curve for users to use SaaS and improve the usability of the product. "In the next three years, most unicorns will appear in the SaaS circle, because one user can complete three jobs." This is the expectation from the other side of the Pacific Ocean.
2. Development efficiency
The integration of large models and SaaS will not only improve product ease of use, but also increase the productivity of engineers and developers to a greater extent.
You must know that engineers spend a lot of time on SaaS development. SaaS companies solve complex long-tail requirements, and developers have to write custom code for this. However, the access to large models allows developers to generate output codes through natural language, thereby liberating engineers and allowing them to focus more on product-level issues.
The most anticipated results are that the development cycle is shortened, the development cost is reduced, the software quality is improved, and the product becomes more refined and humanized.
3. Gross profit margin
However, in the next 3 to 5 years, the gross profit margin of SaaS will decline accordingly.
According to the OpenAI GPT-4 charging standard, the fee is $0.03 for every 1000 token statements. Converted into natural language, it is equivalent to 750 English characters, or 400 Chinese characters. In addition, SaaS needs to access large models, and calling APIs also requires costs. And with the increase of users and the improvement of the complexity of AI tasks, its use cost will further increase.
Therefore, in the short term in the future, the gross profit margin of SaaS companies will decline. Companies need to carefully evaluate AI usage scenarios and make trade-offs between controlling costs and improving product usability. But in the long run, the improvement of the ease of use of SaaS products, the acceleration of product problem solving, the satisfaction of individual needs, and a good user experience will bring about further development.
4. Development mode
The emergence of large models has brought two possibilities to the SaaS development model: the first is to change the development model of PaaS, and the second is MaaS+SaaS.
At present, domestic top SaaS companies have developed their own PaaS platforms, serve customers in the way of PaaS+SaaS, and provide personalized SaaS products for customer business. However, the development cycle of PaaS is generally long, and many domestic SaaS companies still dare not try it, and it is difficult to get big customers in the end. Now, the emergence of large models has ignited hope for these SaaS companies.
The first way: change the development model of PaaS
Developers use natural language programming to develop PaaS platforms, or use GPT technology to make programming more intelligent. Although the low-code PaaS platform can reduce development costs and improve efficiency, there are still many limitations in this approach. Not only many customer needs cannot be met, but the product itself also has many problems.
But MaaS is different. If the MaaS platform is built between PaaS and SaaS, the development of PaaS will be based on the MaaS platform instead of directly on the IaaS provided by cloud vendors. The role of the MaaS platform itself is to provide personalized functions, because the data source of MaaS is the data of enterprise customers themselves in the cloud. In this way, it not only improves the development efficiency of PaaS, but also better meets the individual needs of SaaS customers.
The second way: MaaS+SaaS
This approach goes a step further, allowing MaaS to directly replace PaaS. The structure of cloud computing will be adjusted to: IaaS--MaaS--SaaS. The data of IaaS is directly fed to MaaS, and the data used by MaaS is completely based on the customer's own business. The final output of SaaS will be that the product functions and services each customer gets are generated according to their own business, just like Now the algorithm recommendation logic on the C-end is the same, and each user gets different functions, fully meeting the needs of all customers.
5. Delivery method
If the above-mentioned assumptions are realized, half of domestic SaaS can be realized, and the future delivery method will also change from a subscription model to pay-as-you-go. Even, it may become SaaS-based AI, that is, SaaS is hidden in AI products as a microservice, and the attributes of SaaS are extremely low.
The improvement of product usability and the satisfaction of personalized needs, these two points alone are enough to subvert the delivery method of SaaS. Customers can use natural language to call functions that meet business needs, and there is no need for learning costs. The degree of personalization is higher, so SaaS products will show viral growth.
On the other hand, the cost of SaaS accessing large models increases, and the more functions customers call, the SaaS company should also charge more. At that time, pay-as-you-go will be more conducive to the development of SaaS.
This is why, the next ten years will be the golden decade of SaaS, more unicorns will be born in China, and giant companies will be born abroad.
3. Large model, a new generation of "operating system"
In the next ten years, large models will become the "operating system" of the AI era.
Zhang Yaqin, Chair Professor of Tsinghua University and Dean of the Intelligent Industry Research Institute (AIR), said in his speech "AI Large Model Era", " In terms of industry, GPT+ and other large models are the "operating system" of the artificial intelligence era. , refactoring and rewriting the above application."
From 2000 to today, the Internet has transitioned from the PC era to the mobile Internet era for more than 20 years. The current wave of large-scale models is pushing all enterprises to the next era—the era of artificial intelligence.
Each era has its own operating system, and different server architectures and applications will grow on different operating systems. The operating system in the PC era is Windows, the chip architecture is x86, the server is C/S, and the upper layer is a web browser and installed software.
In the era of mobile Internet, the operating system has changed to iOS and Android, and new chip architectures have emerged, namely CPU and GPU. The server is cloud computing, and the upper layer also has apps and applications that grow in the cloud. That is to say, in the later stage of the development of mobile Internet, cloud computing appeared, which led to IaaS, PaaS, and SaaS.
In the future, in the era of artificial intelligence, the underlying operating system will be a large model, and the chip architecture will also revolve around GPU, CPU, and XPU. The upper cloud computing architecture will change from IaaS--PaaS--SaaS to IaaS--vertical model /MaaS/Basic Model -- SaaS.
Every change of industry platform will produce exponential effect. Academician Zhang Yaqin said, "The industrial opportunities in the mobile Internet era are at least 10 times larger than those in the PC era, the artificial intelligence era is at least 100 times larger than the PC era, and 10 times or more larger than the mobile Internet era."
Cloud Computing Architecture Reshaped in the Big Model Era, Source: Institute of Intelligent Industry, Tsinghua University
The picture above shows the reshaped cloud computing architecture in the era of big models. The bottom layer is still the IaaS layer provided by cloud vendors, including computing, storage, network, and database. Biomedical models, protein analysis models, and more. The last is application SaaS on a vertical model.
At present, breakthroughs have been made in the above-mentioned fields, such as the automatic driving model of Baidu Apollo; the domestic Internet hospital "Medicine" and the Intelligent Industry Research Institute of Tsinghua University have also launched the vertical model "MedGPT" for the medical industry; in the field of protein analysis , as early as 2020, Alphafold came out and achieved 98.5% protein analysis.
In recent months, Internet giants, start-up companies, and research institutions have all been busy in the "Hundred Model Competition". The attitudes of each company are very different. Some major Internet companies are more modest, while others are making wild claims that they will "catch up with ChatGPT within a few months". Among them, only a few can survive, and only three or two companies can really bring about industrial changes in the end.
In the new era of artificial intelligence, maybe they are the next "BAT" giants.