DeepSeek ist populär geworden. Die Goldsucher sind enttäuscht, während die Schaufelverkäufer feiern.
Layoffs, contractions, and executive upheavals. Enterprises can't burn their way to a "large model ticket."
Text | Intelligent Emergence
Early this morning, Amazon Web Services announced the launch of the fully managed, serverless DeepSeek - R1 model on the Amazon Bedrock platform. It is the first overseas cloud provider to offer DeepSeek - R1 as a fully managed, officially commercially available model. Therefore, for enterprises going global, Deepseek on Bedrock is an excellent choice. On the other hand, as the first domestic large model to land on Amazon Bedrock, since its launch at the end of January this year, thousands of customers have deployed the DeepSeek - R1 model through the custom model import function of Amazon Bedrock.
Since the Spring Festival in 2025, the hottest topic globally has been DeepSeek.
At the beginning of the year, the DeepSeek - R1 paper emerged out of nowhere and almost instantly dominated the "battle of a hundred models."
Overseas, the "Magnificent Seven" tech stocks plummeted, creating a deep hole in the three major US stock indexes. In China, almost all AI large - model players "stalled." Advertising spending contractions were the first to be hit, and there were also many team layoffs and executive upheavals.
The market's reaction is in sharp contrast to that at the end of November 2022. At that time, ChatGPT also "emerged out of nowhere," but it triggered a global tech industry frenzy. Tech stocks at home and abroad soared, and huge amounts of capital poured into the "battle of a hundred models," creating a booming situation.
Now, with the emergence of DeepSeek, it has not only explored countless possibilities for the large - model path of reinforcement learning but also fully and unreservedly open - sourced itself, allowing everyone to "stand on the shoulders of DeepSeek" to continue innovating. This can at least be regarded as the industry's "AlphaGo moment" (AlphaGo Zero defeated the human Go champion Ke Jie through reinforcement learning). So why is the market reaction so pessimistic?
In fact, the popularity of DeepSeek, the market's pessimism, and the various reactions of competitors are just the results, not the reasons for these phenomena.
The truth is, it has become increasingly "unprofitable" for enterprises to develop their own large models.
Can't Burn Their Way to a "Large Model Ticket"
Just recently, on March 1st, DeepSeek officially disclosed that according to the token pricing level of DeepSeek - R1, the company's total daily revenue was $562,027, with a cost - profit ratio as high as 545%.
A number that makes most large - model players despair.
On one hand, there are high R & D costs, with well - paid doctors, post - doctors, and chief scientists. On the other hand, there are low daily active users and huge investments with disproportionate ROI. More and more enterprises can't burn their way to a "large model ticket."
In contrast, the competition among large models is becoming increasingly fierce.
According to AppGrowing data, as of November 2024, the total number of advertising placements of ten domestic AI applications such as Kimi, Doubao, Xingye, and Yuanbao exceeded 6.25 million, with the converted amount exceeding 1.5 billion yuan. Among them, Yuezhianmian was the most aggressive, with a cumulative investment of over 540 million yuan.
However, DeepSeek stunned all its competitors overnight. According to Feifan Chanyan data, as of February 5th, 2025, the daily active users of the DeepSeek mobile app had exceeded 40 million. On the same day, the daily active users of the ChatGPT mobile app were 54.95 million, and DeepSeek had reached 74.3% of ChatGPT.
According to foreign media such as The Information and The Wall Street Journal, OpenAI is in talks with investors for a new round of financing, with a maximum financing amount of up to $40 billion (about 290 billion yuan).
At this stage of financing, no matter who is the CEO of OpenAI, they must be so anxious that they tweet every day.
△ Comparison of the daily active user (DAU) data of ChatGPT and DeepSeek mobile apps as of February 5th, 2025. Image source: Feifan Chanyan
Facing a strong competitor, domestic competitors such as Alibaba, Baidu, and Tencent have all extended an olive branch to DeepSeek. The "full - version DeepSeek" has become the latest traffic code on the Internet.
Among them, Tencent was the most aggressive. Different from its competitors, which connected to DeepSeek at the cloud computing service level, Tencent directly integrated DeepSeek into its own AI application, "Yuanbao," making it jump to the second place on the free app download list of the Apple App Store in China. It even surpassed DeepSeek, which had long occupied the top spot, and became the first.
In a recent media interview, Zhu Xiaohu, the managing partner of Jinshajiang Venture Capital, changed his previous distrust of AGI and said, "DeepSeek is making me believe in AGI."
In addition, he emphasized again, "Start - up companies should never develop underlying models. Just capture users and scenarios on top... As the underlying models progress, just use the best and latest models."
This view may sound harsh and a bit one - sided, but for some enterprises, it may not be completely unreasonable.
The large - model competition has entered the next stage.
Gold Diggers and Shovel Sellers
In sharp contrast to the reshuffle of large - model enterprises, there is a soaring global corporate demand for AI, and the outstanding performance of major cloud - computing giants selling AI large - model services.
According to the latest quarterly earnings report, Google Cloud's revenue in Q4 2024 was $11.955 billion, a year - on - year increase of 30%. Google said that this growth was mainly due to the strong performance of the core products of the Google Cloud platform, AI infrastructure, and AGI (generative artificial intelligence) solutions.
Microsoft's Intelligent Cloud business had a revenue of $25.54 billion in the latest quarter (including Azure), a year - on - year increase of 19%. The revenue of Azure and other cloud services increased by 31%, of which 13% of the revenue came from AI - related businesses (the specific revenue data of Azure was not disclosed).
Amazon Web Services, the global cloud leader, had a revenue of $28.8 billion in Q4 2024, achieving a year - on - year revenue growth of 19% for the second consecutive quarter. Andy Jassy, the President and CEO of Amazon, said in an analyst conference call that he was optimistic that Amazon Web Services' cloud services would support most of the global AI workloads.
△ Market share of global cloud giants. Image source: Synergy Research Group
Moreover, a more interesting phenomenon is that at the beginning of the year, when DeepSeek stirred up the global market, three overseas enterprises, Amazon Web Services, Microsoft Azure, and NVIDIA NIM, announced on the same day, January 30th (US time), that they would list or connect to the DeepSeek - R1 model. A few days later, Chinese cloud providers such as Baidu Cloud and Alibaba Cloud successively announced their connections.
Especially Amazon Web Services. This way of connecting to the latest and strongest models at the first time is not the first time.
As early as March 2024, when the Claude 3 series of models "emerged as the new king" and outperformed GPT - 4 in various data, Amazon Web Services listed Claude 3 in Amazon Bedrock at the first time.
The same story happened to DeepSeek, Meta Llama, Stability AI, Mistral AI, and other models. According to official data, the Amazon Bedrock platform of Amazon Web Services currently has more than 180 foundation models for customers to choose from, truly starting a "battle of a hundred models" on the cloud platform.
This approach of Amazon Web Services is in line with its long - standing "Choice Matters" strategy in the large - model field.
Looking globally, although they are all cloud - computing giants, each has a different approach:
Needless to say about Microsoft. In addition to making a huge investment in OpenAI, the Azure OpenAI service also exclusively supports OpenAI models. In 2024, Microsoft established a distribution partnership with French Mistral AI and invested in it, further expanding its European market. At the same time, Microsoft has the powerful Office suite, and the Copilot intelligent assistant has become the natural dominant area in the cloud AI office scenario.
Google is an established giant in the AI era. The god - level paper "Attention is All you Need" in this round of large - model technological innovation was written by Google. Google Cloud tends to promote its own Gemini family first in the large - model field and performs well in multi - modality and joint reasoning across text, images, audio, and video.
It both digs for gold and sells shovels.
As the pioneer of cloud computing and the leading player in the cloud - computing field to date, Amazon Web Services is keen on expanding the number of cutting - edge models available for customers to choose from, such as DeepSeek, Llama, Mistral, etc. In addition, Amazon made heavy investments in Anthropic at the end of 2023 and the end of 2024, with a total investment of up to $8 billion. Amazon Web Services also released the powerful multi - modality, cost - effective self - developed model Amazon Nova at the re:Invent conference last year.
In fact, Amazon Web Services' strategic thinking in the large - model era has always been more firmly inclined to "selling shovels."
This strategy was taught to them by the market.
According to the report data of Jefferies & Company, currently, only 3% of enterprises use only one language model provider, while 34% use two, 41% use three, and 22% use four. According to Gartner's forecast data, by 2027, 80% of Chinese enterprises will choose a multi - model strategy.
For enterprises, it is extremely important to ensure their right to choose multiple models on the cloud.
On one hand, from the perspectives of security, cost, or complex business adaptability, the multi - model strategy on the cloud will be the best choice for enterprises.
On the other hand, in a more profound sense, this round of generative AI brings not only the iteration of production tools but may also be a change that touches the essence of the business model. What enterprise decision - makers care most about is the business growth and business value brought by generative AI. Whether it is DeepSeek, Claude, Nova, or Gemini, they are all just parts of helping enterprises achieve business value, not the whole. The "Choice Matters" strategy formulated by Amazon Web Services points exactly in this direction. Only generative AI that can bring substantial business value to enterprises makes sense.
Since the explosion of large models at the end of 2022, two and a half years have passed. If 2023 was the year of large - model verification and 2024 was the year of large - model products, then 2025 will be the year when the business value of large models is realized.
Variables
In addition to cost, security, and business value, there is another important reason for enterprises to value the "right to choose on the cloud": there are still a large number of variables in the current large - model industry.
In terms of underlying technology, DeepSeek - R1 has replicated the glory of GPT - o1 as an inference model with a genius technical idea and ingenious engineering methods, and has explored countless possibilities for the large - model path of reinforcement learning. However, the long - delayed GPT - 5 is still a "secret weapon" in its infancy.
In terms of vertical scenarios, although the Diffusion architecture has become the absolute king in the text - to - image field, in the fields of video, 3D, and other vertical applications, no one has been able to dominate. In the more upstream industries such as healthcare, finance, and retail, there is even less consensus on the path.
Among the large - model competitors, the newly launched Claude 3.7 Sonnet by Anthropic is equipped with hybrid reasoning ability. Even in DeepSeek's core area - quantitative investment - there are endless cross - border challengers. Just recently, the Jiukun team, a quantitative giant, successfully reproduced DeepSeek - R1, and the Kuande team issued a recruitment notice for its intelligent learning laboratory on February 24th, diving into general artificial intelligence.
DeepSeek has become extremely popular, and the whole nation is in an uproar, but the large - model competition is far from over.
The next disruptive breakthrough in large - model technology may occur in an unexpected place - it may come from DeepSeek, OpenAI, or a startup team in India, France, Russia, or in Shenzhen, Chongqing, or Shanghai in China. The recently popular Manus is a good example.
However, as mentioned above, in 2025, the year when the business value of large models is realized, most enterprise customers actually don't really care which large - model enterprise their "black or white cat" comes from. As long as these models are usable, good, and at the forefront, it's enough.
For the "shovel sellers" of large models, the major cloud - computing providers, this competition has just begun.
It took 30 years for the Internet to become an infrastructure after its invention, and it took nearly 20 years for cloud computing to be widely applied after its technical exploration. For large models to truly become the "new infrastructure" for all industries, it is still in the distant future.
△ The global AI server market scale will maintain rapid growth from 2023 to 2028. Image source: IDC 2025
According to the latest data from the earnings conference calls of four technology giants, Amazon, Microsoft, Google, and Meta, in 2025, their total capital expenditure (CAPEX) is expected to exceed $320 billion. Among them, Amazon Web Services' capital expenditure is expected to reach $100 billion to seize the "once - in - a - lifetime business opportunity" in the AI field.
The competition at the technical R & D level of each company is no exception.
Although the GPT - 4.5 series of Microsoft OpenAI is controversial, the industry still expects an excellent performance from GPT - 5. Google has just released the Gemini 2.0 series of models, fully realizing native multi - modality input and output. Amazon Web Services presented its latest Amazon Nova model at its re:Invent conference - an all - around, cost - effective multi - modality foundation model, which can be regarded as the most cost - effective "alternative" to Claude 3/3.5, GPT - 4o, and Gemini 2.0, suitable for scenarios such as enterprise customer service, content generation, simple data analysis, and enterprise internal automation.
In addition, each company is trying its