"This is obviously false. Large models are both an aggressive offensive and a necessary defense. Anyway, I haven't heard of any reduction in model priority."

"The boss made it clear that we will never withdraw from the table."

In response to rumors that Baidu will likely abandon its general-purpose foundational large model, two Baidu employees told Cityscape.

The rumored originated from a post: "Baidu is likely to abandon the development of its general-purpose foundational large model (another case of 'getting up early but arriving late'?). Li Yanhong recently clarified the strategy of focusing on applications internally, and Baidu is exhausted at the model layer."

The post gained attention after a brief period of circulation, ultimately forcing Baidu to respond. On September 9, Zhang Quanwen, head of marketing at Wenxin Yiyan, posted on WeChat Moments, "The rumor that we are 'abandoning the development of general-purpose large models' is purely false! Wenxin Yiyan has just completed a comprehensive upgrade of its functions. We will continue to increase investment in R&D in the field of general-purpose large models."

Afterward, an informed source close to Baidu told Cityscape that this sparked widespread discussion within Baidu.

A Baidu employee candidly told Cityscape that it was not surprising to hear such rumors: "In the SuperCLUE overall rankings (August 2024), ERNIE-4.0-Turbo-8K's score has fallen behind many other models. If we shift to the application domain, opportunities will be even slimmer."

Over the past six months, market attention on AI large models has shifted from the model technology itself to the practical implementation of intelligent applications, and AI expectations have also begun to adjust. Therefore, making trade-offs between investments in foundational large models and the implementation of AI applications has become an unavoidable challenge for all large model players.

01 The Path to the Birth of the "Rumor" from an ROI Perspective

"Or let me put it this way, training the next generation of models is not Baidu's top priority," said the aforementioned poster, implying that "abandonment" is not an accurate term.

His previous logic for judging Baidu's large model direction was:

GPT-4 has not yet been caught up with, and a model at the GPT-5 level could cost up to $3 billion (rough estimate, with some fluctuation, which is almost Baidu's annual net profit). And you have to know how to do it, and consider whether the ROI (return on investment) will be positive. A failure would be a huge waste of resources and an opportunity cost loss, which is already a very real problem. A company of Baidu's size can no longer afford the dream of models.

From an ROI perspective, looking solely at the output side, i.e., the monetization of large models, Baidu is indeed facing certain commercialization pressures at present.

According to the second-quarter financial report, Baidu's largest source of revenue, online advertising, generated 19.2 billion yuan, a year-on-year decrease of 2%. The reason for this performance in the advertising business is that Baidu's push for AI innovation in its search business has caused considerable internal pressure. Currently, 18% of Baidu's search results are provided by generative AI, which has yet to be monetized.

An industry insider told Cityscape, "Baidu's investment in AI technology is paving the way for the future, but the content generated by AI has not yet found a suitable monetization model, which has somewhat affected advertisers' decision-making."

In the cloud business, Baidu Cloud generated 5.1 billion yuan in revenue for the quarter, of which 9% came from external customers' demand for large models and generative AI-related services. According to calculations, generative AI contributed approximately 459 million yuan to Baidu Cloud's revenue during the quarter, compared to 324 million yuan and 274 million yuan in the previous two quarters, respectively.

While revenue from generative AI is showing an upward trend, its contribution to overall performance remains limited and is not yet sufficient to drive Baidu Cloud's overall growth curve.

The commercialization pressure faced by large companies like Baidu in monetizing large models also comes from the price war in large models initiated actively or passively by domestic cloud vendors.

Recently, Tencent launched its new generation of large language model, "Hunyuan Turbo," further intensifying the price war by reducing prices by 50%. Prior to this, Baidu had also been continuously reducing the inference costs of its models to cope with market price pressures.

In May this year, Baidu Intelligent Cloud announced that its pre-installed services for the ERNIE-Speed, ERNIE-Lite, and ERNIE-Tiny series of models would be freely accessible to customers. During WAIC in July, Baidu Intelligent Cloud announced significant price reductions for its two flagship models, ERNIE 4.0 and ERNIE 3.5, with ERNIE 4.0 Turbo fully open to enterprise customers at prices as low as 0.03 yuan/1,000 tokens for input and 0.06 yuan/1,000 tokens for output.

With the fierce competition in the domestic large model market, the price war has led to a sharp compression in the invocation costs of large models, and the industry has fallen into the so-called "negative gross margin era." Both open-source and closed-source models currently face the same dilemma - the large model business is difficult to achieve direct profitability.

According to Caijing magazine, "The elimination round in the large model market has already begun." An executive in charge of a large model business at a Chinese cloud vendor analyzed that the negative gross margin of large model invocations means that in the short term, the more invocations, the greater the losses for cloud vendors.

Taking these factors into consideration, "it is somewhat unrealistic to achieve a positive ROI solely through the current level of commercial development of large models. Therefore, it is entirely possible that Baidu will redefine its strategy," said a Tencent AI algorithm expert to Cityscape.

02 Baidu's AI Path is "Both and Both"

Although Baidu denied the rumor that it would "abandon the development of general-purpose large models," it did not deny the priority given to application development. Compared to other large companies, Baidu attaches great importance to AI application development.

In the speeches of its executives, they have consistently emphasized that "rolling out large models is meaningless, and there are greater opportunities in rolling out applications." At the recent WAIC, Li Yanhong even stopped emphasizing whether models should be open-source or closed-source, "None of these are important. Basic models without application support are worthless, whether open-source or closed-source."

An industry insider close to Baidu told Cityscape, "Li Yanhong's intention is to encourage entrepreneurs to develop more AI applications based on the Wenxin large model, rather than wasting resources on too many foundational large models."

"The idea is right, but the question is how to implement it. Baidu itself is engaged in foundational large models, so why should others listen to them?" he added.

This issue has also been discussed among Baidu employees. Some employees have expressed the hope that Robin (Li Yanhong) would refrain from making public statements that are prone to ridicule. "Only Baidu's large models are not a waste; other companies doing the same are wasting resources. As an industry leader, this attitude is not broad-minded enough," they said.

From the current market perspective, Baidu's ideal goal in both large models and application fields is actually a two-pronged approach.

In the field of applications, which has always been emphasized, Baidu has launched three AI development tools: AgentBuilder (for intelligent agent development), AppBuilder (for native AI application development), and ModelBuilder (for model customization).

Intelligent agents are the most promising direction for AI applications at Baidu. At the 2024 Baidu AI Developer Conference, Li Yanhong revealed that 10,000 Baidu customers have successfully launched intelligent agents, covering over 30 industries such as education and training, real estate and home furnishing, machinery and equipment, and business services.

In a recent internal sharing session, Li Yanhong reiterated the importance of intelligent agents, arguing that their low threshold is a direct approach to application development.

However, there may not be a consensus within the company on the direction of intelligent agents. According to sources close to Baidu, some employees privately expressed dissatisfaction during the Baidu AI Developer Conference.

Some employees said that the so-called "intelligent agents" showcased on the Wenxin Intelligent Agent platform are mostly like toys, with only a handful of projects used for demonstration. These "toys" have little practical value, even when distributed through search. Truly meaningful intelligent agents should provide high-quality services or access to private data that is not available on the public network. Calling these "toys" AI native applications would be self-deception.

Li Yanhong may also recognize the shortcomings of intelligent agents. "Many people are optimistic about the direction of intelligent agents, but so far, they are not yet a consensus. Few companies, like Baidu, consider intelligent agents as the most important strategy and development direction for large models," he said.

In addition to the product attributes of intelligent agents themselves, external developers also maintain a certain degree of caution towards Baidu in AI application development.

Judging from currently public cases, most users of Baidu's AI development platform are merchants providing lightweight applications for their consumers, while enterprise-level applications are relatively scarce. Although Li Yanhong has repeatedly promised that Baidu will not get involved in application development, leaving enough space for entrepreneurs, the prevalent "pixel-level copying" in the mobile internet era has made developers more vigilant in the era of large models.

The founder of a vertical model startup once told Caixin that developing on the Baidu platform means exposing algorithms and data to Baidu, while Baidu may not be able to provide sufficient interface support for mid-tier developers, which also means that developers' customer data is difficult to deploy on the Baidu platform.

03 Strategic Shift Amid the Bubble of Large Models

Now that the hype has subsided, more people are starting to examine the actual benefits brought about by generative AI.

At the end of June, Goldman Sachs released a report titled "Generative AI: Much Spent, Little Gained," in which Senior Strategist Allison Nathan raised a crucial question: Are we investing too much in AI while gaining too little in return?

The report mentioned that generative AI technology is seen as transformative for companies, industries, and societies, prompting many large companies to plan to invest $1 trillion in AI-related items over the next few years, such as data centers, chips, and power grids.

This expenditure has been validated. A CoutueAI report also mentioned that infrastructure construction is expected to cost about $1.2 trillion by 2030, primarily for the procurement of approximately 25 million GPUs and other related expenses.

To achieve a positive ROI, with a 25% return rate, the expected revenue needs to reach $600 billion. With an investment of $1.2 trillion, total AI revenue must reach $1.8 trillion by 2030 to achieve a breakeven point. However, according to Goldman Sachs, so far, these huge investments have not yielded significant results beyond marginally improving developer productivity. Even NVIDIA, which has benefited the most, has seen its share price decline.

Of course, this is a macro-level discussion. If we focus on specific large companies, the practical problem they face is that general-purpose large models are gradually evolving into a "super money-burning" game, becoming a "rich man's game" that only giants can participate in. Companies without sufficient funds will have to withdraw from the competition.

According to information published on the official website of the State Internet Information Office, since the first batch of AI large models registered in China was released on August 31 last year, a total of 188 large models have been registered nationwide in just one year.

From a terminal thinking perspective, as underfunded players are eliminated or exhaust their resources in competition, only a few truly competitive large model companies will ultimately dominate the market.

"Essentially, foundational large models are like CPUs, both of which are general-purpose technologies. As long as a few companies become strong enough, they can meet most market demands. This is how competition in general-purpose commodities works - the winner takes all, and only a few vendors will survive and thrive through the baptism of market competition," said a computer industry insider.

This "money-burning war" has played out multiple times among internet giants. Taking the price war for large model invocations among cloud vendors as an example, cloud vendors are betting that as prices for large model invocations drop by 90%, the number of invocations will grow exponentially over the next 1-2 years. In this game, a group of model vendors are bound to collapse due to the price war, while survivors will take over the market and clean up the mess.

Obviously, for large companies to win the "money war" in foundational large models, in addition to having sufficient financial support at present, their ability to generate revenue in the long run is crucial, and AI applications are the key direction to achieve this goal.

In fact, a close examination of the moves made by major vendors this year reveals that while they continue to invest in foundational large models, they are also constantly emphasizing the development of AI applications. AI applications are not only the future revenue-generating tool but also one of the critical paths to ground large models and create real value.

Therefore, "Baidu's dual-track strategy (simultaneously advancing both large models and AI applications) is a prudent approach to address the potential bubble in large models," said a source close to Baidu to Cityscape.

"The only problem is that the direction for the implementation of AI applications remains unclear. Many projects are still in the proof-of-concept stage, and there is still a considerable distance from true commercialization," the source added.