[Top left] Click Follow
Wen 丨 Hao Xin Liu Yuqi
Editor丨Liu Yuqi
"I saw him less than three minutes and thought, ah, 19-year-old Bill Gates guess that's it!"
In the eyes of YC founder Paul Graham, 28-year-old Sam Altman, like 19-year-old Bill Gates, is a school dropout, maverick, and believes that the world can be changed, altruism and extreme ambition, but Altman is more radical, and his ambitions even exceed the boundaries of Silicon Valley.
In 2014, Altman was selected by Graham to serve as CEO of the startup incubator. As we all know, after Altman founded OpenAI, he found his second Bole: Nadella, the current CEO of Microsoft.
In the five years since Altman took over as CEO of YC Incubator, he has opened up a number of new businesses, led YC's total market capitalization to about $150 billion, and invested in more than 4,000 startup parks and more than 1,900 companies. Today, OpenAI, which has been in the two-way extreme evaluation, has also preliminarily handed over its report card.
On August 30, according to foreign media "The Information", OpenAI expects to earn more than $1 billion in revenue from the sale of artificial intelligence software and its computing power in the next 12 months.
As soon as the news came out, there was an uproar.
After all, as early as 3 months ago, OpenAI was still struggling in the "life and death line", Indian news media platform Analytics India Magazine said that OpenAI only runs its artificial intelligence service ChatGPT costs about $700,000 a day, OpenAI is currently burning money, if it does not accelerate its commercialization process, it is likely to have to file for bankruptcy at the end of 2024.
Public data shows that in 2022, OpenAI revenue was about $36 million, but this year, they spent $544 million, which means that last year alone, a net loss of $500 million.
OpenAI, which has always been known as a "gold-swallowing beast", suddenly began to have large-scale revenue, which made everyone wonder: What kind of money-making trick has Altman changed?
More importantly, OpenAI is a barometer of the commercial potential of large language models, exploring the mystery of OpenAI revenue, but also unveiling the long-awaited "Pandora's box" of general-purpose large model commercialization.
For the entire industry, boosting confidence is on the one hand, on the other hand, once OpenAI's business model runs through and sets a new template for the industry, it is expected that soon, there will be another wave of "100-model commercialization war", pushing AGI to enter the second stage.
The business landscape of OpenAI
From the moment OpenAI launched ChatGPT, it set its sights on commercialization with a hunter-like viciousness.
On November 30 last year, GPT-3.5 was born. Just over two months later, OpenAI quickly opened the charging model and launched the ChatGPT Plus subscription plan.
Since May this year, OpenAI has gone further and further on the road of business, frequently releasing "big moves":
- On May 15, the ChatGPT iOS app was launched.
- On May 31, GPT-4 third-party plugins were fully opened.
- On June 21, it was revealed that it planned to launch a large model store similar to the "App store".
- On June 23, it was revealed that it planned to launch a ChatGPT version of "personal work assistant".
- On August 29, OpenAI released features that fully benchmark Bing Chat's enterprise version of ChatGPT.
(Source: OpenAI official website)
According to the official website of OpenAI, at present, its products are mainly divided into two categories: one is API-based products, including callable GPT models, DALL· E model (Venzo graph model); Whisper (speech recognition model) and Chat (conversation), Embeddings (vectorization), Analysis, Fine-tuning (fine-tuning) functions for developers; One category is ChatGPT conversational bot as the core of the product, divided into personal version and enterprise version.
Based on the information on OpenAI's official website and the collation of public data, Light Cone Intelligence found that OpenAI currently has two main revenue pillars.
The first is the per-API charging model that OpenAI relies on most since the birth of ChatGPT. In this mode, users can almost use the multimodal capabilities developed by OpenAI throughout the underlying large-language model, model deployment, model development and other processes, and the price is also very friendly, each call only a few cents. OpenAI officially does not specifically divide whether users are individual users or enterprises, but according to foreign media reports, in addition to a large number of individual users, Jasper, Slack, Salesforce, Morgan Stanley and other well-known companies are its early users.
It is worth mentioning that under this charging model, OpenAI also provides its largest "gold owner" Microsoft, including coding, GPT-4, Wensheng Diagram, ChatGPT and other functions, and integrates into Microsoft cloud services, search, office software and other products. How much OpenAI can benefit from it is unknown, but in the case of the Azure cloud business, the cost of Microsoft's use of the above OpenAI features is consistent with the offer, and all of OpenAI's technology runs for free on Microsoft's Azure cloud infrastructure.
The second is a subscription charging system based on ChatGPT products. In the early days, OpenAI obtained a large amount of training data for free, and with this, TikTok and Instagram user growth data were refreshed within 9 months, becoming the fastest app to reach 100 million users.
"Making a wedding dress" for Microsoft and "tasting" users is not the ultimate goal of OpenAI. If you want to be profitable, you need to find a way to increase your payment rate. In June, the number of ChatGPT users crossed the peak and declined, and OpenAI began to shift its business thinking from the C side to the B side in an attempt to "grab" the business of the gold owner.
OpenAI said that many large enterprises are interested in its enterprise-level new products, which have been adopted by more than 80% of Fortune 500 companies since the launch of ChatGPT, including Block, Canva, Estée Lauder, PricewaterhouseCoopers and other large enterprises have also tried the beta version of ChatGPT Enterprise in advance. Since then, OpenAI will also launch a business version of ChatGPT for smaller organizations with more customization options.
According to Light Cone Intelligence, since the launch of ChatGPT, the most anticipated is the enterprise version. After being criticized for "data privacy security", OpenAI made adjustments to the product.
ChatGPT Enterprise Edition is currently powered by OpenAI's most advanced language model, GPT-4, and enterprise users have priority access to GPT-4 and remove the usage limit, and the execution speed is twice as fast as ordinary GPT-4. In addition, the Enterprise version allows more content to be entered, and the context window is expanded to 32,000 tokens and about 25,000 words.
OpenAI promises that customer prompts and all other data will not be used for model training, that users can control how long data is retained, and that any deleted conversations will be automatically deleted from ChatGPT's system within 30 days.
In addition, in terms of deployment, the Enterprise Edition provides a new management control platform that can manage users in batches, including single sign-on, domain verification, and dashboards with usage statistics, which is suitable for massively scalable deployments. At the same time, it also adds a full set of toolchains such as vectorization tools and advanced data analysis tools.
From this point of view, OpenAI is trying to move from the previous low-fee, low-frequency API model of charging by tokens to a diversified charging model of high pricing, high-sticky 2B subscription fees and customized solution charges.
The more you grow, the more you lose?
Revenue growth does not mean that OpenAI really starts to make money, after all, $1 billion in OpenAI's investment costs can only be "sprinkled water". And after the full commercialization, as OpenAI users and GPT4 continue to study the demand for computing power, the cost will continue to rise with the scale of users, and many technology companies are difficult to get rid of the curse of "revenue growth and loss".
The high cost of OpenAI is obvious to all. According to the intelligent combing of light cones, the current cost investment of OpenAI is mainly divided into the following parts:
- Cost of talent: OpenAI has 375 permanent employees in San Francisco, most of whom are machine learning bigwigs, and pay them $200 million a year alone. According to a survey by foreign salary websites, the median salary of OpenAI software engineers is $920,000.
- Training costs: According to data, they spent $4.6 million training GPT3, and the corresponding cloud resource cost is almost 9 figures (that is, hundreds of millions). According to SemiAnalysis, a semiconductor consulting and research firm, if the cost of OpenAI cloud computing is almost $1 per A100 hours, then under such conditions, the cost of just one training is about $63 million, which does not include all experiments, failed training and other costs, such as data collection, RLHF, labor costs, etc.
- Inference, operating costs: Forbes reported that the operating expenses or inference costs of ChatGPT's large language models "far exceed the cost of training when deploying any reasonably sized model," and "in fact, the inference cost of ChatGPT exceeds the cost of training every week."
- Investment: At the beginning of the year, OpenAI invested in at least 16 companies through a $100 million startup fund backed by Microsoft and other investors, and its accelerator, Converge, invested in 10 companies. Light Cone Intelligence also found that in the first half of this year, OpenAI publicly invested in the name of enterprises 3 times, and before it was Microsoft, Google, Nvidia and other established companies.
- Acquisition: On August 17, OpenAI announced the acquisition of a game company called Global Illumination, which is said to be OpenAI's first public acquisition since its inception in 2015.
According to public information, since the establishment of OpenAI, it has received more than 15 billion US dollars from investment alone, which is used to fill the "hole" of high-cost training and development of large models.
To the destination of AGI, OpenAI does need money, but "burning money" is like a bottomless pit, and "bleeding" may not be exchanged for growth, which is why OpenAI has to accelerate the commercialization process.
But having revenue doesn't mean turning a profit. Founder Securities has carefully calculated the relevant indicators of ChatGPT based on public data, and pointed out in the analysis that the general logic of OpenAI's profit is to increase the proportion of GPT-4 payment and reduce GPT-3.5 costs, which is the most important source of cost for OpenAI.
In the case of GPT-3.5 cost compression, if the daily active monthly activity ratio reaches 35%, and the monthly active payment rate exceeds 12%, it may be able to achieve breakeven. For the GPT-3.5 model and GPT-4 model after cost reduction, if the monthly payment rate is increased by 0.5% per month, it may be able to reverse the loss.
As of July 12, 2023, the daily number of ChatGPT page visits has remained basically flat, maintaining at more than 50 million. As of June 19, 2023, OpenAIChatGPT iOS had an average of 946,000 daily active users in the first 30 days.
data.ai data shows that as of June 19, 2023, the average daily active users of ChatGPT iOS from May 21 to June 19 in the United States were about 946,400, and the cumulative number of paid users was about 41,300. Therefore, the daily active payment rate (monthly paying users/daily active users) is about 4.36% (4.13÷94.64). According to Questmobile data, the proportion of daily active monthly active users of Baidu APP is about 37%, so if the ChatGPT daily active monthly active ratio is 37%, the number of monthly active users is about 2.5578 million (94.64÷37%), and the monthly active payment rate (monthly paid users/monthly active users) is about 1.61% (4.13÷255.78).
The above data shows that the commercialization of OpenAI is still difficult and long, and in the future, only by increasing the user payment ratio and reaching a certain level can we achieve breakeven, and the revenue alone cannot explain the problem.
OpenAI and Microsoft "grab food"
The business model is good, but problems arise. The once "hidden worry" has been directly moved to the front of the stage, that is, the delicate relationship with Microsoft.
Need to make it clear that no matter how sweet the "honeymoon period" OpenAI and Microsoft have had, these are two independently operated institutions and companies, although the cooperation between the two is more special than the relationship between acquisition or investment, and even Nadella did not hesitate to cut down some businesses to build an intelligent computing center for OpenAI, but once OpenAI began to commercialize independently, it means that the two have to share the same cake, and war and fight cannot be avoided.
Unlike the competition, OpenAI and Microsoft are closer and more complex.
In 2019, OpenAI transitioned from a non-profit to a hybrid model, with one called OpenAI LP responsible for commercializing the Open AI Research Lab's development products. That same year, Microsoft invested $1 billion in the partnership through its Microsoft Azure AI supercomputing technology as the infrastructure for training GPT models. In 2022, this partnership will be further strengthened, with Microsoft investing $10 billion in OpenAI.
(OpenAI LP, a for-profit entity established by OpenAI)
According to a person familiar with the terms of the partnership, after OpenAI repays its first investors, Microsoft will receive 75% of the profit until its main investment is repaid, after which it will receive 49% of the profit until the theoretical upper limit is reached. At the same time, according to people familiar with the matter to foreign media, starting around 2025, the upper limit of profit sharing will be increased by 20% every year, instead of setting a hard upper limit on profit sharing - basically their return on investment. Investors briefed on the deal say Microsoft could effectively own more than a third of the company.
Since March this year, Light Cone Intelligence has also learned from many companies that if they want to use the ability of ChatGPT, there are two ways in front of them: one is to directly call OpenAI's API interface and pay OpenAI according to the number of tokens used; Second, on Azure, based on the public cloud computing power to use OpenAI services, compared with the former, Azure services are more secure, the price is flat, the supporting facilities are more perfect, which is also the "difference" that Microsoft cloud sales staff has been emphasizing.
But through the survey, even so, there are still a large number of enterprises choose to directly access API services, one of the reasons is that it is simple and convenient, the pay-as-you-go model is more cost-effective for individual developers and small and medium-sized enterprises, and more importantly, the logic of calling APIs and purchasing cloud services is completely different, the former developers can decide by themselves, while the latter need to be approved and reported and consider the combination with the business.
As mentioned earlier, in the eyes of the outside world, the two sides are complementary, Microsoft provides funds, resources and technical support to OpenAI, and OpenAI helps Microsoft become the top technology giant again, but this situation has also changed with the gradual commercialization of OpenAI in 2023.
In June, according to The Information, an internal Microsoft document instructed Azure salespeople to tell customers that Microsoft could offer more than OpenAI. OpenAI, on the other hand, defended itself by delaying access to Microsoft's product APIs and suspending the provision of the latest models. This is the first time that the subtle rift between the two sides has been shown.
When invested in cruel market competition, conflicts of interest between the two sides are inevitable, and a sales pitch can not prove anything, and the strategy behind it can really explain the problem.
For example, in March this year, after OpenAI preemptively signed contracts with companies such as Snap and Instacart, Microsoft's cloud services took a week to announce a preview of ChatGPT's features; After the release of the GPT-4 language model, Microsoft's cloud services were only after companies such as Duolingo and Stripe had paid directly to OpenAI, that Microsoft's cloud services gained access to GPT-4.
The technology is in the hands of OpenAI, which must be shared with Microsoft, but there is no specific description and regulation of the time, node and extent of sharing, and OpenAI wants to use this "time difference" to grab some benchmark customers.
To put it bluntly, how OpenAI handles the relationship with Microsoft, the two sides delineate boundaries, set standards, find ways and balances for cooperation, and once handled carefully, it is likely to be a common crisis for both sides.
However, in the business world, there are no eternal friends or eternal enemies, only eternal interests, and in the face of interests, smart "ghosts" like Altman and Nadella, people look forward to how they can create miracles again, rather than falling into the cliché of "tearing".
#OpenAI# #微软#