- DeepSeek had revenue of around 475 million yuan in the first seven months of 2026, which is ten times the entire year of 2025.
- In that same period, it spent 11 billion yuan on compute power and AI infrastructure, i.e. roughly 23× more than it earned.
- Despite its rocket-fast growth, the company is still in the red by more than 700 million yuan.
- Overall gross margin is 44.6%, and for API services it’s even 82.9%. So the issue isn’t the sales itself, but the hardware.
- DeepSeek is negotiating a second round of funding of 50 billion yuan at a valuation of around half a trillion, and it’s also considering going public.
For most people, DeepSeek is synonymous with cheap AI: a model that allegedly was created for a fraction of what U.S. companies spend on development. But the first detailed figures about the company’s finances also show the other side of the coin. Running a large language model is insanely expensive in China too, and the gap between what DeepSeek earns and what it burns is massive.
Revenue up tenfold, expenses even higher
The numbers were provided by the Chinese business magazine Caijing, and were also picked up by the US The Information. According to them, in the first seven months of 2026 DeepSeek earned approximately 475 million yuan. It sounds decent until you put it next to the costs: on compute power and AI infrastructure, the company spent about 11 billion yuan in the same period—around $1.5 billion in conversion.
The difference is 23-times, and it’s not just a one-off fluctuation. Over the full year 2025, DeepSeek’s revenue was only about 47,5 million yuan, and it spent 1,2 billion on infrastructure. So revenue jumped year over year by ten times, but costs rose almost as fast.
The margin, however, is excellent. The problem is hardware
What’s interesting is that DeepSeek’s business itself works very well. The company’s overall gross margin reached 44.6% over the period in question, and for model calls via API it climbed all the way to 82.9%. Those are numbers many software companies would envy. So every sold token brings real profit.
Money doesn’t disappear in running the service, but in what’s underneath it. Expenses include leasing servers with AI chips, buying the chips themselves, and other computing equipment. It’s an investment in future models, which only shows up in revenue a year or two later. That’s why the company is still operating at a loss exceeding 700 million yuan despite healthy margins.
Half a trillion yuan and the whisper about going public
To keep going at this pace, the company is looking for more money. According to Caijing, it is in talks with existing and new investors about a second funding round worth 50 billion yuan, which would value it at roughly 500 billion yuan. At the same time, there’s talk of preparing an initial public offering.
It’s the logical continuation of the story. DeepSeek made a name for itself by training a competitive model more cheaply than others. Keeping it running and training further generations, however, can’t be done cheaply anywhere in the world. Even that “cheap” Chinese path costs billions.
Why this matters for you, too
These figures explain why DeepSeek models can be used so cheaply: the company is still subsidizing growth from investors’ money, not from revenue. You’ve seen the same behavior before with delivery apps or taxi services that subsidized every ride. Sooner or later, the prices for using AI will have to move closer to the real costs.
For everyday users, this is good news so far: while investors are pouring billions into Chinese AI companies, access to strong models is cheaper than it should be. A great reason to try them right now.
Follow the China Planet website, subscribe to our monthly newsletter, and be sure to check out our Facebook page, the China Planet group, and join our Telegram, where we add the best offers and coupons from China every single day.
Don’t miss the best coupons 📩
Subscribe and we’ll email you new coupons before they expire. We only send it once a month — no spam.
