DeepSeek’s Annualized Revenue Hits $1 Billion as Startup Finalizes $7.5 Billion Fundraising

DeepSeek’s annualized revenue run rate has hit $1 billion, more than double from less than $500 million a few months ago, buoyed by a recent price hike as well as continued popularity of its models, according to two people with direct knowledge of the matter.

The latest revenue figure, shared by CEO Liang Wenfeng in a recent meeting with investors, could boost investor confidence as the company finalizes its second funding round and as it prepares to go public in the Shanghai Stock Exchange. DeepSeek is aiming to complete the fundraising by the end of October, with a goal of raising 50 billion yuan ($7.5 billion) at a valuation of 500 billion yuan.

DeepSeek’s revenue growth was driven in part by a price hike last month, which increased the costs of its models for customers by 2.3 to 4.5 times. Despite the price hike, DeepSeek’s models continue to cost among the lowest in major language models, and Liang said at the meeting that the price hike didn’t cause a decline in DeepSeek’s customer base, as user demand remains strong, according to the people.

DeepSeek’s V4.1-Flash, released earlier this month, and the older V4-Flash, boast capabilities comparable to much larger models and have attracted customers around the world looking for AI systems that can power autonomous agents without breaking the bank.

While DeepSeek’s revenue is still minuscule compared to U.S. AI model developers such as Anthropic and OpenAI, the company is achieving a healthy gross profit margin. The Information reported in July that DeepSeek’s annualized revenue run rate had reached between $400 million and $500 million, and the gross margin for selling access to its models through its application programming interface in the first seven months of this year was 82.9%, higher than both Anthropic and OpenAI during similar periods. That’s because DeepSeek has managed to keep the costs of running its models low, through methods such as improving the efficiency of its AI infrastructure so that the AI systems can perform tasks using fewer chips.

Despite the growth, generating more revenue hasn’t been a priority for DeepSeek, Liang told investors in the meeting on Sunday, which imposed strict terms for participants to prevent leaks. The company generates revenue almost entirely by charging for access to its models through APIs. DeepSeek’s popular chatbot app, which is free to use with no advertising, generates no revenue.

DeepSeek continues to put the majority of its resources into the development of new models. The company is allocating more than 70% of its computing capacity to model training, while reserving less than 30% of its compute for inference, or the running of existing models, Liang told investors during the meeting, according to the people.

Liang also told investors that DeepSeek’s internal tests show its smaller models can run well on graphics chips designed for gaming, according to the people. He said such models can handle most of the everyday tasks users need. If DeepSeek can use gaming chips to run more of those tasks, it could ease the shortage of computing capacity available for inference without diverting too many of its most powerful chips from training. There is precedent for that: after the release of DeepSeek’s R1 model last year, demand and prices for Nvidia gaming chips on China’s black market surged.

Over the past several months, Chinese AI models have rapidly increased their global presence, as the performance of their affordable open-source models narrows the gaps with that of U.S. frontier models from Anthropic and OpenAI. Earlier this month, Z.ai, the Beijing-based developer of the GLM series of models, said its annualized run rate reached $1.8 billion. MiniMax, another Chinese competitor, said its annualized run rate jumped to $800 million in August from $150 million in February.

Cheaper ways to run existing models may help DeepSeek serve more users, but developing new models still requires substantial compute capacity. For DeepSeek, one of the biggest challenges ahead is how to secure enough compute to train its new models amid U.S. export restrictions on Nvidia’s advanced chips. Liang told investors during Sunday’s meeting that a major priority for his company is to use more domestic chips to train its models, adding he expects Huawei Technologies to start delivering training chips to DeepSeek as early as the fourth quarter this year, The Information reported earlier this week.

Another concern for investors is Beijing’s investigation into potential data leaks to Anthropic. The Information reported on Tuesday that China’s internet regulator is investigating DeepSeek and Moonshot AI after Anthropic alleged both companies had been routing sensitive user data to Claude models.