In 2025, are you no longer afraid of losing money in the “big model price war”?

In 2025, competition in the field of large models is becoming increasingly fierce, and price wars have once again become an important means for companies to compete for market share. This article will deeply explore the current situation and logic behind the large model price war, and analyze why companies still dare to reduce prices to compete when facing the risk of losses.

In the past year, as the top trend in the technology circle, the big model has developed extremely rapidly from its birth to its implementation. As expected, after its application, it has started the price war that the Internet is best at. According to incomplete statistics, the companies involved in the big model price war include ByteDance, Alibaba, Baidu, Tencent, iFlytek...

At the end of 2024, Alibaba once again announced a price reduction for large models, with the price reduction even exceeding 80%.

To be honest, the initial big model price war did create a lot of traffic for the company in a short period of time. Baidu revealed in August last year that the average daily API call times of Baidu Wenxin Big Model were 200 million in May, which increased to 600 million in August; the average daily Token consumption was 250 billion in May, which increased to 1 trillion in August.

ByteDance's Doubao also used more than 500 billion tokens per day in July, a 22-fold increase compared to May. But like all industries, a price war that lasts too long will inevitably backfire on corporate profits. Data shows that before May last year, the gross profit margin of domestic large-model inference computing power was higher than 60%. After major manufacturers cut prices in May, the gross profit margin of inference computing power fell to negative.

This is a scary number, but for some reason in 2025, the price war for large models started again.

1. The confidence of the "price war" has increased a little

In the new year, the price war of large models has become more intense. In addition to Alibaba, ByteDance and Dark Side of the Moon have also joined the new round of price cuts. In 2025, can large model companies that once spent money like water and suffered losses to the point of life and death be able to afford price wars again?

First of all, it is certain that large-scale model enterprises have ushered in some considerable changes in actual implementation, financing and profitability. Since the second half of last year, the domestic large-scale model implementation map has basically had a vague outline, and information processing, customer service sales, hardware terminals, AI tools, learning and education... together spell out a bright future for large-scale models.

From January to November 2024, the inventory results of large model-related winning projects showed that there were 728 domestic large model winning projects with a total winning amount of 1.71 billion yuan, which were 3.6 times and 2.6 times the full-year data in 2023, respectively. Affected by this, large model companies have also begun to generate revenue in this regard.

Baidu data shows that in the third quarter of last year, Baidu Smart Cloud's revenue reached 4.9 billion yuan, a year-on-year increase of 11%, and the proportion of AI-related revenue continued to increase to over 11%. Coincidentally, Alibaba Cloud's quarterly revenue increased to 26.549 billion yuan, a year-on-year increase of 6%. Among them, AI-related product revenue achieved triple-digit growth.

Secondly, after two years of declining financing, the large model field finally recovered in 2024. Data shows that in the first nine months of 2024, the AI ​​field completed a total financing amount of 37.15 billion yuan, more than doubled compared with the same period in 2023. In short, after the application is implemented and small-scale performance revenue is achieved and capital is favored again, money is the confidence to continue the price war in the large model track.

But can the current profitability of large-scale model companies really support a series of disorderly and chaotic price wars?

To this day, the operating costs and subsequent losses of large models remain high. Overseas leading large-model companies such as OpenAI will have operating costs of more than $8.5 billion in 2024, with an estimated loss of about $5 billion, and a total loss of $44 billion from 2023 to 2028.

As for the cost of model training, OpenAI predicts that it will reach as high as $9.5 billion by 2026.

Although China has made some achievements in the large model industry, compared with the average development speed of large models worldwide, it is too early to stop or reduce training and research and development. The Center for Basic Model Research at Stanford University released a ranking in September last year. The top ten model manufacturers include the Claude 3.5 series of AI startup Anthropic, the Llama 3.1 series of Meta, the GPT-4 series of OpenAI, and the Gemini 1.5 series of Google.

Currently, only Alibaba's Tongyi Qianwen 2 Instruct has entered the top ten of China's big models. Currently, the number of AI big models in the world exceeds 1,328. In the future, the investment capital in the domestic big model track will only increase. In 2024, although the overall financing amount of the entire artificial intelligence industry has more than doubled year-on-year, the number of financing transactions has only increased by about 10%.

In other words, the big model track has reached a brutal elimination period. As of November 2024, a total of 309 generative AI products in my country have completed filing. The top companies eat meat and drink soup, while small companies not only cannot get investment, but even have problems with survival. In order to survive, they either have to fight price wars or roll up marketing.

However, the leading companies that are determined to occupy the market in advance are willing to continue price wars while spending money frantically on marketing.

Data shows that Doubao's new round of large-scale advertising spending soared to 124 million yuan in early June last year. Kimi's advertising spending in the first 20 days of October was as high as 110 million yuan. In 2025, the gradually mature large model will certainly have a little more confidence, but the road ahead is long and there are countless places where money needs to be spent.

2. “Computing power” is the primary productive force

In 2024, big model companies used continuous price wars and overwhelming advertising to popularize various big model products in the real world. However, with the increasing number of users, accidents that caused service crashes due to computing power resources once again made the entire big model track fall into contemplation.

Incomplete statistics show that last year, Kimi, Wenxinyiyan, ChatGPT, etc. all had problems with normal use. ChatGPT even temporarily suspended new user registration due to excessive demand. In China, when thesis season comes, Kimi and other text processing products are "paralyzed".

How important is computing power to the development of big models? Computing power, algorithms, and data were once considered the "three horses" of big model technology. In the past two years, the innovation of algorithms has kept the demand for computing power in a state of high growth. Comparing GPT-3 with the latest LLaMA3-405B, although the model size has only increased by 2.3 times, the required computing power has increased by 116 times.

Therefore, computing power has gradually become the primary productive force in the large-scale model track, and the layout of computing power by the world's leading large-scale model companies has already begun.

It is reported that the giant data center project of OpenAI and Microsoft is expected to cost more than 115 billion US dollars and be equipped with millions of GPUs. However, OpenAI seems not satisfied and has reached a cooperation with Oracle to build a data center in Texas that can accommodate hundreds of thousands of NVIDIA GPUs in the future; Meta plans to reserve 350,000 NVIDIA H100 GPUs, and the computing power reserve will reach 600,000 in the future.

The demand for computing power in China has also exploded. On the one hand, user experience requires computing power resources to support it. On the other hand, the products of major companies tend to be homogenized, and there is no certain difference in technology, so they can only raise prices again and again. Computing power may be the key to breaking through in the future.

Some institutions have predicted that by 2030, 100% of the domestic inference needs will need to be met by hyperscale data centers. The global big model track has set off a wave of intelligent computing center fever. As of the first half of 2024, there are more than 250 intelligent computing centers that have been built or are under construction in China. In the first half of 2024, there were 791 bidding events related to intelligent computing centers, a year-on-year increase of 407.1%.

But there is one point in the current domestic computing power supply that cannot be ignored: chips.

Data shows that Nvidia has an 80% market share in the domestic AI training chip market. This is undoubtedly a deadlock that must be changed before the computing power supply chain is formed. The implementation plan of Shanghai's "Computing Pujiang" intelligent computing action stated that by 2025, the proportion of domestic computing power chips used in newly built intelligent computing centers will exceed 50%.

In addition to chips, there are many practical problems to be faced in the actual construction of the 100,000-card group advocated by the global large-scale model track.

First, data centers consume a lot of electricity. According to data, a cluster of 100,000 cards can consume up to 3 million kWh of electricity per day, which is equivalent to the average daily electricity consumption of a city's residents. Second, the larger the computing power cluster, the higher the failure rate. A cluster of 100,000 cards may fail every 20 minutes. Third, computing power is currently in short supply and expensive, but the effective utilization rate of computing power for training large models in many companies is often less than 50%.

Of course, the entire large model track, from enterprises to relevant departments, is trying to solve various accidents in the computing power supply process. First, in terms of energy loss, many international companies overseas have chosen a distributed deployment strategy, and Google and Microsoft are also promoting collaborative training in multiple data centers.

As for chips, many domestic companies are conducting multi-core mixed training. For example, Baidu has achieved 95% mixed training efficiency under the unified management of heterogeneous computing power, and shortened the cluster failure recovery time to minutes. Judging from the utilization rate of some domestic computing power clusters, the situation of computing power waste is improving. The computing power utilization rate of an artificial intelligence computing center in Xi'an is as high as 98.5%.

Various signs indicate that once the global large-scale model market has launched its strategy, there is no turning back. Fortunately, this time, the tragedy of the metaverse should not be repeated in the technology circle.

3. In 2025, is it time for “volume application”?

Starting to create actual value has become the main theme of the big model track in 2025. At present, the application of big models has gradually penetrated into multiple scenarios such as finance, medical health, education and training, search, and office. Li Yanhong once said bluntly that the industry should no longer focus on models, but should directly create application value.

According to statistics from the Economic Observer, as of October 9, 2024, the Cyberspace Administration of China has approved 188 generative artificial intelligence registrations, but more than 30% of the large models have not further disclosed their progress after passing the registration; only about 10% of the large models are still accelerating the training of the models; and nearly half of the large models have directly turned to the development of AI applications.

The reasons for this situation are not difficult to guess. On the one hand, whether the price war in the industry continues or not, the effect behind it is not as good as before. Under the mutual pressure of major giants, the entire market has to tend to healthy competition. On the other hand, the current status of technological development such as computing power resources means that basic models often require hundreds of millions of dollars of investment at a time.

Musk once estimated that the training of GPT-5 might require 30,000 to 50,000 Nvidia H100 chips, and the cost of chips alone would exceed $700 million. Turning to applications has naturally become a major way for a large number of companies to choose to save the country in a roundabout way after failing to compete with technology and capital.

Although the leading enterprises are barely adequate in terms of technical resources and funds, they have already started the "accelerator" for market competition when large models exploded. If they do not seize the opportunity in advance through application, they are likely to be buried in the dust of history. In China alone, general large models and industry large models have emerged in an endless stream in the past two years.

The "Interim Measures for the Management of Generative Artificial Intelligence Services" shows that the general big models that have passed the registration include Baidu Wenxin Yiyan, SenseTime's big model "SenseChat", Baichuan Intelligent's Baichuan big model, and Zhipu Huazhang's "Zhipu Qingyan"; industry big models include Kunlun Wanwei's "Tiangong" big model, Zhihu's "Zhihaitu AI" model, Kingsoft Office's "WPS AI", Future's "MathGPT" big model, and NetEase Youdao's "Zi Yue" education big model.

Some companies have already started the "sea of ​​models" strategy. A typical example is Alibaba. At the 2024 Yunqi Conference, Alibaba not only announced another price cut, but also launched more than 100 models at once, including large language models, multimodal models, mathematical models, and code models. The emergence of large models may be a good thing for the entire track.

However, for a company, the uniqueness of its own products will be greatly reduced if similar products are launched one after another, especially since the current large-scale model track has been stuck in the quagmire of homogeneity. Take Baidu as an example. Although Baidu's large-scale model revenue increased last year, the growth rate dropped significantly.

Data shows that in the third quarter of 2024, Baidu Cloud's month-on-month growth rate dropped from 14% to 11%, and the month-on-month growth rate of generative AI cloud revenue dropped sharply from 95% to 17%. The reason is closely related to the intensified market competition. In order to maintain market share, the value of "applications" must be improved.

However, can enterprises just rush to applications and give up technological progress? One thing that cannot be ignored is that the efficiency of getting orders in the current large model market is closely related to the model itself. In the past year, the number of large model bidding projects has increased sharply, among which Alibaba Cloud, Baidu Cloud, Tencent Cloud, and ByteDance's Volcano Cloud are all frequent winners.

But upon closer inspection, Tencent Cloud won 28 bids with a total amount of 210 million yuan; Alibaba Cloud won 20 bids with a total amount of 570 million yuan; Baidu Cloud won 37 bids with a total amount of 500 million yuan; Volcano Cloud won 24 bids, but the total amount was only 61.86 million yuan.

Why are the four companies so different? This is because although Huoshan Cloud has received orders from all the intelligent body segments, the complexity and customization difficulty of intelligent bodies are not high, so the customer price will fluctuate with the scale of R&D. In other words, the "money prospects" of large models are always positively correlated with technology. In 2025, large models can only continue to roll.