虎嗅

AI Large Models Enter an “Unlimited War”

原文:AI大模型进入“无限战争”

Summary of Key Points

This article focuses on the intense competition in China's AI large-model field in 2026: Technological companies represented by DeepSeek (low-cost and open-source) and Kimi (task-oriented) are competing with giants such as ByteDance, Alibaba, and Tencent, each pursuing different strategies to attract users. Behind this is an ultimate battle about how to price AI intelligence. The industry landscape is still undecided, and no company has a true competitive advantage. This battle will determine how ordinary people will use AI in the future.

1. DeepSeek's Price Revolution: Making AI More Accessible

DeepSeek is the "price disruptor" in this competition. In May 2024, it released its V2 model with an API price that was only one percent of that of GPT-4 Turbo, triggering a price war among domestic large-model providers: Zhipu reduced its prices by 80%, ByteDance's DouBao offered extremely low prices, and Baidu and Tencent even made their models free. In January 2025, DeepSeek launched the fully open-source R1 model, which not only had near-top-tier capabilities but could also be compressed into a smaller size for easier use. This caused panic in the market, as it became clear that AI models didn't have to be expensive or rely on large amounts of GPUs (NVIDIA's core product). As a result, NVIDIA's stock price plummeted by 17% in one day, losing $589 billion in value.

DeepSeek's philosophy is simple: Intelligence should be as affordable as water and electricity, accessible to everyone. The founder, Liang Wenfeng, stated that they set prices based on costs, aiming neither to lose money nor to make exorbitant profits. This approach has transformed AI from a tool in the hands of a few companies into a universal tool, completely changing the industry's pricing rules.

2. Kimi K3's Different Approach: Focusing on Task Completion Rather than Low Prices

Kimi K3 takes a completely different path from DeepSeek. It is the largest open-source model in the world (with 2.8 trillion parameters) and experienced a computational overload just two days after its release due to the high number of users, forcing a suspension of new subscriptions. However, its pricing is high: 20 yuan per million tokens for input and 100 yuan per million tokens for output, which is 23 times that of DeepSeek V4 Pro.

Kimi's approach focuses on completing tasks rather than simply selling the number of tokens. For example, it can help with writing reports, processing local files, or using a browser to search for information. The founder, Yang Zhilin, believes that if a model can reduce an engineer's workload from a week to a day, users won't care about the cost of tokens. This represents a paradigm shift: Instead of competing on who responds best, models now compete on how effectively they complete tasks.

3. Giants' Strategies for Attracting Users

ByteDance, Alibaba, and Tencent don't directly sell AI models; instead, they use them to capture user traffic:

  • ByteDance DouBao: Links its services to the Spring Festival Gala, using emotional engagement to retain users (1.9 billion interactions on New Year's Eve, generating 50 million avatars). The strategy is to make users want to communicate rather than being forced to.
  • Alibaba Qianwen: Spends 3 billion yuan on promotions; users can place orders for milk tea for just 1 cent. The goal is to make AI a central part of daily life, integrating with services like food delivery and ticketing in Alibaba's ecosystem.
  • Tencent YuanBao: bets on "AI + social interaction," launching group chat AI features and using 1 billion red envelopes to attract new users. However, subsidies can only bring short-term downloads; long-term usage habits are harder to establish.

The giants' focus is to use AI to retain users within their ecosystems rather than relying solely on the models for revenue.

4. The Ultimate Battle for Pricing Power

The fundamental question at heart of this battle is: How much should "intelligence" cost? Different companies have different approaches:

  • DeepSeek: Prices based on the number of tokens, which is cheaper and more transparent.
  • Kimi: Charges per task completed.
  • ByteDance: Charges based on user usage time; the longer users use the service, the more valuable it becomes.
  • Alibaba: Charges for cloud resources used; the model runs on Alibaba's cloud, so they charge for computing power.
  • Tencent: Charges a percentage of transactions completed through AI.

In the past, companies like Windows, Google, and WeChat defined pricing power through their platforms, entry points, or ecosystems. With AI still unsettled, the battle will continue. For example, if users get used to paying per task, Kimi will have an advantage; if they prefer low-token prices, DeepSeek will be more competitive.

5. An Era Without Competitive Advantages: Anyone Can Win or Lose

No Chinese AI company currently has a true competitive advantage:

  • DeepSeek's cost advantage could be eroded by new technologies that reduce costs.
  • Kimi's task-solving capabilities could be challenged by stronger models from companies like OpenAI.
  • ByteDance's user traffic might be lost if users switch to other AI apps.
  • Tencent's ecosystem might be compromised by more powerful models integrating with WeChat.
  • Alibaba's all-in approach might not be as effective if its models are inferior or its entry points are less attractive compared to ByteDance and Tencent.

However, this is also a positive development: The uncertainty means that newcomers still have a chance. This battle will determine not only the fate of these companies but also how ordinary people use AI in the future—whether they will pay by the number of tokens, by the number of tasks completed, or if they will use it for free while being bombarded with ads.

Just as Turing asked 75 years ago whether machines could think, we are now asking how to charge for machine-generated intelligence. As long as there's no clear answer, the battle won't end.