OpenAI cut off supplies to China, but Microsoft Cloud left behind

Alphabet list

2024-07-05 17:58Posted in Beijing, creators in the field of science and technology

The full text is 3974 words, and it takes about 12 minutes to read

Highlight the point

01OpenAI will block domestic API calls on July 9, 2023, but domestic AI entrepreneurs have not ushered in an industry earthquake because domestic large models have begun to replace ChatGPT.

02Major manufacturers such as Microsoft Cloud, Baidu, Alibaba, Tencent, etc., have launched OpenAI's zero-cost migration plan to attract developers.

03However, domestic large models face challenges in competing with OpenAI, such as more than 3,000 enterprises that have access to OpenAI's technology services for Azure in Microsoft Cloud.

04In addition, domestic large models need to be continuously innovated and optimized to cope with the competition of large manufacturers in the AI application layer.

Technical support is provided by Tencent Hybrid Model

OpenAI cut off supplies to China, but Microsoft Cloud left behind

"The thunder is loud, the rain is small", AI entrepreneur Zhang Qing described OpenAI as saying that it will block domestic API (application programming interface) calls from July 9 this year.

In his opinion, although "more than half of the top large-scale model products in China in 2023 are 'shell' ChatGPT (that is, calling open-source large-scale model APIs to provide services, rather than self-developed models)", OpenAI's interruption of ChatGPT's call did not usher in the imaginary industry earthquake.

In fact, as early as June 2024, Fu Han, the front-end developer of an AI start-up, changed the large model of product invocation from GPT to moonshort.

In May, when the product was just launched and debugged, Fu Han chose to call GPT 4.0. After all, compared with domestic large models, the "price is almost the same, but the processing speed is faster and the feedback results are better" when calling OpenAI.

But after 5 months, under the high cost of large-scale invocation of OpenAI, as the domestic large model began to be purchased at 0 yuan, Fu Han found that "the text processing of the domestic large model is not inferior to GPT 4.0 within a certain range", in the face of dozens of times, or even hundreds of times the price difference, Fu Han spent two weeks to turn the interface to the domestic large model.

The temptation brought by the huge price difference undoubtedly makes AI entrepreneurs "really fragrant". "The same 1 million Tokens (the smallest processing unit of large model text) can be used by 100 users for 5 months, and it costs 100 yuan to buy GPT 4.0, but the domestic large model DeepSeek only needs one yuan." Fu Han told Alphabet List (ID: wujicaijing) that now under the price war of the head large model manufacturers, domestic AI entrepreneurs use domestic large models, and the cost of Tokens in a single month may be hundreds of times different from calling GPT 4.0.

With the news that OpenAI will cut off supply, major manufacturers have dispatched and begun to "cut off" developers. Up to now, including Zhipu AI, Zero One Everything, Dark Side of the Moon, Baichuan Intelligence, SenseTime, Baidu, Alibaba, Tencent, etc., have launched OpenAI's zero-cost migration plan. At the same time, some large-scale model manufacturers also intimately give away 10 million or even 100 million Tokens.

"Now the SDKs of domestic large models are compatible with the OpenAI format, and even if you switch it, it's just a matter of a few lines of code." Fu Han said.

However, for the domestic large models that are in a hurry to commercialize and want to use this to snatch OpenAI shell users, they also face competition from Microsoft Cloud, the owner behind ChatGPT.

At present, enterprises that access OpenAI technical services through Microsoft Cloud Azure are more than 3,000 companies with daily traffic and tens of thousands of registered users, including Tencent, Byte, and Huawei.

OpenAI's popularity is easy for domestic large models, but it is still a little difficult to eat in your mouth.

OpenAI's supply cut rushed to the hot search, but the AI entrepreneurs who relied on ChatGPT to "eat" were quite calm.

"Through the Microsoft cloud, after July 9, you can also use a Chinese IP to access OpenAI, as long as you select an agent from Microsoft's official channels to sign a tripartite agreement, you can use your Azure enterprise account to apply for Open AI services by yourself." The above-mentioned Microsoft cloud employee told Alphabet List. In Beijing, you can choose Kunlian, in Shenzhen there are SoftwareONE and other companies, and the prices of these Microsoft agents are also completely unified.

As the de facto majority shareholder of OpenAI, Microsoft Cloud has long set its sights on the Chinese market. "Before ChatGPT exploded, Microsoft Cloud had already developed a huge ToB business in China, and Microsoft's relevant qualifications for ToB have also been stored in the knowledge base of major companies," Zhang Qing said, and now buying OpenAI's access services from Microsoft Cloud does not even need to pay extra for the original ToB business.

In 2023, taking advantage of the popularity of ChatGPT, Zhang Qing will access OpenAI and make an AI horoscope applet, "dozens of new users pay every day", using ChatGPT to remind daily horoscopes and provide love horoscope analysis, as the number of users increases, access surges, "ChatGPT's nodes are abroad, and when it reaches its peak, the access speed will slow down", which also prompted Zhang Qing to become an enterprise user of Microsoft Cloud in June 2024.

If you want to access ChatGPT in China, you only need to register an enterprise account, meet the threshold of at least $1,000 per month, and pay $360,000 (about 250,000 yuan) for 3 years in advance according to the cost of $12,000 a year. And since Zhang Qing registered a branch in Hong Kong, he can also be exempted from 6% of the relevant tax.

Different from Zhang Qing, at the beginning of the product launch in May, Fu Han connected the ChatGPT interface to an AI product in traffic safety, and in addition to unstable nodes, price has also become an inevitable consideration for AI entrepreneurs who are still in the cold start stage.

"In the past, when the price of the large model was not reduced, there was almost no difference between the price of calling ChatGPT and using the domestic large model, but now, the price difference between the domestic large model and ChatGPT has reached dozens of times, or even hundreds of times." Fu Han told Alphabet List that when the product was just launched, after testing, GPT 4.0 was the most in line with the response needs of the product, and although the feedback accuracy and response speed of the domestic large model were "not as good as GPT 4.0", after the price of the domestic large model was reduced, he still chose to switch the model.

The domestic large model that fights the price war has attracted AI companies and shell users in the market by becoming a "ChatGPT replacement". "As far as I know, the byte model is ridiculously cheap, and if you do some small scenes, the cost is also very suitable for using the byte model." Fu Han said that after evaluating the domestic large models on the market, in the case of "not having too high expectations for the output results", it only meets the requirements of daily dialogue and processing, DeepSeek is cheap and big, "1 million tokens is only 1 yuan, and it is also compatible with OpenAI's SDK format", MiniMax is fast and suitable for specific scenarios that require speed, and Kimi is cheap and suitable for application scenarios that need to output long text.

Becoming the replacement of OpenAI is becoming the next scenario for domestic large models to compete for.

Before being replaced by a domestic large model, OpenAI was once the best choice for domestic AI entrepreneurs.

"Prepare individual licenses, service number authentication, domain names, rent a domestic server, a foreign server, and about 2,000 pieces can make a small application of ChatGPT." Zhang Qingdao.

In 2023, at the moment when ChatGPT is gaining popularity, Li Yizhou, who has not yet overturned, sold 250,000 copies of the 199 yuan "Artificial Intelligence Course for Everyone", OpenAI not only supported AI knowledge blogger Li Yizhou, but also supported Zhang Qing who rushed to set ChatGPT.

Within a year, Zhang Qing has launched nearly 10 ChatGPT shell applets from AI astrology to AI synthetic photos. For each applet, Zhang Qing only needs to use OpenAI's API interface to complete the interpretation of the cards of the astrology and the post-synthesis of the pictures.

With the help of social platforms such as Xiaohongshu and Douyin, Zhang Qing will go to social platforms to post Amway posts, go to Douyin to do rhythm mixing, and will also go to AI bloggers' account comments to do "interception", that is, "In the comment area, pretend to be ordinary users, and xxAI is very easy to use." ”

Under Zhang Qing's operation, there will be 300-400 users entering his mini program every day, according to the 5% payment rate, 30-50 yuan per month membership fee, you can sell the API account that costs a few yuan to hundreds of people, Zhang Qing's profit margin can reach 100%, and the single-day income is more than 10,000 "easily, even if there are only a few thousand fans, there is a fixed profit of hundreds of yuan a day".

At the time of the most popularity, Zhang Qing prepared more than 3 individual licenses, "If the transaction volume is large, it is easy to be intercepted by the system, and I will prepare more than 3 licenses to pay in turn." ”

In 2023, while the domestic large model is catching up with ChatGPT and flawlessly launching AI applications, Zhang Qing borrows the ChatGPT port, and 2-3 people can earn more than 10,000 small workshops a day, making a lot of money.

At the 2024 Venture Capital Decade Summit, Zhu Xiaohu, Managing Partner of GSR Ventures, pointed out that Klarnr in Northern Europe has optimized the customer service end of the call center through "shelling", reducing 700 customer service positions, and bringing an annual profit increase of 40 million US dollars (nearly 300 million yuan) to the company.

As the head application of AI search, Perplexity, launched in December 2022, also once relied on OpenAI's GPT 3.5 model, in this regard, Perplexity CEO Srinivas bluntly said, "A shell product with 100,000 users is more meaningful than having its own model but no users." ”

However, as large model manufacturers began to launch APP applications, including Byte Doubao, Tencent Ingot Treasure, MiniMax Conch AI and other applications, Zhang Qing's good days have also come to an end.

Before May 2024, Zhang Qing's small program "can have at least hundreds of new users every day, and dozens of paying users", and with the launch of the integrated APP of the big factory, the new users of Zhang Qing's small program are facing a cliff-like decline, "The pictures that need to be paid to generate in the small program can be directly free of charge by large manufacturers, and the functional scenarios are more abundant," In contrast, Zhang Qing's small program developed only for a certain function seems quite chicken, and it is precisely because of this that now" The number of users who enter the private domain of the Mini Program every day is basically halved. ”

Even drainage platforms such as Douyin, which Zhang Qing is good at, have already become the world of byte bean bags, not to mention that MiniMax even hit the conch AI advertisement to Huya and Douyu, and in the face of the big factories that spend money to push applications, Zhang Qing bluntly said "I can't bear it".

Whether it is accessing OpenAI through Microsoft Cloud or switching to a domestic large model, it is a definite fact that the traffic dividend brought by OpenAI's popularity in 2023 is disappearing for AI entrepreneurs with large models in the shell.

In 2023, Zhang Qing's remixed video showing AI synthetic pictures became a hit on Douyin, "On the day the video was released, nearly 1,000 people poured into the private domain of the mini program, and the daily turnover for a week exceeded 20,000 yuan. But now, as AIGC has become an "old topic" from a new social currency, Zhang Qing clearly feels that "there are more and more similar AI products on Douyin, but the traffic is getting worse and worse." Users don't have a sense of freshness anymore. ”

In 2024, from Miao Ya to Yuanbao, Huang Zhang, who was born in the 00s, told the alphabet list that before, every time a new AI model or AI application was released, she would download and try it, and the first time she used Midjourney to draw, she excitedly sent 3 Xiaohongshu notes. As the large model soared to more than 200, Byte, Tencent, Ali, and Baidu all launched their own large model applications, and Huang Zhang was not only difficult to distinguish the specific differences between these large models, but also became more and more lacking in interest, and finally "they rarely opened it again after trial".

Even the classic products of OpenAI, such as PDF.ai, its founder Damon Chen still posted in December 2023 PDF.ai revenue has exceeded $60,000, and then in February 2024, Adobe officials not only directly integrated AI functions in the product, but also spent money to buy PDF AI-related keywords.

In just a year and a half, OpenAI's API dividends have faded. With the improvement of the underlying capabilities of domestic large models, leading manufacturers have squeezed into the AI application layer to compete for the remaining traffic. It is becoming less and less possible to try to make a "big hit" based on the fine-tuning of a large open source model.

"You can consider making models for the specific scenarios of the enterprise's own business," Fu Han said, for developers, having their own developed models can undoubtedly take stability and initiative into their own hands, and avoid uncontrollable unexpected factors such as access restrictions.

However, Bruce, who took an angel investment of 7 million yuan and left a leading factory to start a business, told Alphabet List that they are currently developing their own AI travel planning products. In the product development period, the project should also be based on "whether the big manufacturer will do it". At the beginning, based on the need for compliance, he chose to access the domestic large model (Zhipu Qingyan, Tongyi Qianwen), but when he was about to enter the cold start stage of the product, he was most anxious about, "Are there any users to use it?" How to let more users know about this product? ”

Different from large factories with big businesses, in the account of @小布Bruce, Bruce said that start-up AI companies often have no extra manpower and costs to invest, and it is best to find demand scenarios that large manufacturers do not know how to do, or that are not refined enough. But "the patience of investors is not enough". After the launch, the acceleration became the inevitable choice for Bruce.

However, the struggle to become a domestic large model to replace OpenAI is not easy.

"Last year, we were focusing on improving the dark side of text processing, and this year we have hired a lot of bytes of people to do growth, or the mobile Internet play," Bruce said. ”

OpenAI's supply cut is imminent, and an employee of the AI business of a leading manufacturer told Alphabet List, "The company's stock price has fallen, and its reputation has become worse." Obviously, without OpenAI's competition, the problem of big manufacturers trying to win more user recognition has not been reduced.

(Fu Han, Zhang Qing, and Huang Zhang are pseudonyms in the article)

View original image 86K