Two people familiar with DeepSeek's financials have disclosed a set of previously unpublished numbers: from January to July 2026, the company generated roughly ¥475 million in revenue (about $70.7 million), while posting a net loss of about ¥715 million over the same period. For comparison, the company's net loss for all of 2025 was ¥935 million, and its full-year revenue that year was only about a tenth of what it made in the first seven months of this year.
This marks the first time anything close to a complete set of operating figures has leaked out of the company. Until now, outside estimates have relied almost entirely on guesswork — multiplying per-token prices by usage volume, or working backward from the training costs of its open-weight models — with margins of error running several times over.
Two gross margins tell two different stories
The gross-margin line is what clarifies the structure. Over the first seven months, DeepSeek's overall gross margin came to 44.6%, while its API model-inference service alone posted 82.9%.
82.9% is a software-company number. It shows that selling inference by the token is fundamentally viable — after accounting for hardware depreciation, electricity, and bandwidth, the company keeps more than 80 cents of every yuan it takes in. The 44.6% figure emerges once the free traffic from the app and web interface gets folded in. That 38-point gap is essentially the compute DeepSeek gives away to the public every day.
That also pins down where the losses come from: not API pricing, but training spend and the subsidy running through the free tier. Rough math suggests the 82.9% margin on the API side alone isn't nearly large enough to cover the ¥715 million loss.
The price hike just moves the subsidy around
This month DeepSeek made a sizable price adjustment to the API for its flagship model, V4-Pro, and introduced peak and off-peak tiers. Output pricing now rises to $3.96 per million tokens at peak hours and $1.98 per million tokens off-peak; the previous flat rate was $0.87. That's a roughly 4.5x increase at peak and more than double at off-peak, with the peak rate running exactly twice the off-peak one.
Read against the gross-margin structure above, the logic isn't hard to parse: the API side was already profitable, so the point of raising prices again isn't to make that business more profitable — it looks more like making it carry costs for the free tier, while using the price gap to push load toward the night. Developers' scheduled jobs, batch cleanup, and offline evaluations — anything that isn't urgent — get pushed by pricing into off-peak hours, freeing up daytime capacity for interactive requests. It's a practice utilities and cloud computing have used for decades; this appears to be the first time a mainstream model provider has written it into an official price list for tokens.
The cost falls on developers, who now have to rework their billing. A batch pipeline that used to run during the day, left unchanged, would see its cost jump from $0.87 to $3.96 — 4.5 times over; moved to nighttime, it becomes $1.98. Same code, different hour, and the cost can differ by a factor of two.
The loss is widening, not narrowing
Revenue rose tenfold, but the loss didn't shrink to match. The full-year loss for 2025 was ¥935 million, averaging about ¥78 million a month; over the first seven months of this year the loss was ¥715 million, averaging about ¥102 million a month. Rough math puts the monthly loss up by roughly 30%. That fits the profile of a company still stacking up training investment — its revenue curve is steep, but its cost curve is steeper.
On a global scale, ¥715 million is roughly $100 million. Public figures put OpenAI's annual loss at $38.5 billion — two orders of magnitude apart. But revenue is also two orders of magnitude apart, so the comparison isn't really about restraint; the two companies simply aren't operating at the same cost scale for the same business.
On the funding side, DeepSeek is in talks with new and existing investors for a fresh round, targeting a valuation of ¥500 billion, having just closed a round in June. Talk of a 2027 IPO hasn't gone away either. With this set of gross-margin numbers, investors can at least see one thing clearly: the inference business itself is healthy, and it's training and user growth that need the cash.
Sources: two people familiar with the company's financials cited by Global Market Watch, Sina Finance, CocoLoop, cnBeta, DeepSeek's open platform pricing page; revenue, losses, the two gross-margin figures, and API peak/off-peak pricing are cross-checked against public reporting, with monthly averages and multiples flagged as rough calculations.