Why DeepSeek's $45B Big Fund Round Changes Its Open-Source Calculus

Photo by Xiangkun ZHU on Unsplash

The DeepSeek funding round now in talks with China's state-backed Big Fund would mark the first external capital the lab has ever taken, and the structure of that capital matters more than the headline number. According to TechNode's reporting on the Financial Times scoop, the China Integrated Circuit Industry Investment Fund is leading discussions that could value DeepSeek at roughly $45 billion. Weeks earlier, the same conversations started near $10 billion.

That re-pricing in under a month is the story almost nobody is unpacking correctly.

The Number Tripled in Three Weeks, and That Tells You Who Is Bidding

What began in mid-April 2026 as a $300 million raise at a $10 billion valuation, with Alibaba and Tencent in early talks, escalated within weeks into an FT-reported deal at $45 billion led by the China Integrated Circuit Industry Investment Fund, per The Next Web. TechCrunch's account confirms the Big Fund is leading, with Tencent and Alibaba reportedly still in the mix as participants rather than anchors.

A 4.5x mark-up in three weeks is not a market clearing process. It is a bidder of last resort deciding the asset is strategic and pricing accordingly. Private AI rounds with multiple credible buyers do not move that fast on fundamentals.

The pricing also matters because of where it lands DeepSeek on the global league table. A $45 billion valuation would place DeepSeek among the most valuable AI companies in the world, on par with several well-capitalised U.S. AI startups, despite having taken no venture funding and generating no disclosed revenue, according to Capacity Global. Ground News, citing Asia Trade, reported the actual range as $45 to $50 billion, suggesting the FT figure is a floor.

What the Big Fund Is, and Why Its Mandate Just Quietly Expanded

The China Integrated Circuit Industry Investment Fund is not a generalist sovereign vehicle. It was established to advance China's semiconductor industry and has backed major chip players including SMIC; leading a round for a frontier AI lab represents a notable broadening of its mandate. Capacity Global notes that the Big Fund has not publicly backed any of China's large language model developers before this deal.

That is the part to sit with. A fund built to finance fabs and chip designers is now writing the largest check ever directed at a Chinese model lab. The implication is that Beijing's planners have stopped treating chips and models as separate problems.

The Next Web frames it bluntly: if China cannot acquire Nvidia's leading-edge GPUs, it will finance model labs that produce frontier results without them. The Big Fund's involvement converts that thesis into capital allocation policy.

The Open-Source Position Was Always Subsidised by Independence

DeepSeek's open-weight strategy did not happen because Liang Wenfeng is a romantic. It happened because the lab had no outside investors asking awkward questions about monetisation. Until now, freedom from external investors has been central to DeepSeek's operating philosophy; with High-Flyer as its sole backer, the lab pursued long-term research without commercial pressure, a structure Liang has credited with enabling the architectural innovations behind its low-cost training approach, Capacity Global reports.

DeepSeek was founded in July 2023 by Liang Wenfeng, a 40-year-old computer scientist and co-founder of quantitative hedge fund High-Flyer Capital Management, and has been funded entirely from High-Flyer's balance sheet until now. That balance-sheet model is what made V4-Pro and V4-Flash shippable under an MIT license on Hugging Face. There was no cap table to apologise to.

That changes the moment a strategic state investor takes a position. State capital does not necessarily demand closure of weights, but it does demand alignment between corporate strategy and national strategy. Those two have been congruent so far. They will not always be.

The Best Argument That Open Weights Survive, and Where It Breaks

The strongest case that the DeepSeek $45 billion valuation does not threaten its open-weight stance is straightforward: open weights are the strategy. Releasing V4 under MIT, with full-stack adaptation across Huawei Ascend, Cambricon, Hygon, and Moore Threads on launch day per TrendForce, is exactly how you seed a domestic compute stack. Closed weights would defeat the purpose of state backing in the first place, since the policy goal is ecosystem adoption, not licensing revenue.

CNTechPost lays this out: backing from China's most strategic government chip fund would reinforce DeepSeek's position as a leader in China's frontier AI sector while promoting development of a Chinese tech ecosystem comprising domestic models, software, and chips. Open weights are a feature of that plan, not a bug.

That argument holds until the geopolitical temperature rises. The vulnerability is dual-use export logic. If U.S. or allied governments classify frontier model weights as a controlled item, a state-funded Chinese lab becomes legally riskier to host on Hugging Face than an independently-funded one. The model itself is identical. The provenance is what changes the regulatory analysis. That is the failure mode for the DeepSeek open source future, and it does not require Beijing to change anything about its own posture.

The Hardware Story That Makes the Money Make Sense

The China Big Fund AI investment lines up with one specific fact about V4 that most coverage of the funding round has buried. DeepSeek V4 has been optimized for inference on Huawei's Ascend 950PR chips, with full-stack adaptation delivering up to 2.87x the performance of Nvidia's China-specific H20 processor, per the Asia Trade reporting summarised by Ground News.

China's largest internet groups, Alibaba, ByteDance, and Tencent, have collectively ordered hundreds of thousands of Ascend 950 processors following the V4 release, the same source reports. That is a domestic AI buildout being ordered into existence by a single model release, on chips the Big Fund has spent a decade subsidising.

From a portfolio-construction standpoint, a Big Fund stake in DeepSeek is the missing piece. The fund already owns exposure to Ascend silicon through its SMIC and Huawei-adjacent positions. What it lacked was exposure to the demand side, the workload that justifies the silicon investment. DeepSeek is that workload. For broader context on how compute capital is reshaping AI economics, see our coverage in AI Industry articles and AI Infrastructure articles.

What State-Backed Funding Actually Constrains

DeepSeek state backed funding does not need to ban open weights to constrain the lab. It constrains the boring things first.

• Where compute is procured. State capital arriving alongside Ascend optimisation removes any future option to quietly rebuild on Nvidia hardware if Ascend underperforms.
• Who serves the API at scale. The hyperscaler order book of Alibaba, Tencent, and ByteDance becomes the only realistic distribution path inside China.
• What the V5 release schedule looks like. Winbuzzer flagged that whether DeepSeek's reported funding talks convert into the multi-year capital base needed for a V5 cycle that does not run another 484 days is a real decision gate. State money buys faster cycles. Faster cycles raise the strategic stakes per release.
• Whether weights go up on Hugging Face on day one or after a domestic-platform exclusivity window. That is the most likely first compromise, and it would be almost invisible to most users.

None of those are dramatic. They are the directions a research lab quietly drifts when its largest shareholder is a sovereign chip-policy vehicle.

The Counterargument That Actually Deserves a Response

The strongest objection to this analysis is that DeepSeek is too valuable to Beijing as an open-source standard-bearer to risk closing it. The R1 release in January 2025 reset global expectations on AI training cost and triggered a U.S. tech-stock selloff once DeepSeek shot to prominence in January 2025 after releasing R1, a powerful open-source LLM that trained on a fraction of the compute and cost of U.S. rivals like OpenAI and Anthropic. Forcing it private would burn that asset.

That objection is correct on direction and wrong on magnitude. The choice is not binary between MIT-licensed Hugging Face releases and a closed proprietary API. The middle ground is large: delayed-weight releases, distillation-only public artefacts, region-locked licensing, or a published-weights-but-restricted-training-data posture. Any of those preserves the propaganda value of open while quietly tightening control over downstream use. That middle ground is exactly where state-investor incentives push a lab.

What to Watch Over the Next Two Quarters

The specific test is V5. If V5 ships under the same MIT license, on the same day, on Hugging Face, with the same day-zero adaptation across Huawei Ascend, Cambricon, Hygon, and Moore Threads, the open-source thesis survives the funding round intact. If V5 launches first as an Ascend-cloud exclusive with weights following on a delay, the calculus has already shifted.

The second tell is governance disclosure. State investment in Chinese tech firms typically comes with board observer rights and approval thresholds on strategic decisions. Whether DeepSeek discloses any of that publicly will indicate how the lab wants its independence narrative read internationally.

DeepSeek's open-weight era was funded by one quant fund's willingness to lose money on research for two years. That era is ending whether or not the weights keep shipping. The Big Fund is not buying a stake in a research lab. It is buying the demand-side anchor for a domestic chip industry, and the price of that anchor is that the lab stops being the kind of organisation that releases frontier weights because nobody is empowered to stop it.

The model files on Hugging Face will keep updating for now. Read the licenses carefully when V5 lands.