At a glance: DeepSeek. Founded July 2023 by Liang Wenfeng (also co-founder of quantitative hedge fund Zhejiang High-Flyer Asset Management). Headquarters: Hangzhou, China. Funding history: zero outside capital until 2026 — entirely self-funded from High-Flyer’s investment returns. Current raise: first-ever VC round, targeting $300M+, at a valuation reported near $45 billion. Lead investor: China Integrated Circuit Industry Investment Fund (“Big Fund”), China’s state-backed semiconductor investment vehicle. Other parties in talks: Tencent, Alibaba. Status as of May 2026: not formally closed, still in negotiations. Part of our AI Models & Companies reviews.

On January 27, 2025, Nvidia lost $589 billion in market value in a single trading session.

It remains the largest single-day market cap loss for any company in the history of the stock market.

The proximate cause was a Chinese AI lab called DeepSeek, which had released a model — R1 — that appeared to match OpenAI’s best reasoning model at a fraction of the cost, trained on chips that were hobbled by US export restrictions. The original claim was $6 million in training costs. More careful analyses later put the true figure higher — Semianalysis estimated $500 million or more when all development costs are counted — but the signal was impossible to dismiss: the efficiency gap between American and Chinese AI was narrower than the market had assumed.

What almost no one mentioned in the coverage that week: the company behind R1 had never taken a dollar of outside investment.

A Hedge Fund’s Side Project That Changed Everything

DeepSeek did not begin as a startup in the conventional sense.

Its parent company is Zhejiang High-Flyer Asset Management, one of China’s largest quantitative hedge funds. High-Flyer was co-founded in 2015 by Liang Wenfeng — a graduate of Zhejiang University who had spent years applying machine learning to quantitative trading strategies. By 2021, High-Flyer was buying thousands of Nvidia GPUs, initially as infrastructure for its own trading research.

What started as an internal AI research function became a dedicated organization. In July 2023, High-Flyer’s AI research lab was spun off into an independent entity: DeepSeek. Liang Wenfeng serves as CEO of both companies. The corporate structure is opaque by Western standards — ownership flows through shell companies and the High-Flyer partnership — but the essential fact is simple: DeepSeek spent its first three years operating as an extension of a profitable private fund, with no external investors to answer to and no pressure to raise money it didn’t need.

That arrangement produced R1. It also produced V3, a series of increasingly capable open-weight models that repeatedly outperformed expectations about what was achievable on restricted hardware. By early 2026, DeepSeek’s models had become the dominant Chinese AI option on OpenRouter and other model routing platforms, with usage levels challenging established Western models in several categories.

Why Raise Now?

For the first two and a half years of DeepSeek’s existence, the question wasn’t “when will they raise?” — it was “why would they?”

High-Flyer had the capital. DeepSeek had no product requiring a venture capital growth curve, no series of fundraising rounds to justify its existence, no investor board demanding a roadmap to monetization. It could operate at research speed rather than startup speed, which may be part of why it moved so efficiently.

The answer to “why now” is more prosaic than geopolitics: talent retention.

To grant equity to employees — to let the researchers and engineers who built R1 and V3 participate in the company’s value — DeepSeek needs a formal valuation event. Without external investors establishing a price per share, there is no mechanism for employee stock to mean anything. Liang Wenfeng reportedly decided to raise outside capital primarily to solve this problem: give the people who built the company something to own.

That motivation matters for understanding what kind of deal Liang is willing to accept. Reports from April 2026 indicated he was resistant to Tencent’s proposal to acquire as much as 20% of the company — a stake that would give the tech giant significant influence over a competitor in its own AI product development. DeepSeek’s willingness to take state money from the Big Fund, rather than cede control to commercial tech giants, may reflect a deliberate choice about who to be accountable to.

How the Valuation Got to $45 Billion

The initial reporting on DeepSeek’s fundraising, from late April 2026, described talks with Tencent and Alibaba at a valuation of $20 billion or more. That figure was already significant for a company that had never taken outside money and does not have a conventional revenue model.

Then the Big Fund entered the picture.

Within weeks, the valuation in discussions had more than doubled — to approximately $45 billion. The China Integrated Circuit Industry Investment Fund, established by the Chinese government in 2014 specifically to build domestic semiconductor capability in response to US technology restrictions, was reported to be in talks to lead the round.

The Big Fund’s involvement is different in kind from Tencent or Alibaba’s interest. Commercial tech companies invest for returns and strategic positioning. The Big Fund is a policy instrument — a vehicle for deploying state capital in industries the Chinese government has designated as strategically essential. Its previous investments have been concentrated in semiconductor manufacturing, chip design, and the equipment supply chain. DeepSeek would be a new category: AI software infrastructure.

The signal is not subtle. Beijing is treating DeepSeek as national AI infrastructure, not just a successful startup.

The Huawei Connection

There is a hardware dimension to this story that the valuation figures alone don’t capture.

After the initial export restriction rounds in 2022 and 2023, Nvidia’s H100 and A100 chips — the standard training hardware for large AI models — became effectively unavailable to Chinese AI labs. DeepSeek’s efficiency breakthroughs were, in part, a response to this constraint: if you cannot get the best chips, you optimize harder.

That optimization extended to Huawei’s Ascend chips. DeepSeek’s inference pipeline has been substantially optimized to run on Huawei hardware, and Huawei’s Ascend 910 series — particularly the 950PR — has reportedly become one of the primary inference platforms for DeepSeek’s deployed models. In parallel, Huawei’s AI chip revenue has grown significantly, with some reports indicating it has overtaken Nvidia in Chinese market share for AI inference.

The result is a domestic AI stack: DeepSeek models running on Huawei hardware, trained and deployed without meaningful dependence on US technology at any layer. If that stack holds its own against US-trained models — and the benchmark evidence suggests it does in several key areas — it represents the first credible proof that China’s AI trajectory is not terminally constrained by export controls.

The Big Fund’s apparent interest in DeepSeek is coherent against this backdrop. An investment in DeepSeek is also, indirectly, a bet on Huawei’s AI chip business, on the domestic stack’s viability, and on the proposition that China can build frontier AI capability on indigenous infrastructure.

What We Don’t Know

As of late May 2026, the round has not been formally announced.

Valuations reported by the Financial Times, Bloomberg, and TechCrunch — ranging from $20 billion to $45 billion — are based on conversations in progress, not signed term sheets. The figures could change. The round could be restructured. Liang Wenfeng could decide he doesn’t need outside capital after all.

Several important details remain unclear:

  • Round size: $300 million has been cited, but this figure has shifted alongside the valuation. A $45 billion valuation with a $300 million round implies a roughly 0.7% stake, which is unusually small for a lead investor. The terms may be different from what has been reported.
  • Revenue model: DeepSeek does not have an obvious commercial revenue model. Its API is priced at a fraction of comparable US services. Its models are largely open-weight. The company has not announced an enterprise product, a cloud partnership, or a consumer service that generates material revenue. How a $45 billion valuation is sustained — or justified — by future revenue is not publicly explained.
  • Liang’s control appetite: Existing reporting suggests he is reluctant to cede significant ownership. How much influence the Big Fund requires in exchange for leading the round is unknown.

What This Round Would Signal

If the round closes near $45 billion, it would make DeepSeek the most valuable AI startup in China and one of the largest AI companies in the world by private market valuation — behind Anthropic’s $900 billion fundraising discussions and OpenAI’s approximately $300 billion valuation, but ahead of most other Western AI companies.

More importantly, it would formalize something that has been true in practice for several months: DeepSeek is no longer just a research lab. It is a strategic asset.

The nature of the Big Fund’s investment — if it closes as reported — means DeepSeek will have a direct channel to Chinese state infrastructure: to national compute resources, to state-controlled distribution networks, and to policy environments that could accelerate domestic deployment. That is a different kind of capital than Tiger Global writing a check for Sierra.

For anyone trying to understand the long-term shape of the China-US AI competition, DeepSeek’s first funding round is worth watching more carefully than its model benchmarks. The benchmarks tell you what the models can do. The funding structure tells you who is treating AI as national infrastructure — and what that actually means for how it will be built, deployed, and governed.

This analysis is based on reporting from the Financial Times, Bloomberg, TechCrunch, and The Information. The funding round described has not been formally announced, and figures may change as negotiations continue. We did not interview company representatives or obtain information beyond public reporting.