Company story6 min read
How a stock trader came to own DeepSeek
Liang Wenfeng still controls DeepSeek. His hedge fund paid for it for three years, and the outside investors who bought in this June hold under 9%.
For three years DeepSeek had two shareholders.
Since June it has four.
DeepSeek's shareholder register at three dates, as a share of registered capital.
At registration1
- 99% Ningbo Cheng'en
- 1% Liang Wenfeng
After Liang's own capital increase2
- 66% Ningbo Cheng'en
- 34% Liang Wenfeng
After the first outside round3
- 60.19% Ningbo Cheng'en
- 31% Liang Wenfeng
- 8.52% Hangzhou Chengli
- 0.28% National AI Industry Investment Fund
Where these facts come from. DeepSeek publishes no accounts. Its shareholders come from China's business registry as Chinese outlets reported it, its models from its own papers and announcements, and what happened around it from the press. Every fact is footnoted to the source list at the end, and one that only a single outlet reported says “reportedly”.
Liang Wenfeng does. He holds 31% of DeepSeek in his own name, and a partnership called Ningbo Cheng'en holds 60.19%.3 Who the partnership's partners are is not public, but before this year's round Yicai put Liang's effective ownership at 84.3%.2
For its first three years DeepSeek sold nothing to anybody. It was paid for by High-Flyer, the quantitative hedge fund Liang also controls.4 In June 2026 it sold for the first time, and its outside investors, Tencent and a state AI fund among them, now hold 8.8% of it.5
The register as reported in July 202635; the value is our arithmetic on the filing6.
A trader who wanted machines to pick the stocks
Liang was born in 1985 in Zhanjiang, in southern China, and studied electronic information engineering at Zhejiang University. In 2015 he co-founded the firm that became High-Flyer, in Hangzhou.7 On 21 October 2016 it put on its first stock position chosen by a deep-learning model, and by the end of 2017, the firm says, almost all of its strategies ran on AI models.8
Models that trade need computers that train them, and High-Flyer built its own. Its first cluster, Fire-Flyer, cost nearly ¥200 million and held 1,100 accelerator cards; the second cost ¥1 billion.8 DeepSeek's own paper on that second machine counts 10,000 Nvidia A100s.9 They were bought in 2021, Forbes reported.10 In August 2022 the US government told Nvidia that selling the A100 to China now needed a licence.11 High-Flyer already had its ten thousand.
A lab with nothing to sell
In April 2023 High-Flyer said it would set up an independent organisation to research artificial general intelligence.12 DeepSeek was registered in Hangzhou on 17 July 2023, 99% owned by Ningbo Cheng'en and 1% by Liang.1
It hired young. “We are mostly fresh graduates from top universities,” Liang said in 2024, describing a lab with no hierarchy and no separate departments. ChinaTalk, which translated the interview, called it fully funded by High-Flyer with no plans to raise money.4
That left one thing to sell, an API, and DeepSeek priced it near cost. In March 2025 it published a day of its own serving costs: $87,072 of GPU time, costed at $2 an hour, against $562,027 of revenue had every token been billed at R1's price, adding that its real revenue was substantially lower.13
The models
The first, DeepSeek Coder, came out on 2 November 2023, half a year after the lab was announced.1 After it came a run of papers, each adding one idea the next model was built on.
| Model | What was new | Size |
|---|---|---|
| DeepSeek Coder2 Nov 2023 | Code models trained from scratch on 2T tokens, 87% of them code. 14 | 1B to 33B |
| DeepSeek LLM29 Nov 2023 | General models on 2T English and Chinese tokens, with the lab's own study of how they scale. 15 | 7B, 67B |
| DeepSeekMoE11 Jan 2024 | Many small experts plus shared ones, so each token uses a fraction of the model. 16 | 16B |
| DeepSeekMath5 Feb 2024 | GRPO, a cheaper way to train a model with reinforcement learning. 17 | 7B |
| DeepSeek-V26 May 2024 | Multi-head Latent Attention: 93.3% less memory per conversation than the 67B model. 18 | 236B / 21B |
| DeepSeek-V326 Dec 2024 | Training at eight-bit precision at scale, and predicting several tokens at once. 19 | 671B / 37B |
| DeepSeek-R120 Jan 2025 | Reasoning learned through reinforcement learning, released under the MIT licence. 20 | 671B / 37B |
| DeepSeek-V3.121 Aug 2025 | One model that either thinks first or answers directly. 21 | 671B / 37B |
| DeepSeek-V3.2-Exp29 Sep 2025 | DeepSeek Sparse Attention, which made long contexts cheaper to serve. 22 | 671B / 37B |
| DeepSeek-V424 Apr 2026 | A million tokens of context by default, in a Pro and a Flash. 23 | 1.6T / 49B |
| DeepSeek-V4.1 Flash10 Sep 2026 | An encoder-decoder design whose cache needs a quarter of the memory. 24 | 552B / 8B to 16B |
Total parameters, and where a model is a mixture of experts the parameters active for each token. Dates are DeepSeek's own, from its repositories, papers and announcements.
DeepSeek-V2 was the one that moved the market. Multi-head Latent Attention compresses what a model holds in memory about a conversation, and against its 67B model DeepSeek reported 93.3% less of that cache and 42.5% lower training cost.25 It priced V2 at one yuan per million input tokens.26 Within three weeks Alibaba, Baidu and others had cut their prices or made models free.27
“We just accidentally became a catfish.”
Liang Wenfeng on the price war, 2024 interview, translated by ChinaTalk4
DeepSeek-V3, on 26 December 2024, was trained largely at eight-bit precision on 2,048 Nvidia H800s. Its report put the final training run at $5.576 million of GPU time, priced at $2 an hour, and said the figure leaves out the research and experiments before it.19
DeepSeek-R1, on 20 January 2025, was a reasoning model. Its first version, R1-Zero, learned to reason through reinforcement learning alone, without the supervised training labs usually start with. It shipped under the MIT licence, free to use and build on,20 at $0.55 per million input tokens and $2.19 for output, against $15 and $60 for OpenAI's o1.28 A later version of its paper put R1's own training at $294,000, on top of V3.29
The V4 report says one part of its system was tested on Huawei's Ascend chips as well as Nvidia's.30 One model never came: no R2 appears anywhere in DeepSeek's change log.31
The week everyone noticed
On Sunday 26 January 2025 DeepSeek was the top free app in the US App Store.32 On Monday Nvidia lost $593 billion of market value, the largest one-day loss any company had taken on Wall Street.33
Governments followed. DeepSeek's privacy policy says it stores personal data in the People's Republic of China.34 Italy's privacy regulator restricted it that week,35 and Australia ordered it off every government device.36 So did accusations: OpenAI said it was reviewing signs that DeepSeek had distilled its models,37 and in February 2026 Anthropic said DeepSeek had run over 150,000 exchanges with Claude in what it called distillation attacks.38
Then it sold, a little
In February 2025 The Information reported that state funds had approached DeepSeek; a company source called it a rumour.39 In April 2026 Liang raised DeepSeek's registered capital by half with his own money, lifting his direct stake from 1% to 34%.2 On 15 June 2026 DeepSeek completed its first outside round, about ¥51 billion, from Tencent, CATL, JD, Monolith, IDG, Loyal Valley Capital and the state-backed National AI Industry Investment Fund.3
Liang reportedly put in about ¥20 billion of it himself, and the investors reportedly came in without votes and with a five-year lock-up.40 They hold their 8.52% through a partnership set up for the round, Hangzhou Chengli; the state fund holds its 0.28% directly.5 A listed company's filing put a price on the whole: ¥40 million bought a look-through stake of 0.0114%, which values DeepSeek at about ¥351 billion.6
It may not stay this small. A second round of at least ¥80 billion was reported in October,41 and in September Reuters reported that DeepSeek had hired CITIC Securities to prepare a listing on Shanghai's STAR Market.42
What changed
Two months after the first outside money, DeepSeek raised its API prices and began charging double at peak hours, announcing the change with the V4 Pro release.43 Yicai reported that it blamed recent outages on an imbalance between the supply of computing power and demand.44
Liang still controls the company, and its models still ship under the MIT licence. It is no longer a lab with nothing to sell. What it charges now, model by model and seller by seller, is in how DeepSeek went from price war to price rise.
Who owns the other big labs, and who holds their votes, is in who owns the AI companies.
Sources
- 1PressReport on DeepSeek's registration and first shareholders (in Chinese)The Paperread
- 2PressDeepSeek founder injects own funds to lift Chinese AI firm's registered capital by 50%Yicai Globalread
- 3PressReport on DeepSeek's first financing, citing Tianyancha (in Chinese)China Securities Journal, via Eastmoneyread
- 4InterviewDeepSeek CEO interview with China's 36Kr, translatedChinaTalkread
- 5PressReport on DeepSeek's shareholders after the round, citing Qichacha (in Chinese)Time Weekly, via Eastmoneyread
- 6FilingReport on Anhui Korrun's disclosed indirect stake in DeepSeek (in Chinese)Anhui Korrun, via 10jqkaread
- 7PressProfile of Liang Wenfeng (in Chinese)QQ Newsread
- 8PageHistory (in Chinese)High-Flyerread
- 9PaperFire-Flyer AI-HPC: a cost-effective software-hardware co-design for deep learningDeepSeek, arXivread
- 10PressHow much AI firm DeepSeek and its founder are worthForbes Australiaread
- 11FilingUS government license requirement for A100 and H100 exports to China, 26 August 2022Nvidia, Form 8-Kread
- 12PressReport on High-Flyer's new AGI research organisation (in Chinese)Yicairead
- 13RepoDeepSeek-V3/R1 inference system overview, Open Source Week day 6DeepSeek, GitHubread
- 14RepoDeepSeek CoderDeepSeek, GitHubread
- 15RepoDeepSeek LLM 7B and 67BDeepSeek, GitHub (repository opened 29 Nov 2023)read
- 16PaperDeepSeekMoE: towards ultimate expert specialization in mixture-of-experts language modelsDeepSeek, arXivread
- 17PaperDeepSeekMath: pushing the limits of mathematical reasoning in open language modelsDeepSeek, arXivread
- 18RepoDeepSeek-V2DeepSeek, GitHub ("2024.05.06: We released the DeepSeek-V2")read
- 19PaperDeepSeek-V3 technical reportDeepSeek, arXivread
- 20ReleaseDeepSeek-R1 ReleaseDeepSeekread
- 21ReleaseDeepSeek-V3.1 ReleaseDeepSeekread
- 22ReleaseIntroducing DeepSeek-V3.2-ExpDeepSeekread
- 23ReleaseDeepSeek V4 Preview ReleaseDeepSeekread
- 24ReleaseDeepSeek-V4.1-Flash: Smarter, Faster, More EfficientDeepSeekread
- 25PaperDeepSeek-V2: a strong, economical, and efficient mixture-of-experts language modelDeepSeek, arXivread
- 26ArchiveAPI pricing, archived 8 May 2024DeepSeekread
- 27PressChinese tech giants slash prices of language models used to power AI chatbotsReuters, via BusinessDayread
- 28ArchiveAPI pricing, archived 21 January 2025OpenAIread
- 29PaperDeepSeek-R1, version 2, with training costsDeepSeek, arXivread
- 30PaperDeepSeek-V4 technical reportDeepSeek, arXivread
- 31PageChange logDeepSeekread
- 32PressDeepSeek displaces ChatGPT as the App Store's top appTechCrunchread
- 33PressChina's DeepSeek sets off AI market routReutersread
- 34PagePrivacy policy, last updated 10 February 2026DeepSeekread
- 35OrderOrder limiting DeepSeek's processing of Italian users' dataGarante per la protezione dei dati personaliread
- 36OrderPSPF Direction 001-2025: DeepSeek products, applications and web servicesAustralian Governmentread
- 37PressOpenAI says DeepSeek may have inappropriately used its dataNBC Newsread
- 38PageDetecting and preventing distillation attacksAnthropicread
- 39PressDeepSeek reportedly denies US report that it seeks outside fundingYicai Globalread
- 40PressReport on the terms of DeepSeek's first round (reported, unnamed sources)36Krread
- 41PressDeepSeek to raise at least $12 billion in Tencent-backed fundingKhaleej Timesread
- 42PressExclusive: China's DeepSeek taps CITIC to prepare for IPOReuters, via Yahoo Financeread
- 43ReleaseDeepSeek-V4-Pro GA ReleaseDeepSeekread
- 44PressDeepSeek to raise API prices as demand for computing power growsYicai Globalread
More posts
Why a Google SERP API now returns five results, not ten
Google now unlocks each result link a little later than the one above it, the tenth about 24 seconds after the page. Five results arrive before the long waits: 2,273 of 2,500 live searches answered, 12 at a time, at a median of 4.19 seconds.
Claude API pricing: how Opus fell from $75 to $20
A million tokens of Opus output cost $75 in May 2025. Opus 5.5 launched at $20 in September. Every Claude model's price today, every change since, and what one coding session costs.
Who owns the AI companies, and who still holds the votes
Elon Musk owns under half of SpaceX, home of xAI, and casts 82% of its votes. Microsoft owns about a quarter of OpenAI and names none of its directors. How the money and the votes came apart at 13 AI labs.
How Beam 1.0 got past its tokens-per-second ceiling
Beam 1.0 runs at 348.8 tokens a second across its hosts and 537.3 on the fastest. One token a pass stops near 300. What sets that ceiling, and the ways past it.
How DeepSeek went from price war to price rise
DeepSeek's cheapest model costs $1.20 per million output tokens at peak, up from $0.28 in April. Every seller's price today, and every release's price and score.
