According to an October 6, 2026 CNBC report, DeepSeek is considering expanding its current funding round from an initial target of RMB 50 billion (about US$7.4 billion) to as much as RMB 100 billion (about US$14.9 billion). The change is being driven by intensive subscriptions from state-owned funds, investment arms of listed companies, and venture capital institutions, rather than DeepSeek actively raising its target—according to people familiar with the matter, the fundraising process is still underway, and the final amount may be adjusted again before negotiations conclude.
This is DeepSeek's second round since it first opened to external financing in 2026. The first round closed in June 2026, raising about RMB 50 billion, with an implied valuation of about RMB 350 billion based on investor filing documents. In this round, DeepSeek's target valuation is RMB 50 billion, equivalent to about US$75 billion—in just half a year, the valuation anchor has moved up by about 40 percent.
State-Owned Capital Leads as Individual Money Is Shut Out
The most notable structural feature of this round is the investor screening logic. According to CNBC, citing people familiar with the matter, DeepSeek is subjecting potential investors to strict review, actively rejecting private funds dominated by individual investor money, and essentially limiting participants to government-backed funds and corporate entities.
The institutions confirmed to have committed funds include Tencent Holdings, CATL, Geely Automobile, and asset managers Monolith Management and Loyal Valley Capital. According to a previous Bloomberg report, Tencent contributed about RMB 10 billion in the first round and CATL about RMB 5 billion; both are also the largest contributors in this round.
By shutting individual money out, DeepSeek can, on the one hand, avoid excessive equity dispersion before a listing and reduce compliance friction for a future IPO; on the other hand, bringing in state-owned capital and industrial capital can create a more solid ecosystem endorsement during a sensitive period for China's AI policy. The participation of CATL and Geely Automobile—both core players in the hardware and energy ends of the industrial chain—suggests that such cross-shareholding often signals deeper strategic collaboration rather than purely financial investment.
Where the Money Will Go: The Thousand-Card Cluster in Inner Mongolia
According to a report by the U.S. media outlet American Bazaar, DeepSeek is building a large data center in Inner Mongolia and plans to deploy at least 160,000 Huawei Ascend high-end AI accelerator cards. Based on the current market price of the Ascend 950DT chip at about RMB 111,000 per card, the purchase of this batch of chips alone would amount to nearly RMB 18 billion (about US$2.5 billion), equivalent to one-third of the total first-round financing.
These figures explain why the second round must be larger: the marginal investment by large-model companies in inference infrastructure is growing at a nonlinear rate. Simply maintaining the external service capacity of existing models already requires continuous additions of compute, let alone simultaneously advancing pretraining for next-generation models.
Betting on the Huawei Ecosystem: A Technology Strategy and a Geopolitical Hedge
DeepSeek's ties with Huawei go far beyond what outsiders generally realize. On September 30, 2026, DeepSeek open-sourced a complete programming toolchain optimized for Huawei Ascend chips, including TileLang—a high-level programming language positioned as a domestic alternative to Nvidia's CUDA—as well as low-level computing libraries such as DeepGEMM, FlashMLA, and TileKernel. The release of this toolchain means DeepSeek is actively lowering the barrier to entry for Ascend chips across China's entire AI developer ecosystem.
Under U.S. export controls, Nvidia's H100/H800 series has become a scarce resource for Chinese AI companies. Through deep cooperation with Huawei, DeepSeek is attempting to achieve a kind of chip agnosticism at the software layer: enabling more Chinese AI developers to reach comparable inference efficiency on Ascend hardware without relying on Nvidia's CUDA ecosystem. This is consistent with the high-efficiency training narrative conveyed by its R1 model in early 2025.
From an investor's perspective, the combination of the Inner Mongolia data center and Huawei Ascend 950DT is also telling the capital market a story that does not depend on Nvidia. Against the backdrop of continuously escalating U.S.-China technology friction, the strategic valuation premium logic of a Chinese AI company not constrained by Nvidia's supply chain is clear.
Model Side: V4.1 Flash Is the Direct Trigger for the Financing
The unexpectedly strong subscriptions for this round are closely related to DeepSeek's latest model, V4.1 Flash, released in September 2026. According to reports, this model uses a mixture-of-experts (MoE) architecture with 552 billion parameters and natively supports multimodal visual understanding, with improvements in both inference throughput and capability ceiling. After the model's release, subscription intentions from investment institutions clearly accelerated.
The CNBC report mentioned that the first round was delayed because some potential investors disagreed on price, but negotiations are now nearing their end. This detail indicates that DeepSeek did not concede on pricing—the RMB 50 billion target valuation has been held firm, and the expanded quota came because investors actively came knocking.
Scale Benchmarks and Market Coordinates
In reporting this financing, international media broadly invoked a comparison framework with Anthropic. In May 2026, Anthropic completed a US$65 billion Series H funding round, with a valuation approaching US$1 trillion, and its cumulative historical funding total already far exceeds the US$14.9 billion cap of DeepSeek's current round.
More useful comparative data comes from within China's AI sector: the scale of this DeepSeek round is equivalent to twice its previous round (closed in June 2026, about US$7.4 billion), and it comes only about four months after the previous round closed. This capital density is an extremely rare pace even in the history of Chinese technology.
The purchase of Ascend chips for the Inner Mongolia data center alone is already equivalent to about one-third of the total first-round financing. This means that if the entire second round is successfully completed, about two-thirds of the funds will need to be used for new compute beyond existing infrastructure, model research and development, and possible overseas market expansion. DeepSeek founder Liang Wenfeng once emphasized at an investor meeting that the company will continue to develop open AI models and take artificial general intelligence as its long-term goal, rather than focusing on short-term monetization.
The Pre-IPO Outpost: The STAR Market's Potential Demonstration Effect
According to American Bazaar, DeepSeek has hired CITIC Securities to prepare for a listing on China's STAR Market (the Science and Technology Innovation Board), with a target time of 2027. Structurally, the current funding round is the last large-scale private placement before the IPO: by bringing in state-owned funds and industrial capital, it completes the policy-compliance groundwork for the shareholder register while anchoring the valuation at RMB 50 billion to form a reference frame for STAR Market pricing.
The strategic significance of the STAR Market for AI companies is not merely a financing channel. Once DeepSeek successfully lists, it will become a landmark case of a Chinese domestic large-model company on the A-share market and could drive a reshaping of the valuation system for the entire domestic AI industry.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接