For years the knock on DeepSeek was always the same. The people who refused to take it seriously would nod wisely and repeat a version of the same line: cheap AI does not pay the bills. You can undercut every American lab on API pricing, shatter the cost-per-token charts, and set the open-weight world on fire, the argument went, but sooner or later gravity arrives, and gravity keeps a spreadsheet. That story just took a direct hit.
New financial disclosures about DeepSeek's run through July 2026 paint a very different picture, and honestly, I think most of us in the AI commentary business need to sit down and revise our priors. The Chinese lab generated roughly 475 million yuan, about 70.7 million dollars, in revenue during the first seven months of the year. That is close to a tenfold jump over its entire full-year 2025 figure. Tenfold. In seven months.
The narrative versus the numbers
Here is the uncomfortable part for the skeptics. DeepSeek's API gross margin is reportedly sitting near 83 percent. That is not a charity. That is not a company burning cash to buy attention. That is a real, defensible, software-margin business hiding inside a company the west convinced itself could never monetize discount AI.
And yet.
There is always an "and yet" with this company, and it is the part worth chewing on. Despite that gross margin, DeepSeek still posted a net loss of more than 700 million yuan across the same window. Let me put that plainly: a laboratory that is technically brilliant at squeezing margin out of inference is still, at the bottom line, losing around a hundred million dollars. Gross margin and net profitability are not the same thing, and the gap between them is the entire story.
Where does the money go? It goes into the furnace. Compute for the next frontier model, the data-center buildouts that have been all over the headlines, and a research lab that never stops training. DeepSeek is choosing to reinvest essentially everything it earns, and the market is betting that is the correct strategy.
Why the fundraise finally makes sense
That reinvestment bet is what is driving the real financial drama. DeepSeek is reportedly closing in on a fresh pre-IPO funding round at a valuation around 500 billion yuan, roughly 74 billion dollars, and targeting a listing in 2027. After pausing its second round earlier in the summer, it reopened the talks with this revenue report as ammunition.
Look at the sequence and it becomes clear why an investor would take the meeting. The tenfold revenue surge, the margin profile, the nine-figure loss, and the 74 billion dollar valuation are one coherent bet, not four separate stories. Consider what each number is actually telling you:
- The revenue growth proves the product has genuine, paying demand, not just open-source enthusiasm.
- The 83 percent gross margin proves the underlying unit economics are real.
- The 700 million yuan loss proves DeepSeek is spending aggressively to widen its moat.
- The valuation prices all of that in before the IPO window opens in 2027.
I want to be the first to admit the frame I held before this was wrong. I treated DeepSeek's pricing as a permanent discount, a subsidy that could never be raised into profitability. The company has since bumped prices, and the revenue climbed anyway. Margin is high. Demand did not flee. That is the opposite of a company in trouble.
None of this means the risks vanished. Losing 700 million yuan against a 70 million revenue base is still a real burn, and the whole structure only works if the training treadmill keeps producing models people will pay for. One bad roadmap decision, one plateau, and the economics flip from growth at scale to cash incineration.
But the headline takeaway today is not the loss. It is that DeepSeek has finally put real numbers on the table proving that discount AI can become a real business. The "cheap AI cannot pay the bills" narrative did not survive contact with revenue reporting. It was always fair criticism on the surface, and as DeepSeek's financials show, it was also always a little lazy.
I am not ready to crown DeepSeek the winner of the AI economics wars yet. What I am ready to say is that the game changed in the last month, and anyone still telling you that open-source pricing is a doomed business model is arguing from a position that no longer matches the data.
Comments