OpenAI Reaches $1 Billion in Annual Revenue: How Far Has Large Model Commercialization Come?
-
OpenAI has evolved once again.
At midnight Beijing time on September 21, OpenAI announced the release of DALL·E3, the third version of its text-to-image AI tool, 17 months after the previous version.
It is reported that DALL·E3 will integrate with ChatGPT, allowing users not only to design AI-generated images using prompts but also to modify the generated images through conversation. This enhances both the prompt-based image generation experience and the model's ability to understand user instructions, while also significantly improving image quality.
In simple terms, DALL·E3 makes it easier for users to translate their ideas into accurate images, bringing AI image generation closer to the ChatGPT experience. Even if users don't know how to craft prompts, they can simply input their ideas, and ChatGPT will automatically generate detailed prompts for DALL·E3.
DALL·E3
OpenAI CEO Sam Altman stated that DALL·E3 will be available to all ChatGPT Plus and enterprise users in the coming weeks, with plans to release it to research labs and API services in the fall. However, OpenAI did not mention whether a free public version would be released.
Unintentionally, OpenAI has added another revenue stream. Prior to this, Silicon Valley's authoritative media The Information reported that OpenAI's annual revenue has exceeded $1 billion.
According to The Information, OpenAI incurred a loss of $540 million in 2022 but is expected to generate $1 billion in revenue this year. In just 10 months, OpenAI has transitioned from losses to profitability, demonstrating highly efficient commercialization progress.
Burning $700,000 Daily
When OpenAI launched ChatGPT, it led the advancement and development of the AI era, but its profitability was widely doubted by the outside world.
Several months ago, a report by Analytics India Magazine stated that OpenAI spends approximately $700,000 daily just to operate its AI service, ChatGPT. OpenAI is currently in a cash-burning phase, and without accelerating its commercialization efforts, it might face bankruptcy by the end of 2024.
Most media outlets' skepticism toward OpenAI is not without reason. Silicon Valley's authoritative media, The Information, reported in August this year that OpenAI generated $28 million in revenue in 2022 but suffered a $540 million loss.
In the past few years, especially before the explosive popularity of GPT-3.5, OpenAI was purely a non-profit company. All its revenue came from investments by its 'big backer,' Microsoft.
Since 2019, Microsoft has invested $1 billion in OpenAI and later increased its investment to over $15 billion. Most of these funds were spent on building AI supercomputers and purchasing AI chips for hardware expenditures.
According to Microsoft executives, the AI supercomputer powering ChatGPT was built in 2019 with a $1 billion investment. This large-scale, top-tier supercomputer is equipped with tens of thousands of Nvidia A100 GPUs, supported by over 60 data centers deploying hundreds of thousands of Nvidia GPUs in total.
Beyond hardware costs, training ChatGPT also requires significant investment. The smarter ChatGPT becomes, the greater the computational power needed.
OpenAI data shows that training a 1.3-billion-parameter GPT-3XL model requires approximately 27.5 PFlop/s per day, while training a 174.6-billion-parameter GPT-3 model demands about 3,640 PFlop/s daily.
Guosheng Securities estimates that in January 2023, ChatGPT averaged 13 million daily unique visitors, corresponding to a demand for over 30,000 Nvidia A100 GPUs. The initial investment cost was around $800 million, with daily electricity expenses reaching approximately $50,000.
Over $15 billion is needed to cover the high costs of training and developing large models. In addition to hardware, computational, and training expenses, OpenAI's labor costs remain persistently high.
OpenAI employs 375 full-time staff in San Francisco, most of whom are leading experts in machine learning, with annual salaries exceeding $200 million. According to foreign salary surveys, the average income for an OpenAI software engineer is $920,000.
Despite Microsoft's support, OpenAI faced severe financial strain due to exorbitant hardware costs, computing expenses, and labor expenditures. Its ChatGPT project, often described as a 'money-burning beast,' was widely perceived as 'struggling for survival.'
Yet within months, the company that was reportedly losing $700,000 daily has not only stabilized but is now projected to achieve $1 billion in profits.
Rapid Commercialization
The Information recently revealed that OpenAI expects to generate over $1 billion in revenue within the next year through AI software sales and computational services. This revelation has stunned media outlets that previously predicted OpenAI's imminent bankruptcy.
A closer examination shows that since ChatGPT-3.5's global breakthrough on November 30 last year, OpenAI has actively pursued commercialization rather than passively waiting for competitors to catch up. On February 2, 2023, the company launched ChatGPT Plus, a $20/month subscription service offering priority access during peak hours.
On April 12, 2023, OpenAI announced the launch of ChatGPT Business, a paid subscription service designed to allow enterprises to use their services while maintaining control over their data.
Recently, OpenAI also revealed plans to release a beta version of its enterprise edition, commonly referred to as the 'custom version,' which will offer more customization options for businesses of all sizes.
After establishing pricing tiers for Plus, Enterprise, and Business versions, OpenAI has adjusted GPT-4 access limits multiple times to boost revenue. Initially set at 150 messages per 4 hours, it was later reduced to 25 messages per 3 hours and then increased to 50 messages per 3 hours in late July. Additionally, OpenAI has upgraded GPT models and reduced prices, including for embedding models and GPT-3.5-turbo.
With these strategic moves, OpenAI has achieved a strong start in its commercialization efforts.
According to OpenAI, as of July, ChatGPT Plus has attracted 2 million paying users in the consumer market. In the business sector, over 80% of Fortune 500 companies have adopted the enterprise version of ChatGPT. Major corporations like Block, Canva, Estée Lauder, and PwC have already tested the custom beta version of ChatGPT, with positive feedback.
It's worth noting that OpenAI's revenue streams are diverse and not limited to ChatGPT. For example, fine-tuned models like babbage and davinci, the vectorization model Ada v2, the speech model Whisper, and the text-to-image model DALL·E3 mentioned earlier each serve different purposes and have varying pricing models.
CEO Sam Altman stated earlier this year, 'OpenAI is projected to achieve $200 million in annual revenue in 2023, with a target of $1 billion by 2024.'
Based on the current base of 2 million paying users and 80% of Fortune 500 companies, OpenAI is on track to complete its 2024 annual plan this year.
Gaming is a new direction
Beyond being an entrepreneur, OpenAI is also a seasoned investor. Since 2015, OpenAI has been investing in AI startups and organizations through its venture capital fund. According to The Information, as of January this year, OpenAI has invested in at least 16 companies.
On the 17th of last month, OpenAI announced on its official website the completion of its acquisition of a young AI startup, Global Illumination, integrating its entire team. This transaction marks OpenAI's first publicly disclosed acquisition since 2015.
So why is Global Illumination so special? The reason is that it is not only an AI company but also a gaming company.
Global Illumination independently developed a game called "Biomes," which is not a popular mobile game but a browser-based web game. Its style closely resembles the classic sandbox game Minecraft.
Minecraft is the best-selling video game of all time, with 238 million copies sold and 150 million monthly active users as of May 2021.
The gaming industry has always been closely intertwined with AI models. Games provide a stable training environment for AI, enhancing its decision-making capabilities. As early as 2016, OpenAI used the real-time strategy game Dota 2 to train AI.
OpenAI can dynamically adjust terrain, vegetation, animals, weather, and other elements to generate diverse and realistic game worlds, training more powerful AI models. This further drives the evolution of GPT models, ultimately advancing toward Artificial General Intelligence (AGI). Simultaneously, AI can enhance games with greater intelligence, offering players richer experiences.
From another perspective, OpenAI's acquisition of a gaming company and potential development of a game demonstrates its strong financial position, dispelling any concerns about bankruptcy risks.
Whether gaming can become the next growth driver after ChatGPT depends on the types of gaming projects OpenAI pursues in the future.