How to Lose $5 Billion in One Year?
Regarding the cost structure, according to informed sources, as of March this year, OpenAI has spent nearly $4 billion on renting Microsoft's server resources to fully support the operation of ChatGPT and its underlying large models. This part of the cost is also known as inference cost.
In addition to the cost of running ChatGPT, OpenAI's training costs (including payment for data usage) are also showing significant growth, expected to climb to about $3 billion this year. It is reported that the company advanced the training of new AI models at an unexpectedly fast pace last year, causing the originally planned $800 million budget to be greatly exceeded. The training cost is expected to double this year as OpenAI continues to iterate its flagship large models and begins training the next generation of models.
In terms of human resources, OpenAI's expansion is equally rapid, with about 1,500 current employees, and the team size is still growing. To attract and retain top technical talent, OpenAI's investment in salaries and benefits is not to be underestimated, with labor costs expected to reach $1.5 billion this year.
According to informed sources, OpenAI had expected its labor costs for 2023 to reach $500 million. By the end of that year, its employee count had doubled to around 800. Since then, the company's employee count has nearly doubled again. With nearly 200 job openings posted on its website, this indicates that more employees will be joining in the second half of 2024.
Combining these costs, OpenAI's total operating costs this year are expected to reach a staggering $8.5 billion.
On the revenue side, although ChatGPT has shown strong potential with its annualized revenue approaching $2 billion, a major challenge for OpenAI is that many users tend to use the free version of ChatGPT. This undoubtedly exacerbates its computational cost burden without translating into direct revenue. Notably, as Apple plans to direct iPhone users' queries to ChatGPT, this trend may further drive up OpenAI's operating costs.
Nevertheless, OpenAI has other profitable avenues. By providing developers access to its large language models (i.e., API services) and allowing them to build their own generative AI applications or coding assistants based on this, OpenAI has successfully opened up a stable source of income. As of March this year, this API business was bringing in $80 million in monthly revenue for OpenAI.
Overall, OpenAI's current total monthly revenue is about $283 million, suggesting that its full-year revenue range is likely to fluctuate between $3.5 billion and $4.5 billion, depending on market performance in the second half of the year. However, even with the most optimistic revenue forecast, facing potential costs of $8.5 billion, OpenAI may still face a huge loss of $4 billion to $5 billion.
These figures respond to comments made earlier by OpenAI CEO Sam Altman, who described the company as "the most capital-intensive startup in Silicon Valley history." Given the current loss situation, OpenAI may soon initiate a new round of fundraising. It's worth noting that Microsoft committed to invest $10 billion in OpenAI at the beginning of last year, but OpenAI may have already lost $2 billion this year, although it also generated about $1 billion in revenue during the same period.
Spokespersons for both companies currently remain silent and have not made further comments.
Anthropic vs OpenAI
Although OpenAI consumes a lot of funds, its financial situation is still in an advantageous position compared to some competitors. This is due to Microsoft's special support - OpenAI enjoys significant discounts on renting energy-intensive servers equipped with expensive NVIDIA GPUs, basically only needing to bear Microsoft's server operating costs.
In contrast, OpenAI's number one rival in the startup field, Anthropic, faces more severe challenges. It is estimated that Anthropic's revenue is only one-fifth to one-tenth of OpenAI's, but its loss amount is expected to be controlled at about half of OpenAI's. Earlier this year, Anthropic's executives predicted that the company would consume over $2.7 billion in cash throughout the year, with computational costs alone amounting to $2.5 billion.
Although Anthropic's projected annualized revenue could reach about $800 million by the end of the year (i.e., $67 million per month), its partnership with Amazon has diluted these revenues to some extent. As a provider of funds and technology, Amazon not only promotes Anthropic's AI models to its customers but also takes a cut. Although the specific proportion is unclear, it can be foreseen that this cooperation model has a significant impact on Anthropic's net income, resulting in its projected post-dividend annualized revenue range possibly reducing to between $400 million and $600 million, or $33 million to $50 million per month.
In other words, although Anthropic shows strong momentum in growth rate, it is clearly unable to compare with OpenAI in terms of operational efficiency and cost control.
$4 Billion to Microsoft Every Year?
The estimated $4 billion annual inference cost for OpenAI comes from an insider familiar with the details of its server cluster rental from Microsoft. Reportedly, this massive server cluster consists of about 350,000 servers equipped with NVIDIA A100 chips, of which over 80% (about 290,000) are fully supporting the operation of ChatGPT.
Microsoft's charging strategy for OpenAI is extremely favorable, with each A100 server costing only $1.30 per hour, far below market price, mainly due to the close cooperation between the two parties. Insiders say that given the inference cluster is operating at full capacity for a long time, the annual operating cost is thus locked at about $4 billion. It's worth noting that if OpenAI can effectively reduce the cost of running models on these high-performance servers, its overall expenditure may decrease.
Ion Stoica, co-founder of enterprise software company Databricks, is optimistic about this, believing that with continuous technological advancements, inference costs are likely to decrease significantly, meaning OpenAI may make considerable profits from providing services for its models in the coming years. In contrast, the cost of training AI models is more complex and variable, and companies may need to continue investing heavily to nurture more powerful models.
As for the scale of the server cluster Microsoft provides for OpenAI's model training, specific details are still unknown. But it's revealed that before Microsoft increased its training resources, the first quarter was equivalent to configuring about 120,000 A100 servers. Additionally, Microsoft recently reached a major agreement with Oracle to lease a massive server cluster to further meet OpenAI's growing computational needs. According to reports, this cluster is expected to come online in mid-2025, and OpenAI will invest an additional $5 billion in its lease over the next two years.
Improvement in OpenAI's Financial Situation
OpenAI is actively seeking ways to cut costs and generate more revenue. The company announced that by adopting cutting-edge model building technologies, it has helped significantly reduce the cost of running AI models and is expected to generate more revenue.
On this basis, OpenAI plans to launch a series of innovative products, including advanced search engines and intelligent agents capable of executing complex computer tasks, aimed at addressing diverse, multi-step task challenges, such as assisting shoppers in precise product searches, or automating expense report filling and integrating it into accounting software.
Looking ahead, OpenAI is full of expectations for its next-generation flagship model set to debut at the end of this year. This model is expected to far surpass the current star product GPT-4 and is likely to become a key force in driving the company's business growth and making up for losses.
More encouragingly, as generative AI technology becomes increasingly popular in the business world, OpenAI's revenue growth has significantly outpaced its cost growth. Over the past year, the company's monthly revenue has achieved a 3.5-fold increase, climbing to $283 million.