DeepSeek doubles its annual revenue to $1 billion on the eve of its IPO

Citing sources familiar with the matter, The Information reports that DeepSeek's annualized recurring revenue (ARR) has exceeded 1 billion U.S. dollars, more than doubling from the figure of less than 500 million U.S. dollars a few months ago. Liang Wenfeng, CEO of DeepSeek, disclosed this data at a recent investor meeting, and explicitly stated that the API price hike has not caused customer churn, with user demand remaining strong.

Evolving from a "technical raider" to a "commercial giant", DeepSeek is completing a critical identity transition. Behind this transition lies a systematic layout covering finance, organization, technology and computing power.

ARR Doubling: Price Hike Validates Business Model

The core driving force behind the revenue doubling comes from the API price adjustment in mid-August. DeepSeek raised the model invocation price to 2.3 to 4.5 times the original level, with particularly significant increases for cache-hit scenarios — taking V4 Pro as an example, the cache-hit input price during peak hours has increased by about 1100% compared to before, and the increases for other items vary by time period and Token type.

The fact that the price increase did not lead to customer churn is itself a form of market validation. Liang Wenfeng stated bluntly at the investor meeting that user demand remains strong after the price adjustment.

What deserves more attention is the revenue quality. According to The Information, in the first seven months of 2026, DeepSeek's revenue reached approximately 475 million U.S. dollars, which is about ten times its total revenue in 2025. During the same period, the gross profit margin of its API interface business was as high as 82.9%, and the overall gross profit margin was 44.6%. This means that DeepSeek's revenue model is not "burning money for scale", but converting technical advantages into profit margins through gradual price increases after establishing market barriers with a low-price strategy.

Up to now, almost all of DeepSeek's revenue comes from charges for model API invocations. Its widely popular conversational bot app remains free and ad-free, generating no direct revenue yet. In other words, the 1 billion U.S. dollars ARR is achieved without activating C-end monetization, leaving considerable room for imagination for future revenue growth.

Post-90s CFO Onboarded, Capitalization Pace Fully Accelerated

In the same week when the revenue data was released, DeepSeek welcomed its first CFO since its establishment. On September 21, Yan Wentao, former partner of GL Ventures, officially joined the company as Chief Financial Officer, ending the three-year vacancy of the CFO position since the company was founded.

Born in 1991, Yan Wentao graduated from Fudan University. He worked at Tencent Investment and H Capital successively from 2013 to 2020, joined GL Ventures in 2020 and was promoted to partner. He has participated in investments in projects including ByteDance, Zhipu, MiniMax, J&T Express and more. The choice of a young financial leader with profound AI investment background sends a clear signal: what DeepSeek needs is not only financial compliance capability, but also strategic talents who can understand technical logic and connect to the capital market.

The timing of Yan Wentao's onboarding is quite meaningful. On September 9, market sources said that DeepSeek has commissioned CITIC Securities to prepare for its STAR Market IPO, and CITIC Securities has entered the due diligence stage, but the two sides have not yet signed a formal listing counseling agreement. The STAR Market of Shanghai Stock Exchange is a board specially set up for scientific and technological innovation enterprises, which is regarded as the first choice for DeepSeek to go public.

The financing pace is also accelerating. In June this year, DeepSeek completed its first round of external financing of about 510 billion RMB, with a post-investment valuation of nearly 4 trillion RMB; currently, DeepSeek is advancing the second round of financing, targeting to raise 500 billion RMB (about 75 billion U.S. dollars) with a target valuation of 5 trillion RMB, which is planned to be completed by the end of October, and the valuation has increased by more than 100 billion RMB in three months.

V4.1 Flash: Price Cut Without Compromising Intelligence

While making rapid progress in revenue and capital, DeepSeek has not slowed down the iteration pace on the product side. On September 10, DeepSeek officially released the V4.1 Flash model, the smallest model in its brand new model architecture series, with native multimodal visual understanding capability.

In terms of technical architecture, V4.1 Flash is a 552B parameter MoE model, adopting a brand new Causal-Encoder-Decoder structure with asymmetric input and output — the input activation is only 8B, and the output activation is 16B, with significantly lower cost than models of the same size. In benchmark tests, multiple indicators of V4.1 Flash outperform flagship models including DeepSeek V4 Pro: it scores 90.6 on Terminal-Bench 2.1, 74.2 on DeepSWE v1.1, and 3471 on Codeforces Rating.

What is more strategically significant is the pricing strategy: after the release of V4.1 Flash, the corresponding API price was cut, with a maximum drop of 60%. On the one hand, the flagship model sees price hikes, on the other hand, the latest lightweight model sees price cuts — this two-tier pricing of "price increase for high-end products, price cut for entry-level products" is helping DeepSeek build multi-level customer coverage across different price bands.

70% of Computing Power Allocated to Training, Betting on Domestic Chips

In terms of resource allocation, Liang Wenfeng revealed to investors that DeepSeek continues to invest most of its resources in new model R&D, with more than 70% of computing power allocated to model training, and less than 30% left for the inference operation of existing models. Against the backdrop that ARR has exceeded 1 billion U.S. dollars, still devoting the vast majority of computing power to training rather than commercial inference reflects DeepSeek's strategic determination to "exchange long-term model capabilities for future market".

The chip strategy is also accelerating adjustment. At present, DeepSeek uses both NVIDIA and Huawei chips, and Huawei chips are mainly used for inference. However, Liang Wenfeng clearly stated at the closed-door investor meeting on September 20 that training models with domestic chips, especially Huawei chips, is "one of the biggest bets" of the company at present, and this work "must succeed". He believes that Huawei chips are expected to catch up with NVIDIA's level in only a few years.

Liang Wenfeng also revealed that DeepSeek is currently training a larger cutting-edge model with a parameter scale exceeding that of the V4 flagship model, and there are plans for subsequent models with even larger parameter scales. Sources say that DeepSeek plans to deploy at least 160,000 new Huawei Ascend 950DT acceleration chips in its under-construction data center in Ulanqab, Inner Mongolia for model inference. If this plan is implemented smoothly, it will become one of the largest known domestic AI chip clusters so far.

From Model Lab to System-Level Company

On September 8, DeepSeek launched a remarkable recruitment campaign: it plans to recruit about 150 engineers, all positions are concentrated in two directions: server-side development engineer and Agent elastic computing R&D engineer, mainly targeting senior back-end engineers with 2 to 10 years of experience, with no AI research positions set at all.

This change in recruitment direction clearly points to DeepSeek's strategic extension: expanding from basic large models to the Agent application layer. Cooperating with DeepSeek Harness officially open-sourced in August this year — an agent operation framework with the core concept of "Agent = Model + Harness", DeepSeek is building a complete technology stack from the model to the application execution layer.

As Cui Tianyi, head of the DeepSeek Harness team, said, "When anything in the computer field scales up, there will be a huge increase in complexity." As AI model capabilities gradually mature, the real challenge has shifted from "whether it can be developed" to "how to make more and more models, Agents and user requests run stably, efficiently and at low cost". Evolving from a pure model laboratory to a system-level company integrating "model + computing power + chip" is exactly the next page that DeepSeek is writing.

Competition and Valuation: The Logic of 163x Multiple

Challenges are equally real. OpenAI's GPT-6 Luna has a lower price per single task than DeepSeek, with only 0.10 U.S. dollars per million input tokens and 0.50 U.S. dollars per million output tokens. However, in the DeepSWE v1.1 evaluation, DeepSeek V4.1 Flash scores 74.2, while GPT-6 Luna scores about 66.6. Domestic competitors such as Z.ai's GLM-5.3 and Moonshot's Kimi K3 also perform well in benchmark tests such as coding.

In terms of ARR volume, according to estimates from Rhodium Group's report, DeepSeek's ARR is about 500 million U.S. dollars, lower than MiniMax's 800 million U.S. dollars and Moonshot's 1 billion U.S. dollars, and also lower than Z.ai's 1.8 billion U.S. dollars. Even calculated based on the latest 1 billion U.S. dollars ARR disclosed by The Information, there is still a huge gap between DeepSeek's revenue and OpenAI's revenue of about 40 billion U.S. dollars.

But the market has adopted a completely different pricing logic. According to the Rhodium Group report, DeepSeek's current valuation/ARR multiple is about 163x, while OpenAI's is about 34x, and Anthropic's is about 21x. Behind this premium is investors' bet on DeepSeek's full-stack capability of "model + computing power + chip", as well as the pricing for the scarcity of Chinese AI assets.

From shocking the world at the beginning of 2025 with low-cost and high-performance models, to now achieving 1 billion U.S. dollars ARR, onboarding the post-90s CFO, concluding the 7.5 billion U.S. dollars financing soon, and entering the due diligence stage for STAR Market IPO, DeepSeek's narrative is shifting from "technical surprise" to "system capability". When an AI company simultaneously has top-level model R&D capabilities, efficient computing power utilization strategies, clear commercialization paths and mature capital market docking capabilities, it is no longer just a model laboratory, but is becoming a real giant in the AI industry.

This article is from the WeChat official account "Tech Business", author: Tech Business, published with authorization from 36Kr.