When AI has not yet reached the stage of being a productive tool, no one knows what kind of basic model will be the ultimate solution, or how paying customers can convert their investments into real revenue.

Regarding what can be profitable from megamodels, no domestic company has expressed an opinion so far. Or, rather, this is not a topic that can be openly discussed at this time.

Author | Si Hang

Editor | Pi Ye

Produced by | Industry Home

"This year, our focus is on AI applications, but valuations cannot be too high," an early-stage investment fund told Industry Home.

A year ago, megamodels were still in a state of "prosperity," whether it was the queuing up for financing by megamodel superstar companies, big names diving into entrepreneurship themselves, or cloud vendors scrambling to secure AIGC positions. However, 19 months later, there seem to be some unusual signals hidden behind this prosperity.

According to the New York Times, as of August this year, OpenAI's revenue surpassed the $300 million mark, representing a year-over-year growth of 1700%, and it is expected to reach $3.7 billion in 2024. By next year, OpenAI's revenue could even exceed $10 billion. Immediately following the disclosure of this financial document, OpenAI secured its largest funding round of the year, raising $6.6 billion and pushing its valuation to $157 billion.

However, behind these capital-focused figures, there is another set of numbers on the other side of the coin: In 2024, OpenAI's losses amounted to $5 billion, including operating and administrative costs, excluding equity compensation; according to The Information, these losses are projected to continue rising to $14 billion by 2026.

The high costs of megamodels make it difficult for even top AI companies with 180 million individual users and 1 million enterprise subscribers worldwide to talk about profitability in the short term.

Looking domestically, it is visibly apparent that the megamodel market is "cooling off." Behind the two core themes of "pre-training" and "AI applications," different AI megamodel companies are making strategic choices for 2025 and even the upcoming Q4: continue burning money or pursue commercial validation?

In the long run, this is not a paradox. However, the current reality is that when AI has not yet reached the stage of being a productive tool, no one knows what kind of basic model will be the ultimate solution, or how paying customers can convert their investments into real revenue.

At a recent media communication meeting of ZeroOne, one of the "Six Little Tigers" of AI, Kai-Fu Lee explicitly stated that they will continue to invest in pre-trained megamodels and further strengthen the capabilities of the foundational megamodel on top of existing investments to build a sustainable profit model for megamodels going forward.

But how long will it take for validation? Will open-source or closed-source model algorithms ultimately be paid for by customers? When will their PMF (Product-Market Fit) be validated? There are no clear answers to these questions.

From a broader perspective, several questions that need to be addressed are: What is the current survival status of China's megamodel companies? How much room for "burning money" is left for these companies? More importantly, after a year-long megamodel battle and nearly six months of price wars, how many players are still committed to this path?

I. From Burning Money to Making Money: How Big is the Gap for Megamodels to Fill?

The "earnings season" in August brought good news for several cloud vendors.

For instance, Alibaba Cloud reported revenue of RMB 26.549 billion in the second quarter, a year-over-year growth of 6%, while Baidu Intelligent Cloud saw a 14% year-over-year growth in the same period. Notably, all cloud vendor earnings reports emphasized the growth driven by AI, which contributed not only to cloud business but also to some C-end businesses.

Similarly, after deep integration with megamodels, China's cloud market expenditure also increased to a certain extent. According to Canalys, China's total cloud infrastructure spending reached $9.4 billion in the second quarter of 2024, a year-over-year growth of 8%.

Behind these positive announcements, what is the true state of megamodels?

From the perspective of customers, their attitude towards megamodels is undoubtedly positive, but they need further exploration on how to integrate megamodels into their own businesses. According to an IBM survey of 3,000 CEOs globally, 75% of them are actively embracing generative AI.

When it comes to specific megamodel projects, the author has analyzed the bidding results of cloud vendors and megamodel startups over the past two years in the article "Behind 190 Megamodels: Uncovering the Truth of China's Megamodel Industry Implementation After 600 Days." Overall, government and enterprise customers have a higher willingness to invest in megamodel development, and the bid amounts this year are more substantial compared to last year.

These are all B2B deals for megamodels. Although currently, the B-end is the main source of profitability for domestic megamodels; however, it is worth noting that besides private deployment of megamodels, another layer of imagination for megamodel vendors lies in the traffic end, i.e., the C2C business.

For the C-end, especially AIGC-type applications with strong user perception in the market, China generally adopts a free business model at present.

The reason for the "free" business model is that there is currently no true killer application in China. Before providing users with a sufficiently strong product experience, a charging strategy is difficult to implement and may even backfire. After all, for C2C applications, having traffic equals half the battle won.

However, it is undeniable that in this direction, the market generally adopts a traffic-driven approach. Both cloud vendors and startups are adopting various strategies to attract end-users. For example, in September this year, Baidu's Wenxin Yiyan officially upgraded to "Wenxiaoyan," intensifying its AI search capabilities; recently, the Dark Side of the Moon also launched the Kimi search version, and according to internal sources from the company, it has been training in the field of text-to-video generation since early this year. Meanwhile, companies are also offering substantial user subsidies on platforms such as Bilibili and Xiaohongshu.

In reality, although the overarching goal of text-to-video generation is to attract end-users, different vendors adopt different strategies based on their positioning, leading to varied results.

For instance, companies led by Zhipu AI embed text-to-video generation into their own AIGC applications, which are mostly seen in scenarios where the AIGC application itself has good traffic and can further enhance its influence through text-to-video. In contrast, other megamodel companies, such as MiniMax's Hailuo AI, treat text-to-video as a standalone application, hoping to create a killer application through the powerful features of their product.

Whether it's AI search or text-to-video generation, which megamodel vendors have been "convoluting" over the past six months, both correspond to continued investment in megamodel pre-training. However, for these vendors training AIGC applications and foundational megamodels, they bear significant costs.

Data shows that the daily operating cost of ChatGPT can reach $700,000, which involves not only card costs but also energy costs and other training costs. For example, GPT-3, with 175 billion parameters, consumes 1,287 MWh of energy, while GPT-4, with 280 billion parameters, consumes up to 1,750 MWh. In other words, each response from GPT-3 consumes 0.0003 kWh, while GPT-4 consumes 0.0005 kWh per response.

"For megamodel startups, their ability to continue training models is most tested by their financing capabilities," Feng Bo, managing partner of Changlei Capital, told Industry Home.

Under such continuous capital burning, even dollar funds with a more positive attitude towards megamodels can easily find themselves stretched thin. For example, Inflection AI, which raised $1.5 billion in funding, shut down its operations with zero revenue; Stability AI also began mass layoffs, with the CEO even leading the exodus; even unicorn Character AI abandoned its self-built AI model.

According to PitchBook, investors have "selected" 26,000 AI and machine learning projects over the past three years, investing a total of $330 billion.

In terms of return on investment, for domestic C2C free AIGC applications, it is difficult to generate revenue in the short term; meanwhile, Meta has also stated that it is prepared for losses in the coming years but will continue to invest.

Regarding recent reports that "at least two of the 'Six Little Tigers' of megamodels are abandoning megamodel training," Feng Bo believes that "on the one hand, this is a severe test of financing capabilities; on the other hand, it is highly likely that further training will not yield fundamentally different results, so it may be appropriate to pause and avoid senseless and costly continuous training."

II. Megamodels at a Crossroads

What happens when costs and benefits are extremely unbalanced?

It is certain that the imagination of megamodels goes far beyond what we see today. However, within different vendors, the roles played by megamodels vary significantly.

Firstly, for cloud vendors, whether for cloud services, including their catalytic effect on public clouds, or for other SaaS or even C-end businesses, megamodels, as "money-gulping beasts," primarily serve as levers to drive revenue from other businesses through AI.

In addition to the cloud businesses mentioned above, such as Baidu Intelligent Cloud, Alibaba Cloud, Huawei Cloud, and Tencent Cloud, AI is driving growth in two core areas: cloud services and upper-layer applications.

Over the years, cloud vendors have primarily extended their imagination based on the three-tier architecture of IaaS, PaaS, and SaaS. After announcing their integration, cloud vendors' commercial value is more evident in the underlying resource layer and the intermediate PaaS layer, commonly referred to as basic cloud products, database containers, and other middleware products, which constitute the bulk of their revenue.

However, with the emergence of AI, the existing IT architecture is being restructured. For instance, CPU products are now being upgraded to "CPU+GPU" products, and the mode and efficiency of computing and storage capacity invocation are changing. Furthermore, database layers now require more vector retrieval capabilities, creating new enterprise demands.

This is precisely one of the reasons why major cloud computing companies are now all-in on AI – they need to strategically reconstruct their core market products based on AI and optimize service products and systems beyond their own businesses.

Beyond AI strategies, peripheral businesses such as search and collaborative office tools are also being mobilized. Over the past year, collaborative office vendors such as DingTalk, Feishu, and WeChat Work have all integrated AI functions, exemplified by DingTalk's "Magic Wand" and Feishu's "My AI."

At Feishu's September 2024 press conference, the company released data showing that its ARR (Annual Recurring Revenue) reached $200 million in 2023 and is expected to exceed $300 million in 2024. Xie Xin, CEO of Feishu, stated that although the company is still operating at a loss, it is narrowing significantly.

Apart from leveraging megamodels as a lever for B2B products, their most noticeable and impactful presence for end-users is in AIGC applications, such as Wenxiaoyan, Doubao, Tencent Yuanbao, and Tongyi Qianwen.

To gain a foothold in the market for their AIGC applications, these vendors have not only chased after ChatGPT but have also added functions like AI search and text-to-video generation, as mentioned earlier, to enhance their influence. After all, for the C2C market, a true "killer application" is most likely to generate commercial value.

Despite a thriving ecosystem, lower prices, and increasing user numbers, the cost of megamodels, which are "money-gulping beasts," remains a significant challenge for both cloud vendors and megamodel startups.

If well-established cloud vendors with years of experience in cloud services need to pull out all the stops to address megamodels, megamodel startups, which rely on financing, face even more challenges.

Unlike cloud vendors, startups cannot leverage megamodels as "levers" to drive other businesses. This means they can only monetize and validate their models within the narrow lanes of model computing and commercialization. However, as domestic megamodel price wars intensify, a range of solutions and models, from tokens to upper-layer data governance and RAG, are becoming harder to sell to enterprises, making it more difficult for them to generate corresponding commercial returns.

Nevertheless, the long-term layout continues. For instance, the recent launch of the Kimi search version signals the company's aspiration to secure a position in future AI search, which could significantly enhance its advertising and transactional value in the market. In Feng Bo's view, "By reaching as many end-users as possible, these companies with software, applications, and traffic have the potential to evolve into new internet companies, at which point commercialization will no longer be an issue."

In the short term, though, megamodel startups need to focus on both B2B and C2C to survive. Apart from continuously exploring AIGC-type applications, startups led by Zhipu have embarked on an aggressive expansion into the B2B market. As evidenced by the overall bidding results in 2024, Zhipu has won the most bids among the "Six Little Tigers" of megamodels, although its average bid amount is lower than that of large players like telecom operators and Baidu Intelligent Cloud.

Feng Bo told Industry Home that, on the one hand, the biggest demand for megamodels comes from government and enterprise customers, "who naturally prefer telecom operators with state-owned backgrounds. On the other hand, brand reputation is crucial for these big customers when selecting service providers."

From another perspective on the survival situation of megamodel startups, it is worth noting that in Zhipu's latest round of financing, the lead investor was Zhongguancun Science City, a market-oriented investment platform established by the Haidian District Government to promote technological innovation. Moreover, three state-owned enterprises participated in the funding round. Notably, Zhipu recently co-founded a fund with two other companies.

Examining Zhipu's investment portfolio reveals that it covers companies across the entire megamodel industry chain, from infrastructure firms like Xingyun Integrated Circuits, Wuwen Xinqiong, and Jiliu Technology, to model-layer companies like Mianbi Zhineng and Shengshu Technology, and finally to upper-layer application companies like Milelu Zhineng for the legal field and Muyan Zhineng, a startup founded by former Miaoya Product Manager Zhang Yueguang.

Unlike cloud vendors, as a megamodel startup, Zhipu's smarter approach to making a name for itself in the B2B market is to establish its own ecosystem.

The fact that megamodel companies are personally investing as LPs sends another signal: while accelerating the construction of their ecosystems, they are also seeking new profit models.

III. Beyond the Burning Money: Where Will China's AI Megamodel Companies Go Next?

"For early-stage funds like ours, the focus will be on AI applications," Feng Bo, managing partner of Changlei Capital, told Industry Home. On the one hand, AI applications can generate revenue in the short term; on the other hand, unlike megamodel companies, AI applications do not command high valuations.

In fact, AI applications and other agent intelligence companies are not the only ones focusing on this area. Today, AI applications have become one of the top priorities for megamodel companies. Fu Sheng has also publicly stated, "I can't imagine how megamodels can make money, but putting a megamodel shell on something can work – as long as there are users, there's money to be made."

Moreover, Kai-Fu Lee, CEO of ZeroOne, one of the "Six Little Tigers" of megamodels, has also publicly expressed his stance on AI applications:

The pursuit of AGI and the implementation of model capabilities are not contradictory, and can even be said to complement each other. From an industry perspective, only the prosperity of the application layer can guide the entire ecosystem towards a virtuous cycle; from the company's own perspective, successful applications can bring stable operating cash flow and become a commercial foundation to support AGI exploration.

However, is there a conflict between the development of foundational large models and AI applications? On this issue, the views of cloud vendors and large model startups may also differ.

From a practical perspective, firstly, it can be seen that cloud vendors have already begun to deploy upper-layer applications internally. For example, in September this year, Baidu Intelligent Cloud upgraded its intelligent customer service solution, Baidu Keque, for intelligent customer service scenarios; another example is JD.com's AI digital human for e-commerce scenarios; and Huawei Cloud's CodeArts Pangu Assistant, launched to improve programmer productivity.

Although such applications may re-enter the SaaS battlefield, they may differ in terms of product competitiveness, pricing models, and target audiences when advanced from an AI perspective.

Secondly, for large model startups, they have also begun to focus on commercialization issues and have prioritized AI applications. For example, recently ZeroOne Everything also announced its toB business plan, which includes AI digital humans for retail and e-commerce scenarios.

From the perspective of winning bids, the conclusion is even clearer. An analysis of winning bid projects previously covered by Chanyejia shows that although the winning bid amounts have been larger and more scaled this year, a closer look at the types of projects won by startups and cloud vendors reveals that cloud vendors have begun to secure orders for AI applications, such as intelligent customer service and digital humans.

In contrast, for Zhipu, the only large model startup known to have won large-scale bids, the orders they have received are now more focused on pre-training for large models.

The reason for this is that the services provided by cloud vendors are more systematic and "one-stop," with advantages over large model startups in terms of cloud services and AI infrastructure at the bottom level of large models, model training services, and upper-layer applications.

However, building an ecosystem takes time, which does not mean that large model startups have no advantages at all. Different genes and positioning naturally determine their future development paths. As a large model startup, it can either become a part of the cloud vendor's ecosystem or build its own ecosystem, and the two are not conflicting.

In addition to rethinking their profit models, large model startups also need to enhance their toB service capabilities. Unlike cloud vendors' extensive industry know-how accumulated from serving B-end customers in the past, large model startups are still in the early stages of exploring toB opportunities.

Looking back at the cloud computing era, there are similarities between the current large model era and the past in terms of the competitive landscape.

In Feng Bo's view, large model startups are facing competition from two fronts: cloud vendors and traditional internet companies such as iFLYTEK and SenseTime; and IT vendors engaged in industry digitization, who, although lacking large models, can deliver solutions to customers using open-source model training based on their years of industry know-how.

Therefore, it can be seen that in addition to profitability challenges, large model companies also face formidable competitors.

According to OpenAI's financial report disclosed by The New York Times, ChatGPT currently has 10 million users paying a monthly subscription fee of $20, which translates to a monthly subscription revenue of $200 million. OpenAI has also announced a $2 price increase by the end of the year, with the subscription fee expected to rise to $44 over the next five years. With this financial model, OpenAI predicts profitability by 2029.

As for what specifically large models in China can profit from, no domestic company has yet to make a public statement. In other words, this is not a question that can be openly discussed at present.