Chinese AI lab DeepSeek has closed its first-ever external funding round, raising more than 50 billion yuan, or roughly $7.4 billion. The valuation exceeds $50 billion — making the Hangzhou-based company China’s most valuable AI startup, according to consistent reporting. By international standards, however, the figure remains modest: US rival OpenAI recently closed a round of $122 billion at a valuation of around $852 billion, while Anthropic raised $65 billion.

The disparity is largely geopolitical. DeepSeek can only raise capital within China and has limited access to American hardware. According to an analyst cited by Reuters, it therefore makes little sense for the company to try to match the multi-billion-dollar compute budgets of its US rivals.

Where the Money Comes From

Until this round, DeepSeek had operated without outside investors. The lab was founded in July 2023 by Liang Wenfeng, who drew the necessary capital from his quant hedge fund High-Flyer and from private resources. Unlike Western competitors that relied early on venture capital or Big Tech cloud subsidies, DeepSeek financed itself entirely on its own until now.

The single largest backer in the current round is the founder himself: according to people familiar with the matter, Liang contributed roughly 20 billion yuan (about $3 billion). Tencent invested around 10 billion yuan (about $1.4 billion), and battery maker CATL contributed approximately 5 billion yuan. Rounding out the group is China’s National Artificial Intelligence Industry Investment Fund — the state vehicle Beijing uses to channel capital into strategic technology sectors.

A Deal Structure Unlike Anything Before

More notable than the sum is how the round was constructed. Commercial investors were not allowed to put their capital directly into DeepSeek; instead, it flowed into a limited partnership managed by CEO Liang Wenfeng. As a result, they receive no direct equity in the company, no voting rights, and are subject to a five-year lock-up — secondary sales or quick exits are off the table. Tencent’s $1.4 billion check accordingly buys the conglomerate no governance influence whatsoever.

There is one exception: the state-linked National Artificial Intelligence Industry Investment Fund received direct corporate ownership, voting rights, and freedom from the lock-up. It bypasses the limited-partnership wrapper entirely. Liang reportedly also personally vetted the identities of the backers behind the investing funds, to ensure that unknown — for instance, foreign — capital could not end up holding shares.

The structure is plainly designed to preserve the founder’s control and to seal off the investor base. Observers therefore read the round — anchored by the founder and a roster of major domestic firms — as much as a strategic statement as a financial one. It remains open whether Liang’s tight grip on control — no external board, no investor pressure — becomes an asset or a liability as DeepSeek scales further. Critics also point to staff turnover: several key researchers have recently left the company for higher-paying roles at Xiaomi or ByteDance.

Where DeepSeek’s Models Stand in the Benchmarks

DeepSeek’s rise is built on an open model strategy: by publishing its model weights freely, the lab has effectively turned the model layer into a commodity, forcing the market into a competition over cost efficiency rather than raw capability.

In the current rankings, DeepSeek’s models position themselves as the most attractively priced option in the near-frontier range, without quite reaching the very top. On Chatbot Arena (LMArena), which is based on millions of blind user comparisons, the leading models sit in an Elo range of roughly 1,450 to 1,560; DeepSeek’s reasoning models land at the lower edge of this frontier band and score particularly well in the math and logic categories.

On the Artificial Analysis Intelligence Index, which combines reasoning, speed, and cost, DeepSeek V4 Pro ranks well above the median for comparable open-weight models, but in absolute terms remains behind the leading proprietary systems from Anthropic, OpenAI, and Google. Its central selling point is price: at around $0.45 per million input tokens, DeepSeek V4 Pro is among the cheapest models in the near-frontier segment, undercutting the top providers by roughly an order of magnitude. On specialized comparison lists of Chinese models, DeepSeek V4 Pro leads ahead of competitors such as GLM-5.1, Kimi K2.6, and Qwen.

In short: powerful and extremely cheap, but not yet quite at the absolute top.

Anthropic Accuses DeepSeek of Illicit Distillation

There is, however, an accusation hanging over DeepSeek that drew considerable attention in February 2026. US provider Anthropic, the developer of the Claude model, accused DeepSeek — together with the Chinese labs Moonshot AI and MiniMax — in a blog post of extracting capabilities from Claude on an “industrial scale.” Specifically, the three firms are alleged to have generated more than 16 million interactions with Claude through roughly 24,000 fraudulently created accounts — a violation of the terms of service and of regional access restrictions, given that Claude is not commercially available in China. To circumvent the blocks, the companies allegedly relied on commercial proxy services.

At the center is the concept of distillation: a process in which a smaller “student” model is trained on the outputs of a stronger “teacher” model. Anthropic stresses that distillation is a widespread and legitimate training method — for example, when labs create smaller versions of their own models. It becomes illegitimate, the company argues, only when competitors use it to siphon off others’ capabilities in a fraction of the time and cost. Anthropic attributed the bulk of the traffic to MiniMax (more than 13 million interactions); DeepSeek is said to have conducted around 150,000 targeted exchanges aimed at Claude’s reasoning abilities and at generating censorship-compliant answers to sensitive topics. Anthropic says it bases the attribution on IP correlations, metadata, infrastructure indicators, and corroboration from industry partners.

Anthropic tied the allegations to national security: illicitly distilled models could lack key safety safeguards and might be used, for instance, for offensive cyber operations, disinformation, or mass surveillance. The company called on AI firms, cloud providers, and lawmakers to coordinate a response.

The accusation should be placed in a broader context: OpenAI had earlier charged DeepSeek with “free-riding” on the capabilities developed by American frontier labs, and Google had noted an increase in “distillation attacks” on its models. The named Chinese companies — DeepSeek included — have so far not responded publicly to Anthropic’s allegations. Critics, for their part, pointed to the ambivalence of the debate: the line between legitimate and illicit learning from model outputs is legally blurry, and labs such as Anthropic have themselves trained their own models on large amounts of third-party data. Some observers also read the escalation as positioning toward Washington against the backdrop of the ongoing debate over chip export controls.