The price war of AI large models starts, and the winner will be decided in a year?

Wu Xiaobo Channel

2024-05-25 09:00Posted on the official account of Zhejiang Hangzhou Bajiuling Cultural and Creative Co., Ltd

"When industry competition appears in front of most people in the form of price wars, it often means that the enterprises in the industry have reached the point where they are on the same page."

Text / Ba Jiuling (WeChat public account: Wu Xiaobo channel)

You may often see such a scene in film and television dramas: when several factions are fighting for territory, guns are raised, confronting each other, everyone does not dare to pull the trigger, and everyone wants to pull the trigger, a fire fight, and it is about to break out.

This is called the "Mexican stalemate," and it refers to the delicate balance between the opposing parties as they contain each other.

Now this deadlock has been broken in the field of AI. The person who pulled the trigger was OpenAI.

On May 14, OpenAI released an upgraded version of GPT-4, GPT-4o, and announced that it would be free for everyone.

A day later, the sound of gunfire from across the ocean finally came, and sparks streaked across the fertile soil.

On May 15, ByteDance announced that its main bean package model has an input price of 0.0008 yuan/1000 Tokens (Token is the smallest unit of input data used by the large model, such as words or characters, etc.).

Bullets began to whizz by.

Six days later, on May 21, Alibaba announced a significant price reduction for the main model of Tongyi Qianwen, and the API input price of the main model Qwen-Long was reduced from 0.02 yuan/thousand tokens to 0.0005 yuan/thousand tokens, a direct drop of 97%, effective immediately.

Four hours later, Baidu skipped the price cut and directly announced that its two entry-level products, ERNIE Speed and ERNIE Lite, were free of charge.

A day later, iFLYTEK and Tencent, who were aware of the hindsight, also announced that the iFLYTEK Xinghuo API capability is free and open, the iFLYTEK Xinghuo Lite API is permanently free and open, and the price of the iFLYTEK Xinghuo pro/Max API has dropped to 0.21/10,000 Tokens;

Tencent's hybrid model has been reduced in price across the board, and the price of one of its main models, the hybrid element-lite model, has been adjusted from 0.008 yuan/1,000 tokens to full free.

……

Gunshots rang out, people were killed, and it didn't seem to happen in this scenario. But no one doubts that in the coming time, such a thing will not happen.

After all, when the competition in an industry appears in front of most people in the form of a price war, it often means that the enterprises in the industry have reached the point where they are on the same page.

However, companies also have their own interpretations of price wars, such as Tan Bei, president of Volcano Engine, who said: "Swapping losses for income is not sustainable, and we will not do it." ”

Another example is Liu Weiguang, vice president of Alibaba Cloud Intelligence Group, who said: "Only when the cost of AI inference is reduced by 10 or even 100 times per year can we truly promote the explosion of AI applications in all walks of life." ”

If you want to go back further, the price of the entry-level large model of the private equity fund DeepSeek and Zhipu AI has been reduced earlier:

On May 7, DeepSeek announced that it would reduce the price of its large model DeepSeek-V2 to 1 yuan per million tokens and 2 yuan per million tokens (32K context);

On May 11, Zhipu AI reduced the price of its entry-level large model GLM-3-Turbo (context length 128k) from 0.005 yuan/1000 tokens to 0.001 yuan/1000 tokens, while the GLM-3 Turbo Batch API was 50% cheaper, reaching 1 yuan and 2 million tokens.

However, does this mean that ordinary people have more opportunities to use large models at low or even zero cost?

The answer, I'm afraid, is no.

Nowadays, there are two main commercial charging models for large models:

The first is to let C-end users pay, that is, ChatGPT Plus $20/month membership subscription model.

The second is to let B-end users pay, that is, developers call the service API, so that developers can connect to the "faucet" of the large model and use the "water" inside.

This time it is the second type of price reduction.

Comparatively speaking, the B-end market is much smaller than the C-side, so it looks like a huge price war, and the cost pressure is not too great for enterprises.

For ordinary people, the good time for everyone to be able to use AI at a low cost has not yet come.

At the same time, further controversy followed: while foreign companies are trying to run hard to maintain technological leadership and run to the sea of stars, domestic companies are commercializing and gaining market share?

For example, Huang said, "our company never talks about market share, which means people are doing the same thing; Musk said that I don't pay attention to technical barriers, I only focus on the speed of innovation; Altman said that AI development is like a tornado, and OpenAI wants to break through the limits.

The more you look at the above words, the more encouraging they become, but in the real environment, when the industry leader is in a leading position in technology, you can only keep leading in order to continue to get the most resources in the industry.

At this time, followers have many ways to catch up, and full commercialization can provide financial stamina for continuous followers.

Before a marathon reaches the finish line, it is not uncommon for followers to overtake the frontrunners.

Of course, there are still all kinds of questions, and we have also invited professional leaders to answer.

These days, the price of large models has been reduced for free, free of charge. If it continues to develop like this, will there be an upside down post in two days, and what will you choose when the time comes.

My choice is that whoever delivers the eggs, I will choose whomever chooses. Before I got to the point where I didn't send eggs, I would like to start from rational logic and talk about three points:

◎ First of all, the price war has brought about the popularization of AI. Some people say that AI is like air, and we can't live without it in the future. And the price war of large models aptly plays the role of "air purifier".

Giants such as Baidu and Alibaba have cut prices and even provided them for free, so that ordinary people can easily access better fresh air and better AI tools.

In fact, the fundamental reason for the price reduction is that the cost of large model inference is also gradually decreasing, and OpenAI's price reduction has reached an astonishing 90% in the past year or so.

It's just that OpenAI has dropped so much in a year, and we have dropped so much overnight, so it feels like the shock wave is even greater.

The inner OS of the domestic large model: Technically, I can't surpass you for the time being, and I can't surpass you at the speed of price reduction?

◎ Secondly, the price war has also brought about internal friction and homogenization concerns in the industry.

Large factories are fighting hotly, and small factories may face a "catastrophe".

Imagine a small AI company in a price war, like a mouse tap-dancing at the feet of an elephant, both to maintain elegance and avoid being trampled on the "rat mud".

Price wars can also lead to misallocation and waste of resources in the market, and excessive price competition will force manufacturers to cut costs, resulting in insufficient investment in R&D and innovation.

This is not only detrimental to the long-term development of the entire industry, but may also lead to excessive market concentration, the formation of oligopoly, and further inhibit the vitality and innovation momentum of the market.

◎ Finally, in order to win the price war, we must insist on innovation and provide differentiated services.

The price war is only the first round of the 100-model war, and the real "ultimate showdown" lies in the innovation of technology and services.

Just like the competition in the food industry, discounted noodles may temporarily capture the stomach of consumers, but what really makes people remember is the bowl of "secret beef noodles" with a unique flavor.

The same is true in the field of large models, after the price war, whoever can make that bowl of "secret beef noodles" will be able to stand out in the future competition.

The price war of large models is not terrible. There are two main reasons for the emergence of price wars.

The first reason is that the computing power of large technology companies is surplus.

When the friction between China and the United States escalated, China's major technology companies were afraid of cutting off supply, so they frantically hoarded Nvidia's chips and quickly piled up computing resources.

Later, Huawei's Ascend series chips were made, and large manufacturers made a large number of purchases. It can be seen from the financial reports of Baidu and Ali that the expenditure in this area is very large, up to billions of yuan.

However, the promotion speed of AI applications is not as fast as they imagined, resulting in the computing power being idling and the resources that were smashed in the early stage being wasted.

Now large manufacturers are trying to find ways to dilute costs and make computing power roll.

It is worth noting that the Nvidia chips in the second-hand market are not easy to sell now.

If you want to sell it to the government, the government won't want it, they will only use domestic chips. If you want to sell to the cloud platform, the cloud platform has surplus computing power, and there is no need to buy new chips.

I got several calls from the United States asking if I could help with them, and they had Nvidia's A100 or H100 in their hands.

The second reason is technological and service advancements.

In fact, ChatGPT is not the same concept as Alibaba Cloud and Baidu Cloud at all.

ChatGPT does not do the public cloud, does not consider the dynamic distribution of the public cloud, nor does it consider the enterprise application scenarios, but frantically piles resources on the large model.

On the other hand, China's cloud service providers need to consider enterprise users and let them use artificial intelligence and large models on their own platforms, so the services of China's large manufacturers will be better.

Whether it is Alibaba Cloud or Baidu Cloud, it is already possible to dynamically allocate the original computing resources with greater value and more efficiency.

For example, the original 100 cards can serve 150 users, but now after technical upgrades, 100 cards can serve 250-300 users, which virtually improves the utilization efficiency of the card.

The surplus of computing resources and the improvement of application technology, the superposition of the two has the basis for a price war.

Throughout history, the current battle of large models, like the battle of thousands of regiments and the war of taxi apps in the Internet era, are subsidized and low-priced, and in the end, there are only two or three giants left.

It is expected that in the first half of next year, the domestic large model will decide the winner, and the public cloud market will also change.

Whoever has the upper hand in this area is likely to take down the public cloud business as well. This is also an important reason why Alibaba Cloud took the lead in cutting prices, and it wants to maintain its dominant position.

There are three main reasons why domestic large models suddenly launched a price war.

First, domestic manufacturers use price wars to circle users;

Second, the more powerful foreign low-end competing products have been free, and the previous tariff standards of domestic manufacturers have been invalidated;

Third, long text applications and multimodal applications require more tokens, and if the tariff remains the same, the user's use cost will be ridiculously high, and the manufacturer will become a platform island.

Price wars are certainly not a good thing, although it is inevitable that prices will go down. Free is not the point, the principle iteration, level improvement, and multimodal evolution of the model are the key.

The massive number of users accumulated for free may be lost overnight because the intelligence level of the model is left behind.

Now, the price war has become a bright spot for domestic manufacturers. This is not right, what everyone should be concerned about is the innovative breakthrough in the iteration of model principles, the improvement of standards, and multimodality.

Companies of different sizes have different breakthrough directions. Startups with shallow capital should focus on application, model fine-tuning or RAG or scenario-based professional models.

However, powerful manufacturers must dare to make breakthroughs at the bottom. If you don't do your homework well, you won't be able to copy your homework in the future, and there is a high probability of copying mistakes.

It's not okay to copy homework, and it's not okay to copy experience. If you want to bring Internet thinking to the wave of superintelligence, I am afraid that you have a misunderstanding of the value source and development paradigm of AI, and the result is tantamount to self-congratulation.

China's Internet has not been able to move towards the next era in time, and behind it is this kind of application, traffic, and fast money thinking that is at work.

In the era of artificial intelligence, technology has become the force of value, and in the competition for AGI, EI, II, traffic is only attached to technology.

Traffic in the sense of Internet celebrity will not bring user loyalty to AI products, especially the general generative model, and users with slightly worse standards will be lost instantly.

Without the continuous evolution of core capabilities driven by underlying technology, building so-called strong intelligence with application scenarios is tantamount to a castle in the air.

In closing, I would like to say that AI must be accessible to everyone, and AI must be human-centered.

Science and technology must have values, science and technology is becoming more and more powerful today and in the future, science and technology and its industry should always think of helping people, serving people, and benefiting mankind, rather than dominating people, let alone controlling people, expelling people, and squeezing people.

With the price reduction of foreign large models and the increasing number of open sources, Chinese large model manufacturers have announced price reductions or free of charge, mainly because they want to expand the number of users.

Large models have a limited number of paying users, making it difficult to train useful AI, so price reductions can attract more users to participate, increase the scale of data, promote model improvement, and also help expand market share.

Price wars can be seen as a strategy for market competition.

On the one hand, it helps consumers obtain lower-cost products or services, and drives technology adoption and innovation.

But on the other hand, excessive price wars may lead to profit compression, affect the healthy development of the industry, and may even trigger vicious competition.

There are not many paying users for large models, and price reductions will not adversely affect themselves.

In the face of the price war of large manufacturers, start-ups should also be prepared to burn money and form their own competitive advantages as soon as possible.

Startups need to focus on market segments, provide differentiated services, enhance technological innovation or improve operational efficiency, and can also seek support from partners or investors to enhance their competitiveness.

In addition to the free strategy, improving the technical level is the core strategy.

In addition, the following measures can be taken to expand the user base: provide high-quality customer service and technical support; Partnering with industry leaders to develop industry solutions; Increase public acceptance of AI technology through education and training.

"Enabling more people to use AI" is a one-way vision goal at the application level. Theoretically, the popularization of AI technology will bring huge social and economic benefits.

However, in fact, the input-output ratio of the large model does not meet this expectation. We are suitable for advocating the long-term goal of defeating large models from the perspective of low-level R&D.

The author of this article | Mei Haoyu | Rao Zufen | Editor-in-Charge | He Mengfei

Editor-in-Chief | He Mengfei | Image source | VCG

View original image 589K

• 

• 

• 

• 

• 

• 

• 

•