DeepSeek Closes In on $13.9B Mega-Round, Doubling Down on Huawei Chips in Race to IPO
China's most closely watched AI startup is set to lock in its second massive fundraise of 2026, with Tencent and CATL leading a round that has ballooned well past its original target — and a domestic-chip infrastructure bet that could redefine the country's AI compute playbook.
DeepSeek, the Hangzhou-based large language model developer that rattled global AI markets earlier this year, is finalizing a funding round of at least RMB 80 billion (US$11.1 billion) with a potential ceiling of RMB 100 billion (US$13.9 billion), according to people familiar with the matter — nearly double the company's initial RMB 50 billion target and representing the largest single fundraise in Chinese AI history if completed at the upper end.
Term sheets have been signed and the process is approaching its final stage, the people said, cautioning that specific terms remain subject to change. The round values DeepSeek at an implied target of approximately RMB 500 billion (US$69.4 billion), a step-up from the roughly US$55 billion post-money valuation established in its June 2026 Series A. The capital will be deployed primarily toward compute infrastructure buildout, research and development, and preparations for an initial public offering targeted for early 2027.
Tencent and CATL Emerge as Anchor Investors, Signaling Strategic Alignment
Tencent Holdings and the CATL ecosystem are poised to be the largest committed contributors in the current round, the people said — a continuation of positions both conglomerates established in June's Series A, when Tencent injected approximately RMB 10 billion (US$1.39 billion) and CATL's network contributed roughly RMB 5 billion (US$694 million).
The broader investor syndicate reflects a deliberate architecture: existing shareholders deepening exposure alongside incoming state-linked and industrial capital. Returning investors set to increase stakes include IDG Capital, Lishang Capital, Shixiang Capital, Zhengxingu, and the National AI Industry Investment Fund. Newly entering the negotiations are Shixi Capital, SMIC Juyuan, Boyu Capital, CPE Yuanfeng, and a new-capital state fund.
The structure is not incidental. The concentration of state-affiliated and deep-pocketed industrial investors around DeepSeek signals that Beijing's AI national champions strategy is consolidating capital into fewer, larger bets — a pattern that mirrors the semiconductor investment playbook of the past five years.
Oversubscription Driven by V4.1-Flash Launch, Repricing Competitive Benchmarks
The trigger for the round's dramatic oversubscription traces directly to product momentum. Following earlier models that repositioned market expectations around DeepSeek's price-performance ratio relative to Anthropic and OpenAI offerings, the company in September 2026 launched V4.1-Flash, emphasizing enhanced reasoning throughput and inference speed. Investor appetite surged beyond initial projections, forcing repeated upward revisions to the fundraise target.
Founder Liang Wenfeng, who personally contributed approximately RMB 20 billion (US$2.78 billion) in the June round — the single largest individual commitment — had famously stated in 2024 that DeepSeek's constraint was never capital but access to high-end chips blocked by U.S. export controls. The company's current fundraising logic represents a calculated inversion of that position: using capital to buy time, talent, and — critically — domestic compute ecosystem development.
This is DeepSeek's second large-scale financing event within roughly six months. The June 2026 Series A closed at RMB 51 billion (US$7.08 billion), setting what was then a record for a single Chinese AI round, with a post-money valuation near RMB 400 billion (US$55.6 billion). If the current round closes at its upper bound, DeepSeek's cumulative fundraising across both events would exceed RMB 150 billion (US$20.8 billion).
Inner Mongolia Data Center and 160,000 Huawei Ascend Chips Anchor Compute Strategy
The financing is inseparable from a capital-intensive infrastructure commitment. DeepSeek is advancing construction of a large-scale data center in Inner Mongolia, targeting deployment of at least 160,000 Huawei Ascend 910C 950DT supernode accelerators — its first foray into frontier model pre-training on domestically produced compute, rather than restricting Ascend chips to inference workloads. At prevailing market unit prices, the near-term chip procurement alone is estimated at approximately RMB 17 billion (US$2.36 billion).
On September 30, 2026, DeepSeek simultaneously open-sourced six infrastructure components: TileLang (Ascend edition), DeepGEMM-Ascend, DeepEP-Ascend, TileKernels, FlashMLA, and DeepSelect. The stack has been integrated into Huawei's CANN (Compute Architecture for Neural Networks) community, covering large-EP low-latency inference, large-scale training, KV Cache pooling, and Agentic Reinforcement Learning pipelines.
Industry engineers note the engineering complexity involved. Scaling pre-training to tens of thousands of chips requires solving operator correctness, mixed-precision alignment, distributed communication stability, and checkpoint recovery — challenges where China's domestic software stack has historically lagged CUDA by a material margin. The incremental maturation of compilation toolchains and the open-sourced component library represent DeepSeek's attempt to close that gap systematically.
If the Ascend-based pre-training pipeline proves production-viable, the implications extend well beyond one company: it would offer the Chinese AI industry a replicable template for frontier model development that is structurally independent of Nvidia hardware and the CUDA ecosystem.
IPO Timeline Crystallizes, CFO Hire and CITIC Securities Mandate Signal Execution Phase
Capital markets preparation is moving from planning to execution. Multiple media reports indicate DeepSeek is in discussions to engage CITIC Securities to advise on a STAR Market listing on the Shanghai Stock Exchange, with the intention to initiate the formal process before year-end 2026, though an advisory agreement has not yet been formally signed.
The hire of Yan Wentao, previously a partner at Hillhouse Venture. as Chief Financial Officer in late September 2026 is the most concrete organizational signal that capital markets execution has moved to the front burner. CFO appointments at this stage in Chinese tech companies' pre-IPO timelines have historically preceded regulatory filing submissions by six to twelve months.
Sector-Wide Fundraising Compression Raises Valuation Floor Across Chinese AI
DeepSeek's trajectory sits within a broader pattern of accelerating capital concentration at the top of the Chinese large model stack. Moonshot AI raised approximately RMB 25 billion (US$3.47 billion) in July 2026, far exceeding its initial RMB 7–14 billion target, and is targeting a Hong Kong IPO in Q1 2027 at a private market valuation of roughly US$50 billion. Zhipu AI raised a combined approximately HK$70 billion (US$9.0 billion) across two tranches and a convertible bond offering in July and September 2026.
Meanwhile, Manus, developed by Beijing Butterfly Effect Technology, recently closed a funding round exceeding US$500 million led by Boyu Capital and IDG Capital, with continued participation from Tencent, Sequoia China, and ZhenFund — reportedly doubling its valuation toward US$4 billion.
The aggregate picture is one of concentrated capital chasing a narrow set of companies deemed capable of reaching commercial scale before the IPO window opens. For DeepSeek, the compounding effect of fundraising momentum, domestic compute investment, and product velocity creates a reinforcing narrative — but also concentrates execution risk. The validation of its Ascend-based compute stack is now the single most consequential near-term variable for both its valuation trajectory and its relevance as an industrial policy benchmark.
Related Coverage:
DeepSeek Desktop Agent Disrupts Enterprise SaaS Ecosystem
DeepSeek, Kimi Wage Asymmetric War for China's Scarcest AI Minds