How does OpenAI make money?

As the leader of the entire AI big model industry, OpenAIIIIIIII's business model determines the subsequent AI capital expenditure of other enterprises, and its profitability also provides reference for other AI enterprises. Founder Securities pointed out in its July 21 report that the overall logic of OpenAIIIIIIII profit is to increase the proportion of GPT-4 paid and reduce GPT-3.5 costs. * * In the case of GPT-3.5 cost compression, if the daily living and monthly living ratio reaches 35%, the monthly living payment rate exceeds 12%, or can achieve break-even. For cost-compressed GPT-3.5 and GPT-4 models, a 0.5 per cent monthly increase in the monthly rate may turn around. **** If the OpenAIIIIIIII is profitable, it may stimulate all overseas giants to increase capital spending. * * For the entire AI ecology, driven by OpenAIIIIIIII charges, or help users to develop a sense of payment, all kinds of AI-related applications are expected to blossom. ## Cost Analysis: The main cost OpenAIIIIIIII by "10% accuracy vs 50% calculation force" is calculation force cost. OpenAIIIIIIII may control the cost by sacrificing accuracy. The study pointed out that for every 10% decrease in model accuracy, the calculation force may be halved. According to Founder Securities, the relationship between calculation force and accuracy: according to the paper research of YufanLiu et al., the calculation force may be halved for every 10% decrease in the accuracy of the model. * * The size of the model is determined by its parameter quantity and its precision. The precision is usually FP64, FP32, FP16, BF16, TF32, INT8, IN4, etc. The decrease in precision will not only expand the load of the calculation force, but also lead to a decrease in performance to a certain extent.>> ! > > According to the paper by TimDettmers et al., the team used a large model GPT-4 as a referee, scoring the answers of different models, with the GPT-3.5 score as 100 percent, and finally GPT-4 its own score of 114.5 percent. MichalKosinski studies have shown that GPT-3 can solve 70% of the theory of mind tasks, while GPT-3.5 solve 93% of the tasks, with GPT-3.5 grades as a 100, and the GPT-3 score is about 75(70/93* 100). Therefore, with different accuracy of the model, Founder Securities points out:> Under the assumption of 25 million DAU, if each user uses 10 times, the number of GPUs required for GPT-3.5 is about 272000 A100. If the calculation force is divided by 2 according to the logic that the accuracy is reduced by 10%, it can be calculated that the calculation force required for the GPT-4 is about 664000 (27.2GPU/(2 log0.9(114.5/100)) under the conditions of 25 million daily life and 10 average usage times.>> ! For OpenAIIIIIIII, Founder Securities pointed out that GPT-3.5 is the main source of cost, and the calculation cost can be reduced by reducing accuracy:> Assuming that the ChatGPT has 60 million DAU and the average number of requests per user is 8, the overall calculation cost of uncompressed GPT-3.5/GPT-4 for the whole year is about US $6.55 billion.>> * * If the accuracy of the GPT-3.5 drops to 95.7 percent of the original, the cost drops by about 25 percent, and the overall power cost of GPT3.5 is about $4.58 billion under the condition of 60 million DAU and an average of 8 requests per user. ** > > * ! * ## Payment Analysis: Multimodal and Third-Party Plug-ins Push Users to Pay Active Users and Pay Scale:> As of July 12, 2023, the daily traffic of ChatGPT web pages has basically remained flat at over 50 million. As of June 19, 2023, the average daily active users in the OpenAIIIIIIIIChatGPTiOS-end U.S. region for the first 30 days were 946000.>> ! > > according to data.ai data, as of June 19, 2023, the daily living payment rate for the first month of ChatGPT iOS launch was about 4.36%; If the daily living and monthly living ratio is 37% in ChatGPT, the monthly active number is about 2.556 million (94.64 ÷ 37%), at this time, the monthly live payment rate (monthly paid users/monthly active users) is about 1.6 (4.13 ÷ 255.8).>> ! On how to improve the payment rate:> **GPT-4 multi-modal functions and CodeInterpreter provide users with work assistance and promote user payment. **On July 11, 2023, OpenAIIIIIIII launched the code interpreter (Codeinterpreter) function, whether it is code writing or data analysis, chart generation, performing mathematical operations, even generating videos and analyzing the stock market, you can get high-quality results.>> **GPT-4 a unique third-party plug-in and" code interpreter "function to drive users to pay. * * In May 2023, OpenAIIIIIIII fully opened GPT-4 third-party plug-in functions (plugins), which are only open to Plus users and promote users' payment. On July 11, 2023, the beta version of ChatGPT "code interpreter" (CodeInterpreter) was officially opened to all Plus users, specifically including functions such as helping to analyze data, create charts, edit files, perform mathematical operations, etc. The code interpreter expanded the functions of ChatGPT, bring users a better interactive programming experience and powerful data visualization capabilities, which may drive users to pay.>> ! ## Profit and Loss Line: 35% Daily Live Ratio 12% Payment Rate From the overall logic, the key to OpenAIIIIIIII a turnaround is to increase the GPT-4 payment ratio and reduce GPT-3.5 costs. Founder Securities pointed out that in the case of GPT-3.5 cost compression, the 12% payment rate may be the profit and loss line. * *> the proportion of daily and monthly life is 35%, and when the payment rate exceeds 12%, the annual income may reach 4.94 billion USD/year;>> if the accuracy of the GPT-3.5 decreases by about 5% and the cost decreases by about 25%, the overall computing cost of Open AI is about us $4.74 billion under the condition of 60 million DAU and an average of 8 requests per user.>> ! Judging from the growth trend of monthly payment rate, Founder Securities pointed out that the monthly payment rate increase of 0.25 percent per month may not be sustainable, and the monthly increase of 0.5 percent may turn a loss.> In the case of the OpenAIIIIIIII $10 billion, sensitivity hypothesis analysis: if the cost-compressed GPT-3.5 model and the total GPT-4 power cost of $10 billion may be lost to January 2026 under the condition of 60 million DAU and a 0.25 percent monthly increase in the payment rate.>> **If the cost-compressed GPT-3.5 model and GPT-4 overall are likely to turn a profit in November 2025 under the condition of 60 million DAU and a 0.5 per cent monthly increase in payment rates. In addition, Founder Securities from a single user break-even analysis:> If the daily throughput of a single user is about 7-8 times per GPT-4, the daily live-to-live ratio is about 30%-35%, in a profitable state. * * > > if the daily living and monthly living ratio reaches 50%-55%, when the user uses more than 14 times, there may be a loss. After the introduction of GPT-4 access restrictions, the use of "high-frequency users" is limited, and the average number of requests may drop.>> ! ## OpenAIIIIIIII profits or stimulate giants to increase capital expenditure It is worth mentioning that OpenAIIIIIIII the profit and loss situation or stimulate other giants to increase "calculation capital expenditure":> * * 1, "search scenario company": * * Google, Baidu Google: Google search scenario is similar to OpenAIIIIIIII, if OpenAIIIIIIII to achieve profitability, Google can add member payment to Bard, or increase its calculation capital investment. Baidu: Baidu's search scenario is also similar to OpenAIIIIIIII's. OpenAIIIIIIII's profit may affect Baidu's capital investment and AI function may be added to its search.>> 2." Social Scene ": Meta, Tencent Meta:Meta has social scenes. If OpenAIIIIIIII can make profits, Meta may add chat robots to social scenes. Tencent: Tencent has social scenarios such as WeChat, and if OpenAIIIIIIII can make a profit, Tencent may also be able to access AI features in social scenarios such as WeChat, affecting its capital expenditure.

• • 3, "cloud computing vendors" * *: Microsoft, Ali Microsoft: Microsoft mainly provides API services. if OpenAIIIIIIII with strong support from Microsoft's cloud computing resources can make profits, the profit expectation of third-party AI applications will be affected, and Microsoft will continue to increase capital investment in the future. Ali: Ali, as a cloud computing vendor, also provides API services. If OpenAIIIIIIII is profitable, the profit expectation of third-party AI applications will also increase, which may increase the demand for API and may also affect Ali's capital expenditure.>> * * 4. Third-party companies accessing API: * * Only when the OpenAIIIIIIII achieves a break-even can we further increase the investment in API. The demand for computing power is first provided to oneself. The users of one's own application are direct users of the OpenAIIIIIIII, with strong stickiness, while the users of third-party applications are indirect users, with weak stickiness. Therefore, direct users are given priority, and third-party companies that access the API are considered second. Therefore, only when profits OpenAIIIIIIII be realized can further increase the investment in API be affected.