DeepSeek has not taken outside money since it was founded in 2023. For two years it turned down every venture capital firm and major tech company that came calling, funding its research entirely from the returns of its parent hedge fund, Zhejiang High-Flyer Asset Management, which reportedly posted a 56.6% return in 2025. That era is ending. The lab is now in the final stages of discussions for its first external financing round, targeting roughly 70 billion yuan — approximately $10.29 billion — that would value it at around $45 billion pre-investment.

The round is still in flux. Final investment amounts, participants, and the pre-money valuation have all been described as subject to change. But the direction of travel is clear enough that it is worth thinking through what this means for DeepSeek’s trajectory, and what it signals about the broader competitive landscape.

What the Round Actually Looks Like

The headline number is the total round target of around $10 billion, though the actual external check size is considerably smaller. DeepSeek is seeking at least $300 million in outside capital; founder Liang Wenfeng may personally contribute around 20 billion yuan (~$2.75 billion) of the total. The National Artificial Intelligence Industry Investment Fund — a Beijing-backed vehicle for strategic AI projects — is reportedly in talks to invest approximately 10 billion yuan. Tencent, IDG Capital, Monolith Capital, JD.com, and NetEase are among the names cited as close to or in discussions about participating.

The pre-money valuation has moved quickly. Reports from early May 2026 put it at $20 billion. By late May, Bloomberg was citing a figure closer to $45 billion. That kind of jump in a matter of weeks reflects how much investor interest has accelerated once it became clear that Liang was willing to accept outside capital at all.

The Pitch: AGI, Not Near-Term Revenue

What Liang has told potential investors matters as much as the numbers. According to people familiar with the discussions, the core message has been that DeepSeek intends to remain research-first, will keep releasing open-source models, and is pursuing artificial general intelligence as its primary goal rather than short-term commercialization.

That framing is striking for a company raising at a $45 billion valuation. It positions DeepSeek closer to the self-described lab missions of OpenAI and Anthropic than to a typical growth-stage enterprise software company. Whether investors price that as a feature or a risk depends on what you think the economics of frontier AI development look like over the next few years.

The practical reason for seeking capital is straightforward: training runs at the scale DeepSeek now operates have moved beyond what even a highly profitable hedge fund can self-finance indefinitely. The Information has separately reported that the lab’s longer-term ambitions could involve raising $7 billion or more as it moves toward recurring revenue. The current round is the first step in that direction.

DeepSeek’s Leverage Coming Into This Round

The leverage DeepSeek brings to investor conversations is unusual. The R1 model released in January 2025 matched OpenAI’s top reasoning model at a reported training cost of around $6 million. A figure that erased roughly $600 billion from Nvidia’s market capitalization in a single session when the implications landed. V3 and its successors extended that reputation for doing more with less.

The V4 family, released in April 2026, continued the pattern. DeepSeek-V4-Pro is a 1.6-trillion parameter Mixture-of-Experts model that ranks competitively against frontier closed-source systems on LiveCodeBench and SWE-bench Verified. Notably, the V4 series was optimized to run on Huawei Ascend and Cambricon silicon as well as Nvidia. That constraint has, counterintuitively, produced architectural innovations (Compressed Sparse Attention, mixed FP4/FP8 precision training, the Muon optimizer) that make DeepSeek’s models more efficient per FLOP than many of its peers.

For investors, the bet is that this track record of efficiency compounds. A lab that produces frontier-quality models at a fraction of US lab costs, while remaining open-source and therefore building ecosystem goodwill, is a different kind of asset than a closed-model API business.

What This Changes — and What It Probably Does Not

The most immediate consequence of the round, if it closes as reported, is that DeepSeek gets a meaningful runway for the next generation of training runs. Frontier model training costs are scaling rapidly; the kind of compute investment required to stay at the frontier in 2027 or 2028 is materially higher than what was required for R1 in early 2025. Capital matters here in a direct way.

What is less clear is whether the open-source commitment survives at scale. Liang has stated explicitly that releasing open weights is the plan. The V4 family shipped under an MIT license. But the economics of releasing model weights change when training a single run costs hundreds of millions of dollars. US labs that started with open intentions walked those commitments back as model costs climbed. It would be worth watching whether the same pressure eventually applies.

The investor composition is also worth watching. The National Artificial Intelligence Industry Investment Fund is not a passive LP. It is a policy instrument. State participation at this scale in a frontier AI lab creates alignment with national priorities that may or may not map to what international developer communities find most useful. DeepSeek has so far navigated this tension better than many Chinese tech companies but the capital structure is changing.

The Broader Signal

DeepSeek’s financing round does not exist in isolation. It is happening alongside a broader reorientation of the AI industry around what the open-weights ecosystem can produce. In 2024, the conventional wisdom was that frontier AI required closed models because the investment was too large to give away. DeepSeek’s output has challenged that assumption at the level of model quality. This round, if it closes, challenges it at the level of capital formation too — demonstrating that you can raise at a $45 billion valuation on an open-source, research-first pitch.

For developers building on open-weight models today, the implication is relatively clear: the lab that has been one of the most consistent suppliers of accessible frontier-quality models is about to have significantly more resources. If the research-first, open-source posture holds, the downstream effect is more capable models available sooner and at lower inference cost than a world where this capital flowed entirely to closed-model API providers.

Whether that plays out depends on a lot of moving parts that are not settled. The round has not closed. The composition of investors is still in flux. And the relationship between research ambition, open-source commitments, and investor expectations at a $45 billion valuation is a dynamic that no lab has navigated at quite this scale before.

What is not in question is that DeepSeek is one of the most consequential AI labs operating right now, and that the capital it is about to receive will determine what it can build next.

DeepSeek Models on DeepInfra

DeepSeek V4 Pro and V4 Flash are available now on DeepInfra. V4 Pro is priced at $1.30 per million input tokens and $2.60 per million output tokens, with cached input at $0.10 per million — significantly below closed-source alternatives with comparable benchmark performance. You can access both models via the DeepInfra API with your existing credentials.

Explore DeepSeek on DeepInfra: deepinfra.com/deepseek