The most expensive AI lab in the world and the most efficient one are both raising money at the same time. OpenAI filed for a trillion-dollar IPO in late May while losing $1.22 for every dollar of revenue. DeepSeek, the Chinese lab that claims it trained a frontier-class model for $6 million, is raising $7.4 billion in its first outside funding round at a valuation between $52 billion and $59 billion.
The round is notable for two reasons: it ends three years of complete self-funding, and the investors include a battery manufacturer, a gaming company, and China’s sovereign AI fund. That composition reveals more about the state of the global AI race than the valuation itself.
Three Years of Self-Funding, Then a $7.4 Billion Round
DeepSeek was founded in 2023 as a side project of High-Flyer, the quantitative hedge fund run by Liang Wenfeng. In 2021, Liang began stockpiling thousands of Nvidia GPUs for AI research. Two years later he spun out the AI lab that would become one of China’s most consequential AI companies.
For three years, DeepSeek operated without a single external investor. Liang funded everything from High-Flyer’s trading profits, giving him complete control over research direction and release strategy. That independence produced the V3 and R1 models in early 2025, which reshaped global assumptions about AI compute efficiency by demonstrating that near-frontier performance was possible at a fraction of Western training budgets.
The maiden round targets 50 billion yuan ($7.4 billion), with a post-money valuation of 350 to 400 billion yuan ($52 billion to $59 billion). CNBC reports the round is expected to close within weeks, with fewer than ten investors participating.
Liang himself is committing 20 billion yuan ($3 billion) of personal capital, roughly 40% of the total raise. That commitment keeps him firmly in control of a company he has run with unusual independence, even by Chinese tech standards.
The Investor Composition Is the Signal
The remaining $4.4 billion comes from a strategically curated group of domestic backers, each bringing something beyond capital.
Tencent is considering 10 billion yuan ($1.5 billion), which would make it the largest external investor. Tencent operates WeChat (1.3 billion monthly users), one of China’s largest cloud platforms, and a massive gaming infrastructure. For DeepSeek, Tencent provides distribution at scale and cloud infrastructure for API serving. For Tencent, the investment secures access to China’s strongest foundation model as the company rebuilds its own AI product stack.
CATL, the world’s largest electric vehicle battery manufacturer, is in for 5 billion yuan ($740 million). A battery company backing an AI lab makes little sense until you consider what both companies need: massive amounts of electricity. CATL has been building energy storage systems for data centers and positioning its power technology for the AI infrastructure buildout. The DeepSeek investment gives CATL a commercial relationship with a company that will need gigawatts of power as it scales compute.
China’s National AI Industry Investment Fund is in final discussions to participate. State backing signals that Beijing views DeepSeek as a strategic national asset, not just a commercial venture.
NetEase (gaming) and JD.com (e-commerce) are also in late-stage talks, along with IDG Capital and Monolith Capital. The investor base is entirely domestic. No Western venture capital, no Gulf sovereign wealth funds, no New York crossover hedge funds. The Next Web reported that this is deliberate: the cap table is tight, the alignment is strategic, and the participants represent a cross-section of China’s tech, energy, and government infrastructure.
Six Percent of Anthropic, 90% of the Benchmarks
At $59 billion, DeepSeek is valued at roughly 6% of Anthropic’s $965 billion and 7% of OpenAI’s $852 billion. The valuation gap is enormous. The benchmark gap is not.
DeepSeek V4 Pro, released in April 2026, scores 90.1% on GPQA Diamond (graduate-level reasoning), compared to 93.6% for GPT-5.5. On SWE-bench Pro (real-world GitHub issue resolution), V4 Pro hits 55.4% versus 58.6% for GPT-5.5. On Humanity’s Last Exam, DeepSeek scores 37.7% to GPT-5.5’s 41.4% and Claude Opus 4.7’s 46.9%. VentureBeat described V4 as delivering “near state-of-the-art intelligence at one-sixth the cost” of Opus 4.7 and GPT-5.5.
The pricing tells the sharper story. DeepSeek V4 Flash runs at $0.14 per million input tokens. V4 Pro costs $1.74 per million input tokens. Anthropic’s Opus 4.7 charges roughly $15 per million input tokens for comparable capability. For developers and enterprises running inference at scale, that differential compounds fast.
DeepSeek built this competitive position under U.S. semiconductor export restrictions that limit Chinese companies’ access to advanced Nvidia chips. The company trained its models on older-generation H800 GPUs and developed software optimization techniques that squeeze more performance per chip than Western labs typically achieve. The Stanford AI Index 2026 placed China just 2.7% behind the U.S. on composite AI capability metrics, and DeepSeek’s benchmarks are a primary reason why.
127 Million Users, $470 Million in Profit
The business fundamentals explain why investors are writing large checks at what looks, by Western AI standards, like a modest valuation.
DeepSeek reached 127 million monthly active users by March 2026 and generated an estimated $470 million in net profit during 2025, representing a 28% to 32% net profit margin. OpenAI, by comparison, lost approximately $5 billion in 2025 on roughly $4 billion in revenue.
The profitability is structural. DeepSeek’s open-weight model release strategy drives adoption at minimal customer acquisition cost. The company charges for API access at prices low enough to undercut every Western competitor, yet margins remain healthy because the underlying training and inference costs are proportionally lower.
The open question is whether $7.4 billion in outside capital changes this dynamic. External investors bring commercial expectations, and DeepSeek’s identity was built on cheap, open models released freely. Liang’s $3 billion personal stake and controlling position suggest the strategy will not change immediately. But the introduction of Tencent, CATL, and a sovereign fund into the cap table creates stakeholder pressure that did not exist when one hedge fund manager was financing everything himself.
What This Round Changes
Three implications stand out.
First, the performance gap between Chinese and Western AI models continues to narrow while the valuation gap widens. A 16:1 valuation premium for Anthropic over DeepSeek only makes sense if Western labs maintain a permanent quality lead. V4’s benchmarks suggest the lead is measured in months, not years.
Second, China’s AI investment thesis is converging around infrastructure. The DeepSeek round looks less like a startup fundraise and more like a coordinated national project, with tech platforms (Tencent, NetEase, JD.com), energy companies (CATL), and the state all directing capital toward a single lab. This mirrors the sovereign AI strategies emerging in Europe, but with deeper coordination and faster execution.
Third, every open-weight model that matches closed-source competitors at one-sixth the price puts structural pressure on Western lab revenue. If V4’s successor narrows the remaining three-to-six-point benchmark gap while maintaining that pricing, OpenAI and Anthropic face their most serious margin threat yet.
DeepSeek spent three years proving that frontier AI does not require frontier spending. Now it has $7.4 billion to test whether the same thesis holds at the next scale.