How to Value DeepSeek: Throw Out the Model Company Playbook
The biggest mistake you can make: treating it like a model company's fundraise.
Within three weeks, DeepSeek’s valuation appeared in public reports four separate times. Each time, it was higher.
Early April: a seed round at roughly $10 billion. April 22: reports put it above $20 billion. May 6: word leaked that the “Big Fund” was in talks to lead at around $45 billion. Last week, a ceiling above $50 billion surfaced — which, if it closes, would be the largest single financing round in the history of Chinese AI.
The biggest check isn’t coming from a VC or an internet giant. It’s reportedly from DeepSeek’s founder, Wenfeng Liang, himself: up to 20 billion RMB (roughly $2.8 billion), about 40% of the total raise. By injecting capital directly, he would reportedly lift his direct stake from 1% to 34%, and with indirect holdings factored in, control approximately 84.29% of the company.
The “Big Fund”, formally the National Integrated Circuit Industry Investment Fund, is the state capital that built China’s semiconductor industry. It backed SMIC, China’s largest chipmaker, and YMTC, the country’s leading memory manufacturer. It has never publicly invested in a large language model company. If this round closes, it would be the first time.
For years, DeepSeek’s most recognizable quality was what it refused to do: no fundraising, no commercialization, no roadshows. High-Flyer, the quantitative hedge fund that owns DeepSeek and bankrolls its research, always had enough capital. DeepSeek’s moves toward commercialization have been essentially zero, which made this valuation look strange even by the standards of other AI companies already trading on fairy-tale multiples.
Put these facts side by side, and no standard model company framework explains them. Why does a valuation quintuple in three weeks? Why does a semiconductor state fund show up? Why does the founder write the biggest check himself?
This round needs a different framework.
DeepSeek is not another “better model company.” It looks like one, but it functions more like an infrastructure company wearing a model company’s clothes. This financing round makes that identity visible for the first time.
1.
When Wenfeng Liang talked about fundraising in 2024, he said: “Our problem has never been money. It’s been the embargo on advanced chips.”
For a typical model company, chips are a cost line, not a strategic constraint. Calling chips the core problem means that from day one, DeepSeek wasn’t only asking how to build better models. It was asking how to rebuild a functioning system under conditions where compute was controlled by others.
Every major technical contribution DeepSeek has made traces back to this. Multi- MLA, MoE, FP8 training, extreme inference efficiency…… beneath all of them is the same question: how do you build frontier models with fewer chips, and with chips you’re not supposed to have? That’s infrastructure R&D logic, not model company logic.
So if DeepSeek isn’t a model company, what should it be priced against? Not Kimi, Zhipu, or MiniMax — the well-funded Chinese LLM startups that are DeepSeek’s obvious peers by surface appearance. The financing structure that leaked points somewhere else entirely.
2.
On April 24, DeepSeek released V4. For the first time, its technical report placed Huawei’s Ascend NPU and Nvidia’s GPU on the same hardware validation list: “We have verified the fine-grained Expert Parallelism scheme on both the Nvidia GPU and Huawei Ascend NPU platforms.” A trillion-parameter model, officially certified on domestic AI chips in a public document.
On release day, eight domestic chip manufacturers, including Huawei Ascend and Cambricon, China’s two most prominent domestic AI chip companies, completed integration. Previous work at this scale had taken months. DeepSeek had spent the preceding months in close collaboration with both companies, reportedly pushing back the release to get it right and rewriting significant portions of the model’s lower-level architecture to make it happen.
The financing kicked off during this same window. The timing isn’t incidental. Most companies raise to do something. DeepSeek raised because it had already done something, publicly, verifiably. Wenfeng Liang used a model release to complete what normally takes multiple investor meetings to establish.
Three checks. Three separate logics. None of them overlapping.
The Big Fund’s entire investment history is things China cannot afford to be without: SMIC, YMTC, the bedrock of domestic semiconductor manufacturing. The reason it never touched model companies before was simple: commercial return timelines for AI models are uncertain, and the logic that works for chip fabs doesn’t translate. DeepSeek’s V4 release dissolved that objection. The fund can now describe this as a bet on “whether the domestic compute ecosystem can actually run” — a strategic inflection point where a frontier model accelerates domestic chip development and closes the loop between domestic silicon and domestic AI. Framed that way, it belongs in the same category as SMIC. That’s a language the Big Fund was built to speak.
Tencent and other industrial investors reportedly taking stakes have a simpler motive: they don’t want to find themselves outside the ownership structure if DeepSeek becomes China’s AI infrastructure layer. Not being on the cap table means being a paying customer. And what they bring isn’t just capital —it’s ecosystem access, enterprise customers, cloud infrastructure, which happen to be exactly where DeepSeek is weakest commercially.
This also explains why several other tech giants with their own infrastructure ambitions probably won’t invest. They already see DeepSeek as a potential competitor at the infrastructure level, not a model company worth a check.
3.
The result is a strange duality. On one side: national strategic capital, locking in the infrastructure identity. On the other: industrial ecosystem capital, unlocking the commercial path. In most companies, these two goals would be in tension. The financing structure suggests DeepSeek is trying to run them in parallel, fully independent, neither one holding the other back.
After this round, in other words, how fast DeepSeek commercializes and whether its larger strategic bet pays off have nothing to do with each other.
The Big Fund piece operates on a five-to-ten-year horizon — compute sovereignty, the domestic chip ecosystem — and doesn’t need quarterly returns. The Tencent piece can move at the same commercial pace as any other model company. Neither waits for the other.
The discussion around DeepSeek’s commercialization problem tends to overlook a few things. Its API pricing has already undercut the industry floor: V4-Flash cache-hit input is listed at less than 0.02 RMB(0.003 USD) per million tokens, roughly 1% of GPT-5.5’s cost. That price looks like it should be impossible to sustain, but DeepSeek has compressed its inference costs far enough that even at this level, the API business doesn’t lose money.
And in a16z’s latest Top 100 Gen AI Consumer Apps report, DeepSeek still spans both Chinese and American users.Web traffic split 33.5% China, 6.6% US, with a mobile user base that hasn’t cratered the way outside observers predicted, even as model updates have slowed. Every monetization path available to other model companies is available to DeepSeek, at lower marginal cost.
But this split between commercial and strategic isn’t new. It predates the fundraising. It was the original design.
High-Flyer makes money. DeepSeek does research. That was the structure from day one. High-Flyer never needed DeepSeek to generate commercial returns. DeepSeek never had to justify ARR (annual recurring revenue) growth to its parent. The closest analogy might be a proprietary trading firm that invests a portion of its profits in basic research — where the research is part of the firm’s reason to exist, not a project waiting to be monetized.
But that arrangement rests entirely on one person holding both steering wheels at the same time. It’s personal, not institutional. Once outside capital arrives, the risk of losing that balance becomes real for the first time.
What this round is doing, then, is writing that informal arrangement into equity structure. The High-Flyer / DeepSeek relationship gets replicated at a larger scale: High-Flyer and DeepSeek are the first layer of separation; commercial operations and infrastructure research inside DeepSeek are the second. Both layers share the same logic and the same designer.
The third check, Wenfeng Liang’s own, makes this most explicit.
Reports put his personal contribution at up to 20 billion RMB ($2.8 billion). For reference, Moonshot AI — the Chinese startup behind Kimi, just closed a round totaling roughly $2 billion. Wenfeng Liang is personally writing a check larger than what most Chinese AI companies raise in an entire round.
The outside 60% isn’t coming in to keep DeepSeek alive or fund the next training run. High-Flyer’s capital is sufficient. So is Wenfeng Liang’s own. Every external investor is entering because DeepSeek stopped being a model company, and they’d rather own a piece of whatever it’s becoming.
His personal stake locks in control over the whole structure. No matter how much outside capital arrives, the core decisions don’t get diluted.
4.
The round has also been read, widely, as a talent retention play. That’s probably giving it too much credit, or at least, too specific a credit.
The departures were real. Bingxuan Wang left for Tencent. Daya Guo joined ByteDance Seed. Fuli Luo was poached by Xiaomi at a reported eight-figure salary (RMB). Chong Ruan joined DeepRoute.ai (a self-driving startup). Together they span base model, inference, OCR, and multimodal — four of DeepSeek’s core technical lines.
But look at the actual numbers. According to detailed data compiled by Caijing (a leading Chinese financial magazine), of the 15 researchers who appear most frequently across 27 DeepSeek papers, only 2 have left. Of the 86 people listed on the original DeepSeek-LLM release, 71 still appear in the V4 author list. Out of a team of roughly 300, about 10 people have departed: a 3.3% attrition rate. For anyone tracking AI talent right now, that number is remarkably low.
Equity pricing does have real retention value, but here it’s a side effect, not the point. DeepSeek’s pattern is consistent: it doesn’t act to solve the immediate problem. It acts to do the right long-term thing, and the immediate problems tend to sort themselves out. They didn’t raise to keep people. Keeping people is just one of the things that gets resolved along the way.
5.
To really understand the valuation, you have to get clear on who DeepSeek is actually competing against.
Nvidia sets the rules of compute in the AI era. Every model company, Chinese or American, operates within the boundaries Nvidia defines. DeepSeek is trying to redraw those boundaries. When V4 was validated on Huawei’s Ascend and published in an official technical document, that ambition was confirmed publicly for the first time.
The $50 billion valuation isn’t pricing model quality. It isn’t pricing annual revenue or user count. For context: Moonshot AI just closed a round valuing the company at roughly $20 billion, up from $4.3 billion six months earlier, with real ARR behind it. Anthropic’s last round pegged it at $380 billion, with annualized revenue reportedly crossing $30 billion in early 2026. DeepSeek has essentially zero commercial revenue, and its valuation sits well above twice Moonshot’s. The largest single contributor to this round is a semiconductor state fund, not a growth investor betting on future subscription income.
The investor mix tells you exactly what’s being priced. This is an infrastructure premium, not a commercialization bet.
If DeepSeek actually builds what it’s trying to build — a frontier AI compute stack that doesn’t depend on Nvidia, an open-source foundation for Chinese AI infrastructure — it has no obvious comparison. You’d need something between Stargate-scale OpenAI and Nvidia itself: a company making frontier models while simultaneously trying to solve the compute infrastructure problem from first principles. No other company in the world is doing both.
By that measure, $50 billion is a record. It’s also a discount.
This round is pricing something that hasn’t happened yet. Whether it does depends on whether Wenfeng Liang can finish what he started building on day one.
The full arc, from High-Flyer to DeepSeek, describes someone who follows through on long-term conviction with unusual consistency. Do the right long-term thing, and most of the near-term problems tend to work themselves out.
This round may be the first time the outside world has seen what that long-term thing actually looks like.