DeepSeek (深度求索) was founded in Hangzhou in 2023. Its founder, Liang Wenfeng, is also a co-founder of the quant fund High-Flyer (幻方科技). In early 2025, it open-sourced the reasoning model R1, and its app hit number one on the US App Store free chart. The whole industry started re-running the math on how much it really costs to train a frontier model.
One line to remember it by: the Chinese company that used quant fund money and engineering efficiency to make near-frontier AI open-source and cheap. Since 2026, it has moved from using only its own money to raising outside funding and preparing to go public, and it now stands on the front line of the US-China AI rivalry.
Key Data Snapshot
| Item | Data |
|---|---|
| Founded | 2023, Hangzhou, China |
| Founder / CEO | Liang Wenfeng (梁文鋒) |
| Listing status | Not public (as of October 2026) |
| First funding round | Closed June 2026, raising more than $7.4 billion at a post-money valuation above $50 billion (according to The Information) |
| Second funding round | In talks at a valuation of about RMB 500 billion (about $75 billion); not closed as of early October 2026 (according to Bloomberg and CNBC) |
| Main models | Flagship V4-Pro, V4.1-Flash; reasoning model R1 (no R2 yet) |
| App monthly active users | About 130 million, down about 20% year on year (June 2026, QuestMobile) |
| Competitors | Global: OpenAI, Anthropic, Google; China: Tongyi Qianwen, Doubao, Kimi, GLM |
A note on the numbers: DeepSeek has no public audited financials. Valuation and user numbers mostly come from media reports or third-party data, so focus on the order of magnitude and the trend. The “545% theoretical profit margin” it published assumes all-day usage billed at list price. It is a theoretical figure, not actual gross margin.
A Quick Tour of Seven Dimensions
① Technology and product: Two lines: reasoning models (R series) and general-purpose models (V series). Its calling card is engineering efficiency. The MoE (Mixture of Experts) architecture turns on only a small part of the model’s parameters each time, reaching near-frontier results with less compute. Its main model weights are released under the MIT license, so anyone can download and run them. V4.1-Flash, launched in September 2026, is the smallest model in the new architecture and can read images natively.
② Customers and positioning: A free app and website serve everyday users, while a low-cost API and open weights serve developers and businesses. Its position is clear: “the cheapest usable frontier model.”
③ Ecosystem and partnerships: DeepSeek doesn’t rely on distribution deals. It lets open weights grow the ecosystem on their own. Its models are on Hugging Face and GitHub, and major inference frameworks and many kinds of chips already support them.
④ Valuation and finances: Before 2026, almost all its money came from parent High-Flyer, which used quant trading profits to buy GPUs for research. Its first external round reportedly included Tencent, CATL, China’s national AI industry investment fund, and Alibaba, and Liang Wenfeng put in money himself. On listing, according to Reuters, it has hired CITIC Securities to prepare a listing on Shanghai’s STAR Market, with size and timing not set; Bloomberg reports the target is early 2027.
⑤ Regulation and risk: Inside China, it has completed generative AI service registration. Overseas, Italy’s data protection authority blocked it from accessing Italian user data in early 2025. In September 2026, a joint advisory from the US NSA, FBI, and CISA named DeepSeek and five other Chinese AI companies, accusing them of large-scale “distillation” of US models (asking huge numbers of questions and training their own models on the answers). That same month, according to The Information, China’s Cyberspace Administration opened an investigation into DeepSeek and Moonshot over user data that may have flowed to Anthropic; no penalty has been announced.
⑥ Geopolitics and supply chain: Advanced chips are the biggest outside variable. Early training ran mainly on the H800, a cut-down chip Nvidia made to meet export controls. In September 2026, DeepSeek announced a partnership with Huawei, open-sourcing programming tools for Ascend chips, and is pushing a “supernode” setup built from 128 Ascend 950 chips. According to Reuters in June 2026, the US had approved but held off on adding it to the export-control Entity List.
⑦ Leadership and governance: Liang Wenfeng runs both High-Flyer and DeepSeek, and decision-making is highly centralized.
Key Milestones
| Time | Milestone |
|---|---|
| 2021 | Parent High-Flyer builds a cluster of about 10,000 A100 GPUs |
| 2023 | Liang Wenfeng founds DeepSeek, spun out of High-Flyer’s AGI research team |
| 2024-04 | Completes China’s generative AI service registration |
| 2024-12 | Releases V3 (the model that made low-cost training famous) |
| 2025-01 | Open-sources reasoning model R1; app hits number one on the US App Store free chart; Italy blocks its access to user data |
| 2026-04 | Releases V4 series preview (V4-Pro / V4-Flash) |
| 2026-06 | Closes first external funding round (reported) |
| 2026-08 | V4-Pro full release goes live |
| 2026-09 | Releases V4.1-Flash; Ascend tools partnership with Huawei; reportedly hires CITIC Securities to prepare a listing |
| 2026-10 | According to Bloomberg, close to raising at least RMB 80 billion (about $11.9 billion) in the second round |
Further Reading
- US rivals playing a different game: OpenAI, Anthropic
- Other Chinese models: Kimi, Tongyi Qianwen (Qwen)
- The full hardware chain: The AI Hardware Supply Chain, End to End