The capital maneuvers at DeepSeek have accelerated markedly over the past six months. In April, the company initiated its first external funding round; by June, that round was completed, raising 50 billion yuan and achieving a post-investment valuation exceeding 350 billion yuan. July saw a temporary halt to the second round, which resumed in August with plans to raise approximately another 50 billion yuan. By September, DeepSeek moved closer to public markets, with reports indicating the company has selected CSC Financial to advance its STAR Market IPO, targeting a valuation around 500 billion yuan and aiming to launch the listing process within the year.
Investors willing to fund DeepSeek have never been scarce. From internet giants, industrial capital, and state-backed funds to venture capital and family offices, capital continues to compete for limited investment shares. Why is DeepSeek choosing this moment to go public, and how will an IPO reshape its operational logic?
The Cost of Intelligence
AI is an expensive business, and DeepSeek's emergence hasn't changed that fundamental reality. What sets DeepSeek apart is its historically restrained approach to participating in the world's priciest technological race. For a long period, the team was far smaller than leading US AI labs, funding came primarily from High-Flyer, and the founder maintained tight control, leveraging algorithmic and engineering efficiency to minimize training and inference costs. The debut of DeepSeek-R1 demonstrated that even without OpenAI or Anthropic-scale funding, a Chinese company could reach the global frontier through technical efficiency.
Yet creating a single R1 model and operating a frontier AI company long-term are entirely different challenges. As competition becomes a protracted war, cost structures inevitably shift. Compute is the first pressure point. Frontier models require not just periodic pre-training but also reinforcement learning, post-training, model evaluation, and ever-growing inference resources. As the user base expands, compute spending no longer occurs only before a model's release but is continuously consumed with every API call. According to tech publication The Information, DeepSeek spent roughly 11 billion yuan on AI infrastructure in the first seven months of this year, covering server rentals equipped with AI chips, chip purchases, and other computing equipment, with 2025 expenditures in this category projected at 1.2 billion yuan.
Business scope and headcount are also expanding. DeepSeek is no longer just training foundation models. Recent job postings and disclosures show its Harness team listing approximately 150 openings, spanning server-side development, Agent elastic computing R&D, and other core technical areas. Public information indicates DeepSeek currently employs around 300 people, meaning this hiring wave equals half its existing workforce. Revenue growth cannot cover these massive outlays either. The Information data shows that in the first seven months of 2026, DeepSeek generated about 475 million yuan in revenue—roughly ten times its full-year 2025 figure—but still incurred a net loss of approximately 715 million yuan over the same period. Sources cited by Reuters suggest IPO proceeds will fund expanded compute infrastructure, model R&D, and talent acquisition. But DeepSeek isn't truly short on cash.
A Complex Capital Structure
If DeepSeek's problem were simply "having money," the answer would be relatively straightforward—there is an overabundance of willing investors. In its first funding round, DeepSeek designed a highly unusual transaction structure. Apart from a few institutions, most investors don't directly hold DeepSeek shares; instead, they invest indirectly through limited partnerships controlled by Liang Wenfeng. These investors have no voting rights and face a five-year lock-up period. The effect is clear: DeepSeek secures external capital while keeping capital as insulated from corporate control as possible, preserving Liang's tight grip on the company.
However, even with the lock-up and lack of voting power, capital keeps seeking entry points. The Financial Times reported that one Hong Kong family office investor received DeepSeek investment opportunities through eight different channels in a recent period. As investment demand balloons while shares remain scarce, this structure creates new problems. Some institutions that secured allocations have set up SPVs (special purpose vehicles) to bring in other investors. In essence, external investors don't directly own DeepSeek shares; they first invest in a vehicle created specifically for DeepSeek, which then funnels capital to an upper-level SPV or fund. As these structures layer atop one another, they create multi-tiered indirect holdings. Some intermediary vehicles charge front-end fees of 6% to over 15% and demand up to 40% of excess returns. This convoluted indirect ownership makes the ultimate sources of capital and affiliated relationships increasingly opaque.
The Financial Times also noted that Liang has begun personally vetting some end investors, aiming to maintain control and reduce governance risks ahead of the IPO. DeepSeek's challenge now extends beyond financing. On one hand, it needs longer-term capital sources and equity liquidity; on the other, as investors and shareholding layers multiply, maintaining clear, stable control while attracting capital has become a governance issue. An IPO cannot automatically resolve these tensions, but it offers a more public, standardized framework. The company gains a sustained financing tool, shareholders receive pricing and exit mechanisms, and the private-agreement-based relationship between capital and founder must now convert into publicly-traded corporate governance rules.
Pricing Beyond 500 Billion
The market's current valuation of DeepSeek is largely built on technological leadership and future potential. But a company valued at 500 billion yuan cannot—and will not—forever justify its price through successive model breakthroughs alone. After R1, DeepSeek's newer models have failed to recreate that industry-shaking moment. The delayed August release of DeepSeek V4 showed progress—officially achieving open-source leadership in reasoning and coding while maintaining notable cost advantages—but it didn't narrow the gap with the most advanced US models the way R1 did. Meanwhile, competition is increasingly shifting toward cost-performance. OpenAI cut GPT-5.6 Luna API prices by 80%, Alibaba Cloud slashed Qwen3.8-Flash call prices, and Tencent Cloud reduced pricing on some Hunyuan models. DeepSeek's sharpest historical edge—offering near-frontier model capability at rock-bottom inference costs—is being rapidly eroded. How long can its cost-performance narrative hold? Once in public markets, investors will scrutinize not just model capability but operational discipline—every question about spending, revenue, and growth becomes central to the company's continued valuation.