In the new round of Internet innovation, large models have become the next competition point, which is not fleeting like the previous Metaverse, but is gradually infiltrating into the application level.

The "2024 China Mobile Internet Semi-Annual Report" was released, showing that in June 2024, the monthly active user base of AIGC apps reached 61.7 million, a year-on-year increase of 653%. Since last year, large models have once again sparked a fierce battle on the Internet. Global tech giants are afraid of missing an important turning point in the era. Whether they are gaming, e-commerce, or social players, they are all striving to keep up with the era of large models.

Some institutions have predicted that by 2025, the global AI market will exceed $6 trillion, with a compound annual growth rate of 30% from 2017 to 2025.

Indeed, the Internet world has been quiet for many years, but thanks to large models, it has finally become lively again. Taobao, Alipay, Douyin, and other top 20 domestic super apps in terms of traffic volume are basically focusing on embedded AI applications, such as intelligent assistants, intelligent search, intelligent shopping guides, and various new gameplays are emerging in an endless stream.

After the loss of innovation ability in the global Internet field, can large models really give new vitality to the Internet? This question deserves serious consideration.

Innovation cognition is degrading, and product development is stagnant

Why have Internet giants failed to incubate new innovative products before the emergence of large models?

In fact, global technology companies are being brutally questioned by this issue. In January last year, the British magazine "Nature" published a paper based on 45 million manuscripts and 3.9 million patents, finding that disruptive technologies are declining globally.

From a corporate perspective, the Internet market has never stopped its research and development progress over the years, and has even continued to strengthen it. However, it seems that the R&D of major factories has lost its meaning. Long-term investment and returns are not proportional, severely depleting the enthusiasm for innovation among capitalists.

This is not unfounded. According to the Shanghai Stock Exchange report, in 2021 alone, the investment in R&D costs of a group of science and technology innovation sector companies represented by Cambricon in China reached 16.7 billion yuan. Unfortunately, the cumulative loss far exceeded this amount. According to statistics bureau data, from 2000 to 2019, corporate investment has exceeded 76% of the total domestic R&D investment, with an annual growth rate of 10%.

Even under the unfavorable external environment in recent years, the domestic R&D growth rate of enterprises can still maintain above 18%. However, it is also an indisputable fact that R&D is difficult to increase revenue. Previously, Baidu's Robin Li publicly stated that Baidu has over 10,000 R&D engineers, and its investment once accounted for 20% of revenue, but the actual performance in return was not ideal.

As a result, giant companies would rather invest in existing projects. Taking Tencent as an example, data shows that Tencent currently has investments in over 800 companies, 160 of which are unicorns valued at over $1 billion. For this reason, there was even a rumor circulating that the investment and intervention of giants had compressed the innovation within the industry.

In addition, the fact that R&D does not yield significant returns has also discouraged large factories from blindly developing new products.

In recent years, the emergence frequency of mini-programs has been much higher than that of independent apps. Previously, Alibaba, Tencent, ByteDance, Baidu, Kuaishou, Meituan, JD.com, and others have successively developed mini-programs, while independent apps have been shut down in large numbers to save trial and error costs. Statistics show that Tencent once shut down more than 40 projects in a year, and ByteDance also shelved Party Island. Up to now, only Internet giants have shut down more than 70 independent products in recent years.

Behind this is closely related to the profitability of the entire Internet industry. According to data from the Ministry of Industry and Information Technology, in the first quarter of this year, the operating costs of China's Internet enterprises above a designated size increased by 5.1% year-on-year, and the total profit reached 27.89 billion yuan, but it decreased by 15.3% year-on-year, with the growth rate of total profit turning from positive to negative.

The emergence of large models is like a ray of dawn shining into the Internet world. Large factories struggling with stagnation have rushed in. It can also be seen from the R&D direction that large models have indeed stimulated the R&D confidence of giants. However, how long can the innovation brought about by large models last?

It is worth noting that it is difficult for the Internet sector, which is currently experiencing innovation fatigue, to produce another phenomenal product or leading technology. After all, having gone through the era of WeChat and Douyin, any ripple can trigger intense competition within the industry. Just as it is now, self-developed chips, big data, cloud computing, artificial intelligence, and other technologies have become the main focus of all giants and even technology startups.

The same plot of homogenization has never disappeared in the Internet world. When AI gameplay can be seen on any app, such innovation is no longer "innovation."

In addition, although the popularity of large models has stirred up some ripples, the problem that Internet giants once worried about most, that R&D and revenue are not proportional, has become even more serious. As global technology focuses on AI, the resulting capital expenditure has also increased. During this period, the financial reports of overseas giants fully revealed the nature of large models burning money.

Some institutions have analyzed that by 2025 and 2026, the training cost of large models will approach $5 billion to $10 billion, among which Meta, Google, and Microsoft may plan to increase the R&D cost of large models to $50 billion.

All signs indicate that the Internet may never have stopped innovating, but the perception of innovation has declined.

The power of large model applications is not as strong as expected

Unlike past innovations, the large model applications collectively provided to users by the Internet have encountered some troubles soon after their launch: Is there a high likelihood that users will need large models? Based on current data, the answer may be more pessimistic than expected.

According to Red Rock Capital, even for ChatGPT, the leader of global large models, its first-month user retention rate is only 56%, with about half of users abandoning it within a month. Similarly, the "2024 China Mobile Internet Semi-Annual Report" also shows that domestic AIGC users are unstable, and the average usage time per person in the AIGC industry has declined by 23.5% year-on-year.

In the final analysis, the infiltration of artificial intelligence into real life is still a "fantasy" of capital.

From a user perspective, the user retention rate of AIGC applications on almost all mainstream apps is lower than that of traditional applications, and engagement is also low. In July, Shell Finance published a survey showing that 52.05% of respondents sometimes use large models at work, 23.97% rarely use them, 20.55% frequently use them, and only 2.05% always use them.

From a corporate perspective, Huawei has a set of forecast data showing that by 2026, the penetration rate of artificial intelligence in enterprises will only reach 20%.

Why is this the case? Technology, cost, practicality, and security are all reasons.

Taking the entertainment industry, where AI is most widely applied, as an example, Jackie Chan's new film "Legend" was released recently. Before its release, AI technology was the biggest gimmick in the film's promotion. It is reported that Bona Film Group used AI to restore Jackie Chan at the age of 27 in the movie, but few audiences bought it.

Data shows that the current Douban rating of "Legend" is 5.4, and it has only grossed over 70 million yuan in more than ten days since its release.

In another major application area, the advertising industry, user feedback has been mixed. According to iResearch Consulting, about half of advertiser enterprises have applied AIGC technology in online marketing activities, with over 90% of them using it for content and creative scenarios. At present, most Internet companies introduce large models into their own products to boost declining advertising revenue.

However, the shortcomings of AIGC have begun to emerge: for example, the production materials are too formulaic, AI effects cause user fatigue, and the well-known issue of AI plagiarism... Previously, an article titled "I Generated an Ad with AI in Five Minutes, but Spent Five Hours Removing the AI Taste" sparked heated discussions on social platforms.

If large models cannot generate a rigid demand effect in users' online lives like social communication and short video entertainment, they will not significantly advance the Internet process. Currently, the primary focus of the Internet sector is to improve the application efficiency of AI implementation.

Capital has also realized this, and investment flows are shifting from the R&D track to the application track. Haitong International Research Report stated that 2024 is expected to be the first year of comprehensive commercial implementation of domestic large models.

Data shows that among nearly 120 global large model investment events this year, large model application enterprises accounted for 69%, more than half. In contrast, AI Infra and general large models only accounted for 16% and 11%, respectively, while large model data services accounted for only 3%. Looking closely at the application areas of large models, AI healthcare, visual/video generation, office assistants, and programming assistants have the densest concentration of funded enterprises, accounting for 15%, 15%, 13%, and 11%, respectively.

In summary, capital is accelerating the popularization of large models in the real world, and it is urgent for large model enterprises to match technology and business needs. Only in this way can the Internet be "saved." Otherwise, the Internet, which has lost its innovative power, will continue to be confused.

The "family fortune" left by the Internet for large models to live off is not much

One issue to note is that at the stage of large models, most of the gameplay is still the same as before, either by continuing to engage in price wars or by eating into their existing traffic "family fortune."

Essentially, the implementation of large models is not much different from the early days of the Internet when companies "circled the land and raced the horses."

In May this year, a group of domestic large model players began officially announcing price cuts. After Ali Group's Tongyi Qianwen's main large model Qwen-Long's API price dropped by 97%, Wenxin's two main large models, ERNIE Speed and ERNIE Lite, became fully free. Subsequently, iFLYTEK also announced that its Xunfei Spark API capabilities would be freely open to the public.

On the side of ByteDance, Doubao took only 30 days from launch to become the top choice. It is reported that the reason Doubao became a "top stream" large model in such a short time is not only because of the support of Douyin, which has a monthly active user base of 794 million, but also because the new round of funding reached 124 million yuan.

Remember the days when domestic Internet giants' most reliable tactic was to throw money around. Today, is the "land-grabbing" approach still applicable?

First, the key reason why large models can only exchange traffic for money at present is technological convergence. Ultimately, it will also be technology that affects user retention. Simply reducing application costs can indeed increase exposure and compete for users in the short term, but in the long run, AI technology services are not takeout delivery or short videos, and relying on money cannot bring a good user experience.

Second, the development of large models is inherently a capital-intensive project with enormous costs. While cash-rich giants may be able to afford price wars, the profitability of large models is still uncertain, and the risks for small companies entering the market cannot be underestimated. This is bound to further reduce the innovation and creativity of the entire industry.

In fact, the price war for large models started overseas first, when OpenAI and Google were the first to announce price cuts. However, in overseas markets, cloud vendors are moving away from traditional service models and using other methods to fill this cost gap. Taking NVIDIA as an example, in May, NVIDIA announced its first-quarter fiscal year 2025 results.

NVIDIA stated that training and inferring AI on NVIDIA CUDA can drive the growth of cloud rental revenue. For every $1 spent on NVIDIA AI infrastructure, cloud service providers have the opportunity to generate $5 in GPU hosting revenue over four years. Whether China can quickly follow suit with this plan is still up for debate.

Of course, besides the tactics that can continue to be "inherited," there is not much "family fortune" left for large models by the Internet over the years. Even globally, beyond funding, the most critical information and data required by large models are already in short supply.

Similarweb data shows that since May 2023, when ChatGPT reached a peak of 1.8 billion global visits, its traffic growth has gradually slowed down. In response, OpenAI decided to relax restrictions on ChatGPT, allowing users to access it without registration for a time.

Unfortunately, this is also one of the current dilemmas facing the development of large models: the existing amount of Internet information is difficult to support the training of so many large models.

During this period, Bytedance and a group of online office companies "feeding" large models have caused dissatisfaction among many users. Public data shows that GPT-4's training involved up to 12 trillion tokens of data. In the future, models like GPT-5 may require 60 trillion to 100 trillion tokens.

According to Epoch Research Institute's predictions, there is a 50% chance that the demand for high-quality data will exceed supply by mid-2024, and this likelihood increases to 90% by 2026. The risk of this data shortage will persist until 2028.

As for how to make up for this significant data gap, the Internet, whose penetration rate is gradually approaching its ceiling, has not found a better solution for the time being.

Dao Zong You Li, formerly known as Waidaidao, is a new media outlet in the Internet and technology circle. WeChat official account with the same name: Dao Zong You Li (daotmt). This article is original and any form of reprinting without retaining the author's relevant information is prohibited.