This piece gives you a quick introduction to DeepSeek (深度求索). In early 2025, its app suddenly shot to number one on the US App Store free chart, overtaking ChatGPT, and all of Silicon Valley started re-running the math on what it actually costs to train a frontier model. But who the company is, how it stays alive — most people can’t really explain it.
DeepSeek was founded in 2023 in Hangzhou, backed by quant fund High-Flyer (幻方科技). Liang Wenfeng (梁文鋒) is the founder and the boss of both. What it’s most known for is hitting frontier-comparable performance with less compute, then releasing model weights under open licenses like MIT.
Its biggest difference from most rivals: it’s cheap and open. When OpenAI and Anthropic took the closed-source, premium-price route, DeepSeek put its model weights out in public and drove API prices to the lowest in its class. “The cheapest usable frontier model” became its own kind of moat.
One line to remember it by: the Chinese company that used quant fund profits and engineering efficiency to make frontier AI open-source and cheap.
There are also some open questions with no settled answer yet. The most-watched external variable is access to advanced chips — Liang Wenfeng himself has said the real constraint isn’t money, it’s the embargo on advanced chips. On top of that, the first external funding valuation is still in negotiation, there are no audited financials, and there are data and export-control controversies overseas. Worth keeping those in mind when reading about DeepSeek.
Key Data Snapshot
Let’s put the key numbers together first. DeepSeek is not public and has no complete open financials, so many of the figures below come from the company’s own public statements or media reports. We try to flag which are officially confirmed and which are outside estimates.
| Item | Data |
|---|---|
| Year founded | 2023 |
| Headquarters | Hangzhou, China |
| Founder / CEO | Liang Wenfeng (梁文鋒) |
| Parent / background | Grew out of quant fund High-Flyer (幻方科技), same founder; primary source of early compute and capital |
| Company type | Private; AI research and commercial API platform spun out of a quant fund’s AGI research team |
| Main products | Reasoning models (R series), general-purpose models (V series); free app/web, open weights, low-cost API |
| Latest valuation | Media reports put first-round valuation at roughly $20–50 billion (not yet confirmed closed, figures vary across outlets) |
| Open-source license | R1 and V3 series mostly MIT-licensed; V4 series officially described as open-source (full license terms to be confirmed) |
| Main compute | Nvidia H800 (primary for early training), H20; also pursuing Huawei Ascend, AMD as alternative paths |
| Main competitors | Global: OpenAI, Anthropic, Google, Meta; China: Tongyi Qianwen, Doubao, Kimi, GLM, and others |
Two things to keep in mind when reading any Chinese AI startup’s numbers. ① DeepSeek has published things like “theoretical profit margin” figures — where all-day usage is calculated at full list price — those are theoretical estimates, not actual financial results; when you see a number like that, first ask “is this an assumption or an actual booking?” ② Private companies have no audited financials, so valuations and revenue figures are mostly media reports or estimates. Getting the order of magnitude and trend right is more solid than chasing exact values.
A Quick Tour of Seven Dimensions
Getting to know an AI company works well through seven dimensions. We’ll break the key ones into standalone pieces later.
① Technology and product direction: Products run in two lines — reasoning models (the R series that made DeepSeek famous in early 2025) and general-purpose models (the V series) — and are now moving toward longer context windows and agent tool use. Its real calling card is engineering efficiency: heavy use of MoE (Mixture of Experts) sparse architecture, where only a small fraction of parameters activate per inference, combined with a range of training optimizations that let it hit frontier-comparable quality with less compute. Model weights are mostly released under open licenses like MIT (the latest generation is officially described as open-source; full license terms pending confirmation).
② Customer structure and market positioning: On one end, a free app targeting global consumer mindshare (which topped the US App Store free chart in early 2025). On the other, extremely low API pricing and open weights aimed at developers and enterprises. Its competitive position is clear: “the cheapest usable frontier model.” Head-to-head competitors globally are OpenAI, Anthropic, Google, and Meta; in China there’s a pack of local models — Tongyi Qianwen, Doubao, Kimi, GLM, and others.
③ Ecosystem and partnership strategy: DeepSeek doesn’t rely on distribution stacking. It bets on “open weights plus broad inference framework support” and lets the ecosystem grow on its own. Models are on HuggingFace, GitHub, and other platforms; major inference frameworks and hardware from AMD and Huawei Ascend have all seen community-built adapters. Early on, capital and compute came from almost a single source: parent company High-Flyer, which converted quant trading profits into GPU clusters and poured them into research.
④ Valuation and financial model: As of May 2026, it is in talks for its first external funding round, with media-reported valuations in the range of $20–50 billion, not yet confirmed closed, and different outlets give meaningfully different figures. Founder Liang Wenfeng holds a dominant position. Another number that gets cited a lot is a “545% theoretical profit margin” — which assumes all-day usage charged at list price — that’s a theoretical estimate, not actual revenue or gross margin. The company is not public and has no audited financials; its actual financial picture is still to be disclosed.
⑤ Commercialization risk and regulation: Inside China, DeepSeek has completed generative AI service registration — one of the few checkpoints that can actually be verified. Overseas there are complications: Italy’s data protection authority blocked its access to Italian user data in early 2025, and South Korea’s personal data authority raised cross-border transfer concerns (per media reports). Sustaining the deep promotional pricing long-term also raises questions about whether its inference costs and margins can hold up.
⑥ Geopolitics and supply chain: Access to advanced chips is DeepSeek’s most-watched external variable. Early training ran mainly on Nvidia’s H800, a downspec version released post-export-control, and it is now pushing Huawei Ascend, AMD, and other domestic or alternative paths (technical adapters exist; commercial scale is still being established). There is also a contested supply-chain allegation: a single anonymous US official claimed DeepSeek had military ties and was circumventing export controls; Nvidia denied this and said the H800 chips were lawfully acquired, and the claims were not independently corroborated. That’s a single-source claim denied by the supplier and not independently verified — worth noting both sides. For how the full hardware chain works, see The AI Hardware Supply Chain, End to End.
⑦ Leadership and governance: Founder Liang Wenfeng leads both High-Flyer and DeepSeek simultaneously, and decision-making is highly centralized — that’s the company’s signature and also a risk. It was spun out of High-Flyer’s AGI research team, and there’s been no public confirmation of governance structures like an independent board. Talent movement is worth watching long-term: some core researchers have already left and moved to other Chinese tech companies’ AI teams.
Key Milestones
Picking out the pivotal moments that brought DeepSeek to where it is today:
| Time | Milestone |
|---|---|
| 2021 | Parent company High-Flyer completes a cluster of roughly 10,000 A100 GPUs, laying the compute foundation |
| 2023 | Liang Wenfeng founds DeepSeek, spun out of High-Flyer’s AGI research team |
| 2024-04 | DeepSeek completes China’s generative AI service registration (a verifiable compliance checkpoint) |
| 2024-12 | Releases V3 (the model that put low-cost training on the map) |
| 2025-01 | Releases reasoning model R1 (under open-source license); app tops the US App Store free chart |
| 2025-01 | Italy’s data protection authority blocks it from accessing user data (first overseas regulatory event) |
| 2025 H2 | R1 and V3 continue iterating; capabilities and agent/tool use strengthened |
| 2026-04 | Releases V4 series (V4-Pro / V4-Flash, supporting million-token-scale context) |
| 2026-05 | First external funding talks become public (media-reported valuation up to ~$50 billion, not yet closed) |
Milestones will keep being updated, with figures based on the latest announcements. This table was last compiled in May 2026.
Further Reading and Upcoming Standalone Pieces
If you want to go deeper, Penchan will break the key dimensions into standalone pieces over time:
- How does DeepSeek get compute under export controls?
- How do you calculate DeepSeek’s valuation? Why can it stay cheap and still survive?
- From registration to the Italy controversy: DeepSeek’s full regulatory picture
- Liang Wenfeng and High-Flyer: how a quant fund grew an AI company with global attention
- Want to learn how to use DeepSeek and get started for free? Tutorial content is separate from this piece (link will be added when live)
- Compare with rivals playing a different game: OpenAI and Anthropic
- For the full hardware chain: The AI Hardware Supply Chain, End to End