DeepSeek makes everyone Perplexity

As early as May 2024, DeepSeek-V2 was called the "price butcher" and the "Pinduoduo of the AI ​​industry" for its price of one percent of GPT-4. At that time, the impact it brought was still limited to specific actions such as pricing in the underlying model competition.

Now, the emergence of R1, after achieving an order of magnitude improvement in actual cost-effectiveness, has once again set off a "zero-yuan purchase" carnival in the large model circle, and directly changed the fundamental strategy of the players who have invested the most money in the competition for tickets to the AI ​​era.

From cloud computing giants to mobile phone manufacturers, from car companies to Internet platforms, the wave of entry brought about by DeepSeek is dazzling.

First, international giants such as Microsoft, Amazon, and Nvidia rushed to connect to DeepSeek, and then domestic cloud service providers quickly followed suit. Alibaba Cloud, Tencent Cloud, Baidu Smart Cloud, Volcano Engine, etc. were ready to go, and even offered free tokens as a gift, for fear that users would miss out on this carnival.

From smartphones to smart cars, many manufacturers are also connecting their products to DeepSeek. Under the dual temptations of performance and cost, instead of worrying about whether to develop in-house or which company to choose, it is better to go all in on DeepSeek.

The most magical plot is yet to come. The latest surprising move comes from the big companies.

In recent days, manufacturers who insist on self-developed large models have begun to integrate DeepSeek into their core products and provide them to users for free. When Wen Xiaoyan's update log clearly stated "Integrated with DeepSeek-R1", this scene was simply a magical turning point in the history of domestic AI development.

Platforms such as Tencent Yuanbao, Zhihu, and iFLYTEK have also embraced DeepSeek, as if they were performing the classic script of "If you can't beat them, join them."

Even OpenAI can’t sit still. Just yesterday, Ultraman revealed that GPT-4.5/5 will be released soon, and it will also open a “zero-dollar purchase” program, giving free users unlimited chat privileges.

As DeepSeek becomes ubiquitous, a true era of AI for all seems to be accelerating. Will this carnival that started with "zero-yuan purchases" become a turning point that changes the industry landscape?

When the words "connect to DeepSeek model" appeared in the update logs of Wenxin Yiyan and Tencent Yuanbao, Perplexity CEO Srinivas's words "If your goal is to build a product-centric company, don't waste time training your own models" became more meaningful.

How much resources have these manufacturers invested in self-developed big models? With tens of billions of R&D investment and thousands of AI research teams, they regard self-developed big models as their core competitiveness in the future. However, when an open source model achieves the same performance at a lower cost, the so-called technology "moat" collapses in an instant. What they are facing is not just a simple question of "whether to access or not", but whether the entire strategic layout needs to be rethought.

If small and medium-sized enterprises can embrace the open source model with ease and without any burden, then these leading manufacturers face a more complex choice: continuing to insist on self-research may cause them to lag behind in product experience in the short term; but easily giving up self-research means that the huge investment before may be wasted, and more importantly, they may lose their voice in the next round of technological revolution.

In this dilemma, the "embracing" of DeepSeek and these "zero-yuan purchases" we see are not so much active choices as compromises that have to be made under market pressure. This may also indirectly confirm a cruel reality: when users can get better services at a lower cost, there is no loyalty.

In this era of "models overturning everything", no one can predict when the next technological breakthrough will occur. Currently, manufacturers who still rely on the Internet era to hold a large number of users have found that the only thing they can do is to firmly grasp these user bases and application scenarios, and use all possible means to maintain sufficient openness and flexibility to ensure that users stay in their product ecosystem. Even if this means admitting their past misjudgments, even if it means becoming an "AI Wrapper" in the words of Perplexity, as long as "you provide so much value, no one will care." After all, in the final analysis, only by staying at the table can we see the future.

So, from this perspective, the response of these big companies is unprecedentedly fast: the research and development of their basic models is still continuing, but the positioning is changing. For example, the positioning of Byte Seed will become more mainstream, and the focus will be on research again. After all, the next major breakthrough in the underlying AI has become open again. There may be opportunities if there is research determination. The role of pre-trained models is finally no longer presented in a rush, which is also a relief for these manufacturers.

At the same time, at the product level, the rapid “if you can’t beat them, join them” approach actually reflects a rare flexibility. No longer relying on a single model may be a better solution to cope with the rapidly changing market environment.

Behind these rapid strategic adjustments is the manufacturers' new understanding of the business model in the AI ​​era after being stimulated by DeepSeek.

The emergence of DeepSeek R1 has made the AI ​​industry experience its "steam engine moment". Just like James Watt's improved steam engine in 1765 was able to more efficiently convert thermal energy into mechanical energy, greatly improving its efficiency, DeepSeek has achieved GPT o1-level performance at 1/10 of the cost.

Economist William Stanley Jevons observed after the advent of the improved steam engine that the improvement in efficiency did not reduce coal consumption, but instead led to a surge in demand due to the lower threshold for use. This "Jevons paradox" seems to be repeating itself in the field of AI.

The precipitous drop in token costs has allowed AI to penetrate into previously unimaginable scenarios. Token pricing is like the "coal price" of the Victorian era, and now this barrier is crumbling, and DeepSeek has greatly accelerated this process. When the token cost approaches zero, just like the evolution of email services from "pay-as-you-go" to infrastructure, the business model of AI products that charge users directly will probably be redefined. Especially in consumer scenarios, the direct charging model may gradually be replaced by indirect monetization methods such as advertising and value-added services.

Srinivas, the founder of Perplexity, once said that "advertising is the greatest business model of the past 50 years". This statement seems particularly prescient at the moment. In fact, as early as April 2024, this company, which once claimed to be "not affected by advertising-driven models", had already begun exploring the integration of advertising into AI dialogue scenarios - presenting advertisements in the form of "recommended questions". This face-slapping change may foreshadow the future direction of AI business models.

As we have seen in the past, not just with the Internet, but also with electricity, radio, computers, cars, cell phones, and search engines. The creators of these technologies have a strong incentive to reduce the price until everyone on the planet can afford it. This is also what is happening in the field of AI, which is why we can use the most advanced generative AI at low cost or even for free, and this situation will continue.

But as Marc Andreessen said, this is not because these companies are stupid or generous, but on the contrary, because they want to maximize the market size and thus maximize profits. In this period of technological change where everything is turbulent, it is the best strategy for each company to tightly protect the resources accumulated in the previous era, not afraid of being "slapped in the face", and take advantage of the fact that DeepSeek and others have not thought about competing for users themselves, and use products to build their own user ecosystem as much as possible under the strategy of free and "join if you can't beat them".