Key Takeaways:
A Chinese developer that spent last year selling inference at a fraction of Western prices has now shown it can charge more. DeepSeek's annualized revenue run rate has climbed past $1 billion, The Information reported Sept. 24, more than doubling from below $500 million a few months earlier, as the company closes a 50 billion yuan ($7.45 billion) funding round at a 500 billion yuan ($74.6 billion) target valuation.
The revenue jump was driven partly by raising model pricing by 2.3 to 4.5 times, the report said, a detail that separates pricing power from a subsidized land grab. Chief Executive Liang Wenfeng shared the figure at a recent investor meeting, where he also said the company allocates more than 70% of its computing capacity to training and less than 30% to inference, according to the report.
Neither DeepSeek nor The Information's sources have published audited financials. Reuters, which reported the same figures, said it could not immediately verify the report and that DeepSeek could not be reached for comment. Run-rate revenue measures what a company would earn over a year if its current pace held, and figures of this kind for private AI labs have historically been revised or contested.
The numbers land as DeepSeek prepares a potential listing on Shanghai's STAR Market. The company hired CITIC Securities to prepare the offering and named a first chief financial officer, a GL Ventures partner, earlier in September, Reuters reported. Timing, valuation and fundraising size remain undecided.
The composition of the growth carries more signal than the headline. A low-cost entrant that doubles revenue while lifting prices 2.3 to 4.5 times is demonstrating that demand for cheap inference survives a price increase — a test that Western model providers have not had to run at the same discount level.
That cuts two ways for the peer set. It confirms a large, price-sensitive market for Chinese inference, which supports demand assumptions behind domestic compute buildouts. It also pressures the pricing assumptions embedded in Western model providers' revenue models, where per-token pricing has fallen steadily as competition intensified.
DeepSeek's own cost base is the other half of the equation. Liang told investors the company is spending roughly 30 billion yuan ($4.5 billion) to expand computing infrastructure, and that training a model comparable in scale to the largest systems OpenAI is developing would require about 50,000 Nvidia GB300 chips or 200,000 Huawei Ascend 950 chips. The company is currently training a 2-trillion-parameter model and has an 8-trillion-parameter model on its roadmap.
For scale, Alibaba's Qwen 3.8 Max carries 2.4 trillion parameters and Moonshot AI's Kimi K3 has 2.8 trillion total parameters under a mixture-of-experts architecture that activates 104 billion per token. An 8-trillion-parameter DeepSeek model would be roughly three times Kimi K3's total scale.
Liang also told investors DeepSeek plans to adopt domestic chips from Huawei and other Chinese suppliers at scale for model training, calling it one of the company's "biggest bets" and saying it "must succeed." Huawei is expected to deliver training chips as early as the fourth quarter, aligning with the Ascend 950DT launch on Huawei's three-year Ascend roadmap. Huawei said in April that its chips had already handled part of the training work for DeepSeek's lightweight V4-Flash model.
Supply is the constraint. Huawei has been limited by shortages of advanced memory and other components, so DeepSeek continues to train new models on Nvidia hardware in parallel. Training places far higher demands on chip stability, interconnect bandwidth and software toolchains than inference, and Huawei's ability to meet those requirements at 200,000-chip scale is unproven.
The read-through for listed AI equities runs through the hyperscalers and model developers rather than DeepSeek itself, which remains private. A verified $1 billion run rate at a $74.6 billion target valuation implies roughly 75 times annualized revenue — a multiple that only holds if the pricing increase sticks and the Huawei transition does not delay model releases. Any confirmation or denial from DeepSeek, and corroboration of the pricing detail, are the next checkpoints.
This article is for informational purposes only and does not constitute investment advice.
