It was revealed that OpenAI lost money this year! $8.5 billion burned

Compile | Li Shuiqing

Edit | Heart

Zhidong reported on July 25 that according to foreign media The Information, OpenAI may face a loss of up to 5 billion US dollars (about 36.3 billion yuan) this year, and it is expected to need to raise funds again soon.

In terms of cost, according to people familiar with the matter, OpenAI's model inference costs, training costs, and labor costs may combine to reach $8.5 billion this year. In terms of manpower, OpenAI originally planned to have 800 employees by the end of the year, but now its number has nearly doubled to 1,500, and the labor cost has reached $1.5 billion.

In terms of revenue, OpenAI recently reached about $283 million in monthly revenue, mainly thanks to ChatGPT and its API interface services, and is expected to reach $3.5-4.5 billion in revenue for the full year.

But OpenAI's financial situation is not bad compared to competitors such as Anthropic. According to The Information, although Anthropic's loss is about half of OpenAI's, its revenue is only 1/5 to 1/10 of OpenAI's.

This is partly due to Microsoft's discount on OpenAI's computing power, such as the rental price of each A100 server was exposed to be as low as $1.3 per hour.

Despite this, OpenAI's financial pressure is still not small, how long can tens of billions of dollars in financing be burned? Attract the attention of the industry.

1. Cost account: the cost of training is 7 billion US dollars, and the labor cost is 1.5 billion US dollars

As of March, OpenAI had spent nearly $4 billion this year on renting Microsoft servers to support ChatGPT and its underlying large model, also known as inference costs, according to an insider.

In addition to keeping ChatGPT running, OpenAI's training costs are expected to soar to $3 billion this year, including data fees.

According to insiders, OpenAI accelerated the training progress of the new AI last year, far exceeding the original plan. OpenAI originally planned to invest about $800 million in training costs, but the actual expenditure was much higher.

Foreign media estimate that as OpenAI trains a new version of the flagship model and sets up new projects, these costs will double this year.

In addition, OpenAI's team of about 1,500 people is still expanding rapidly, and labor costs are expected to reach about $1.5 billion. This is largely due to the fierce competition in the tech talent market from companies like Google.

Of course, the above is only an estimate. OpenAI originally planned to keep labor costs at $500 million in 2023 and expected to double its headcount to about 800 by the end of the year, according to people familiar with the matter. Since then, however, the number of employees has nearly doubled again, and based on nearly 200 vacancies on its website, it is likely that the expansion will continue in the second half of 2024.

To sum up, OpenAI's total operating costs this year could be as high as $8.5 billion.

2. Income account: Revenue may reach $4.5 billion, but still a loss of $5 billion

In terms of revenue, ChatGPT is expected to reach about $2 billion in annual revenue.

One of the challenges OpenAI faces, however, is that millions of users use its services for free, which dramatically increases computing costs without generating additional revenue. With Apple planning to introduce ChatGPT to the iPhone later this year, that cost could climb further.

But OpenAI doesn't rely solely on ChatGPT as a single source of revenue.

It charges developers a fee to access its technology platform to develop their own conversational AI applications or coding assistants, and the API service has generated more than $80 million in monthly revenue in March.

Overall, OpenAI's recent total monthly revenue is about $283 million, and based on this, its full-year revenue could be between $3.5 billion and $4.5 billion, depending on the market performance in the second half of the year.

If the potential cost of $8.5 billion is subtracted from the top estimate of $4.5 billion in revenue, OpenAI would face a loss of $4 billion to $5 billion.

This analysis explains why OpenAI CEO Sam Altman called the company "the most capital-intensive startup in Silicon Valley history."

This also means that OpenAI will soon need to raise money again in order to stay afloat. At the beginning of last year, Microsoft invested $10 billion in OpenAI, which generated about $1 billion in revenue that year, but its losses may have exceeded $2 billion.

So far, company spokespeople for both sides have not commented.

Third, the income is 5-10 times that of Anthropic, and the loss is twice as much

OpenAI may face a lot of losses, but its financial situation is still much better than that of its competitors.

In stark contrast to OpenAI is its strong rival, United States AI startup Anthropic. According to The Information's estimates, Anthropic's revenue is only 1/5 to 1/10 of OpenAI's.

However, in terms of losses, Anthropic may be in a better position, with a cash burn of more than $2.7 billion expected this year, but the loss could be only about half that of OpenAI.

Anthropic is expected to spend as much as $2.5 billion on computing costs this year, according to people familiar with the matter. At the same time, the startup expects to achieve annual revenue of about $800 million by the end of the year, or an average monthly revenue of $67 million.

However, it is worth noting that Anthropic has a revenue sharing mechanism with Amazon, which is both its financial and technical supporter and sells Anthropic's AI models to customers through its cloud services.

It's unclear what percentage of revenue will be achieved through the Amazon channel, but it's expected to be significant. This means that after deducting the share with Amazon, Anthropic's annualized revenue this year could be between $400 million and $600 million, or an average monthly revenue of $33 million to $50 million.

In short, while Anthropic is growing rapidly, it is nowhere near as efficient as OpenAI in terms of operational efficiency.

4. Each A100 server is $1.3/hour, and Microsoft has a large discount on computing power

OpenAI was able to outperform its opponents, largely thanks to the large discount on computing power given by Microsoft.

The Information's estimate of OpenAI's inference cost of about $4 billion per year comes from a person familiar with OpenAI's renting of server clusters from Microsoft.

According to the source, the computing power of the cluster is equivalent to equipping 350,000 NVIDIA A100 chips. Among them, about 290,000 chips, that is, more than 80% of the total capacity of the cluster, are dedicated to supporting the operation of ChatGPT.

Microsoft charges OpenAI quite favorably, at just $1.3 per hour per A100 chip server, which is much lower than what it charges other customers. According to people familiar with the matter, the inference cluster is operating at full capacity almost around the clock, so the annual cost estimate is about $4 billion. However, if OpenAI can reduce the cost of running the model on a server, this fee may be reduced.

Ion Stoica, co-founder of enterprise software company Databricks, said: "The industry consensus expectation is that inference costs will continue to fall significantly, which means that OpenAI is poised to make significant profits from its model services in the coming years. However, the cost of training AI is more complex, and companies are likely to continue to invest heavily in training larger models. ”

As for the size of OpenAI's server cluster for model training, it is unclear at this time. However, it was revealed that the computing power of the cluster in the first quarter was equivalent to 120,000 A100 chips, and Microsoft has not yet sent more training servers to OpenAI.

It is worth mentioning that Microsoft recently reached a partnership with Oracle to rent a huge server cluster to meet OpenAI's computing needs. The cluster is expected to be commissioned in mid-2025, when OpenAI will pay about $5 billion in rent over two years, as reported earlier by The Information.

Fifth, to improve the financial situation, OpenAI is not idle

OpenAI is working to reduce costs and increase revenue. The company has previously announced that it has significantly reduced the cost of running AI models by applying new technologies for building models.

At the same time, OpenAI also sees great potential to increase revenue, and plans to launch new products including search engines and computer usage agents. Once available, these products will be able to handle complex, multi-step tasks, such as helping shoppers research products or auto-filling expense reports and importing them into accounting software.

In addition, the company is confident that the next generation of its flagship will be released before the end of the year. If this new product is a significant improvement over GPT-4, which was launched in early 2023, OpenAI's business growth may be more than enough to make up for its current losses.

In another sign, as more businesses adopt OpenAI's conversational AI technology, the latter's revenue is growing even faster than costs, with monthly revenue surging 3.5 times to $283 million per month in the past year alone.

Conclusion: AI large-scale model startups are under great financial pressure and urgently need continuous blood transfusion

By calculating an economic account for OpenAI, we found that even the most lucrative AI model startups are facing severe financial pressure. OpenAI is sitting on Microsoft's financing of more than one million dollars, and its revenue may reach $4.5 billion this year, but it will lose 5 billion, not to mention other AI startups.

In the next step, OpenAI may continue to raise funds to maintain operations. For more AI startups, how to promote the upgrading of technologies, products, solutions and services, reduce costs and increase efficiency has become an unavoidable topic.