DeepSeek Nears 100 Billion Yuan Funding Round as Tencent and CATL Back Shanghai IPO Push

A financing push doubles in size

DeepSeek is nearing completion of a funding round worth about 100 billion yuan, or $15 billion, with Tencent and battery maker Contemporary Amperex Technology, known as CATL, among its biggest backers. The financing would give the Hangzhou artificial intelligence company substantially more money for models, computing infrastructure and international expansion as it prepares for a possible Shanghai stock market listing in early 2027.

Details emerging on October 22, 2026, show how quickly the transaction has grown. DeepSeek initially sought 50 billion yuan, increased its target to 80 billion yuan and then reached approximately 100 billion yuan. The latest amount is twice its original goal.

People familiar with the transaction described DeepSeek as having raised the money, while others said final completion was expected within days and that financial terms and the investor lineup could still change. DeepSeek had not publicly confirmed the transaction, its final size or its terms by the time of those accounts.

The financing follows a first external round completed in June and a second fundraising effort launched in July. If the latest transaction closes at the reported size, DeepSeek would have attracted roughly 150 billion yuan across the two rounds in about six months.

DeepSeek gained worldwide attention in 2025 by showing that competitive AI models could be developed at relatively low cost, unsettling technology investors and triggering a selloff in technology shares. Its current spending plans reveal the next stage of that challenge: combining efficient software with enough computing capacity to serve more customers and develop more capable systems.

The central business question is now concrete. DeepSeek must turn a much larger pool of capital into reliable computing infrastructure, useful models and revenue that can support the continuing cost of operating them.

Who is investing, and how much is DeepSeek worth?

Tencent, the operator of WeChat, and CATL are returning investors. More than 15 investors are reportedly involved in the latest round, bringing together technology companies, manufacturers and investment funds.

Named participants include appliance maker Midea, food delivery company Meituan, automaker Geely and semiconductor company GigaDevice. Other investors identified in the fundraising include Stony Creek Capital, China Fortune-Tech Capital, CPE, Legend Capital, Boyu Capital and IDG Capital. Some were involved in earlier discussions, while the final membership of the round was still being settled.

Existing shareholders and new investors were expected to provide roughly half the money each. CATL reportedly planned to contribute about 10 billion yuan, following approximately 5 billion yuan in the first round. At those amounts, its latest contribution would be twice its earlier investment and about one tenth of the new financing.

DeepSeek's reported valuation before the new investment is approximately 500 billion yuan. A valuation is the price investors place on the whole business, rather than the amount of cash it holds. On a straightforward calculation, adding 100 billion yuan would produce a value after investment of about 600 billion yuan, with the new capital representing roughly one sixth of that total. The final ownership split will depend on the transaction's terms.

The mix of industrial and technology investors gives DeepSeek a broad financial base. The reported investment agreements, however, contain no public detail about how these companies might use its models or whether they would receive special commercial access.

How the fundraising unfolded

DeepSeek's financing and listing preparations have advanced alongside new models and work on Chinese computing hardware. The sequence helps explain why the latest round became much larger than its initial target.

• April 2026: DeepSeek reportedly began its first external fundraising effort.
• Late June: The first round closed, raising about $7.4 billion, or roughly 50 billion yuan.
• July: DeepSeek began its second round, initially targeting 50 billion yuan.
• September: The company released DeepSeek-V4.1-Flash and hired Yen Wen-tao as chief financial officer.
• September 30: DeepSeek released six software components for Huawei Ascend hardware.
• October 6: Fundraising accounts put commitments at at least 80 billion yuan, with the possible total approaching 100 billion yuan.
• October 17: DeepSeek released V4.1-Flash-1017, focused on completing entire development projects.
• October 22: The latest accounts put the financing at approximately 100 billion yuan, with final completion expected within days.

The negotiations reportedly stretched from July to October as more investors sought to join. By the final stage, the transaction had become both a source of expansion capital and a step toward selling shares to public investors.

A Shanghai listing is taking shape

DeepSeek is reportedly targeting a listing in Shanghai in early 2027 and has selected CITIC Securities to help prepare a potential initial public offering on the STAR Market. An IPO allows a company to sell shares on a stock exchange, opening ownership to public investors and creating another way to raise money.

The STAR Market is the Shanghai Stock Exchange board focused on technology businesses. Preparations reportedly include changes to DeepSeek's corporate structure. An October account said a formal listing guidance agreement had not yet been signed; later accounts described CITIC Securities as having been selected or hired.

Yen Wen-tao, a former venture capital investor, joined as chief financial officer in September, with one account placing his arrival late in the month. A finance chief typically oversees financial reporting, funding and the preparations needed for investors to assess a business before a listing.

Early 2027 is a reported target rather than an announced trading date. DeepSeek has not publicly set out the offering size, final listing valuation or detailed timetable. If its plans proceed, it could reach public markets before some major rivals also considering listings, including OpenAI and Moonshot AI.

Revenue is growing, with a pricing contribution

Founder Liang Wenfeng reportedly told investors in September that DeepSeek's annualised revenue had reached about $1 billion. Some of the growth was attributed to higher model prices.

Annualised revenue takes a recent rate of sales and projects it across a full year. A business generating roughly $83 million a month, for example, would be running at approximately $1 billion a year if that pace continued. DeepSeek's figure describes its recent revenue pace, while actual revenue for a completed financial year would depend on sales throughout that year.

The pricing contribution matters because revenue can rise through higher charges, greater usage or a combination of both. For customers, the relevant calculation is how much useful work a model performs for the money spent. A higher price can still be economical if a system finishes tasks with fewer attempts or uses less computing time.

The reported $15 billion financing is about 15 times the stated $1 billion annual revenue pace. That comparison shows the scale of the resources being committed to expansion. Revenue measures sales, while investment supplies capital that can fund research, equipment and operating expenses over time.

Where the computing money could go

DeepSeek has increased spending on data centres and models whose underlying components are made publicly available. The new financing is reportedly intended to support model research and computing power, both essential to developing new systems and answering requests from customers.

A major proposed project is a data centre in Inner Mongolia containing at least 160,000 Huawei Ascend 950DT accelerators. Accelerators are specialised processors that carry out the large volumes of calculations used in AI. The reported plan would create one of the largest known installations using Huawei AI hardware.

Near term chip purchases were estimated at about 17 billion yuan. Compared with a 100 billion yuan funding round, that would equal roughly 17% of the money raised, before allowing for buildings, electricity, cooling, networking and other equipment. The figure is a procurement estimate tied to the proposed deployment.

Huawei's deputy chairman and rotating chairman, Eric Xu, said in September that the Ascend 950DT system was still being tested, with supply at scale expected at the end of 2026 or early 2027. The next generation 960DT was planned for shipment in the first quarter of 2027. Those schedules make equipment availability a practical constraint on DeepSeek's expansion.

The infrastructure push follows DeepSeek's reported work on an inference chip project aimed at reducing the continuing cost of generating answers. Its own processor design and its proposed Huawei deployment address different parts of the same expense: acquiring and operating enough computing capacity to run a growing AI service.

Making models work on Chinese chips

DeepSeek has also worked with Huawei on programming tools tailored to Ascend processors. On September 30, it released six components, including DeepGEMM-Ascend, DeepEP-Ascend, FlashMLA and an Ascend version of TileLang.

Software compatibility is central to the project. Many AI development tools have been built around Nvidia's CUDA software environment. Moving work to another chip family requires engineers to make calculations, communication between processors and memory management operate reliably on the new hardware.

Training creates a model by processing large amounts of data and adjusting its internal settings. Inference is the work a trained model performs when answering a request. DeepSeek's reported ambition to use Chinese hardware for frontier model training would extend domestic processors into a particularly demanding workload, involving large groups of chips running together for long periods.

China's earlier domestic chip requirements for data centres receiving state funding have also pushed the industry toward local equipment. DeepSeek's planned installation would add a large commercial development effort to that wider shift, making hardware supply and supporting software central to its ability to spend the new capital effectively.

New models aim to use less memory

DeepSeek's September model, V4.1-Flash, was described by the company as designed for greater capability, faster inference and higher throughput. Throughput means the amount of work a system can process in a given time, which affects how many customer requests it can handle.

The model reportedly contains 552 billion parameters, the internal numerical settings that help determine its responses. It activates about 8 billion parameters during input processing and 16 billion during output generation. Using only part of the model at each stage can reduce the computing work required for a request.

DeepSeek also described reductions in the resources needed for its key value cache, or KV cache. This is temporary storage that helps a model retain information from the text it is processing. The reported requirements for high bandwidth memory fell to one quarter of the previous generation, while solid state storage requirements fell to one eighth.

Those figures correspond to reductions of 75% and 87.5%, respectively, for those particular cache resources. For applications that repeatedly consult earlier text, call tools and revise results, lower memory requirements can help more work fit on the same equipment.

The October 17 update, V4.1-Flash-1017, focused on completing projects from beginning to end. DeepSeek described using reinforcement learning with self feedback, a training approach in which a system learns from assessments of its results, and showed examples including a three dimensional model of the Forbidden City and a browser racing game.

Competition moves from cheap answers to completed work

DeepSeek's financing would narrow its resource gap with American competitors, although OpenAI and Anthropic have each reportedly raised more than $100 billion. Each company's cumulative fundraising is therefore more than six times the size of DeepSeek's latest reported round.

Competition within China is also accelerating. MiniMax released M3.1 Flash on October 19, targeting tasks involving programming, computer operation and AI agents. Agents are systems that take several steps toward a goal, such as opening tools, writing code and checking a result.

MiniMax said its new model delivered nearly three times the previous inference speed and reduced usage costs by 70%. Updates from Moonshot AI's Kimi and Zhipu's GLM were also expected in the following two weeks. Developers choosing among these products face a practical comparison: the cost, speed and reliability of finishing a task.

DeepSeek hopes fresh capital will support international expansion, but it faces restrictions in some markets. Government bodies, including some US agencies, have barred its technology from government devices amid scrutiny of security and privacy policies. Such restrictions limit access to those customers even as the company expands its technical capacity.

The next financial milestone is completion of the investment round. Beyond that, the reported early 2027 listing target and Huawei's expected hardware deliveries bring funding, equipment supply and public market preparations into the same period. DeepSeek's challenge is to coordinate all three while continuing to release models that customers find worth paying for.

Key Points

• DeepSeek's latest funding round is reported at about 100 billion yuan, or $15 billion, twice its original target.
• Tencent and CATL are among the biggest backers, with more than 15 investors reportedly participating.
• Final completion was expected within days of October 22, with terms and the investor lineup still subject to change.
• The reported valuation before new investment is approximately 500 billion yuan.
• DeepSeek is preparing a potential Shanghai STAR Market IPO, reportedly targeting early 2027.
• Expansion plans include model research, Chinese computing hardware and a proposed Inner Mongolia data centre.
• Liang Wenfeng reportedly put annualised revenue at $1 billion in September, with part of the growth coming from higher prices.