Company Overview

DeepSeek (深度求索) is a Chinese AI research lab and product company headquartered in Hangzhou, founded in 2023 as a spin-off from High-Flyer Capital Management, one of China’s leading quantitative hedge funds. In January 2025, it released DeepSeek-R1 — a reasoning model that matched OpenAI o1 on key benchmarks at a fraction of the training cost, triggering a global reassessment of AI development economics.

Key Facts

• Founded: 2023, Hangzhou, China
• Parent: High-Flyer Capital Management (幻方科技)
• Employees: ~200 (lean by design)
• Funding: Backed by High-Flyer; no external VC rounds disclosed
• Key models: DeepSeek-V3, DeepSeek-R1, DeepSeek-Coder
• Open source: Yes — weights available on Hugging Face

Why DeepSeek Matters

DeepSeek upended two assumptions the Western AI industry had treated as settled facts: that frontier AI requires billions in compute spending, and that only US labs (or those with access to Nvidia H100 clusters) could build competitive reasoning models.

DeepSeek-R1 was reportedly trained on a cluster of approximately 2,000 Nvidia H800 GPUs — a chip US export controls had already downgraded for Chinese buyers. The training cost was estimated at under $6 million. OpenAI’s comparable models cost orders of magnitude more. The efficiency gap came from architectural innovations: Mixture-of-Experts (MoE) design, Multi-head Latent Attention (MLA), and aggressive quantization techniques.

When DeepSeek released R1 as open source in January 2025, Nvidia lost $600 billion in market cap in a single trading session — the largest single-day loss for any company in stock market history. Markets were pricing in the possibility that the GPU arms race might not be the only path to frontier AI.

What DeepSeek Actually Builds

DeepSeek runs a consumer-facing chat app (deepseek.com) and an API service. Unlike OpenAI or Anthropic, it operates with a skeleton crew — reportedly under 200 people total, including the research team. The company has consistently chosen to open-source its models rather than keep them proprietary, which has accelerated global adoption and sparked a wave of fine-tunes and integrations.

Its model family as of mid-2025:

• DeepSeek-V3 — general-purpose frontier model, competitive with GPT-4o
• DeepSeek-R1 — chain-of-thought reasoning model, trained with reinforcement learning; open-source weights
• DeepSeek-Coder-V2 — coding-specialized model, benchmarks near Claude 3.5 Sonnet on code tasks
• DeepSeek-VL2 — multimodal vision-language model

European Relevance

DeepSeek’s open-source release changed the European AI landscape in a concrete way: European companies and developers can now run a frontier-class reasoning model on-premise, without sending data to US cloud providers. This matters particularly for sectors where data sovereignty is non-negotiable — healthcare, finance, defense, government.

Several European enterprises have already begun deploying self-hosted DeepSeek-R1 instances. The EU AI Office is tracking DeepSeek under the AI Act’s General-Purpose AI provisions, given its systemic risk threshold. How Brussels regulates open-weights frontier models — particularly from non-EU labs — is one of the most consequential open questions in European AI policy.

Controversies and Risks

DeepSeek’s rapid rise has not been without scrutiny. Italian and Irish data protection authorities launched investigations into DeepSeek’s data practices in early 2025, with Italy briefly blocking the service. The core concern: DeepSeek’s privacy policy states that user data may be stored on servers in China, subject to Chinese law. For European users, this creates a potential conflict with GDPR’s data transfer requirements.

DeepSeek’s models also apply censorship filters consistent with Chinese regulatory requirements — topics like Taiwan’s political status, Tiananmen, and Xinjiang receive sanitized or deflected responses. This is expected behavior for a Chinese company, but it limits the models‘ utility for unconstrained analysis.

Bottom Line

DeepSeek is the most important new entrant in global AI since the GPT-3 era. It demonstrated that the frontier is not the exclusive territory of US hyperscalers, and it forced a global repricing of what AI development actually costs. For European decision-makers, the implications run in two directions: opportunity (cheap, self-hostable frontier models) and risk (data sovereignty, censorship, regulatory uncertainty). Both are real. Neither should be dismissed.

---

Inside China AI covers China’s AI and technology landscape for European professionals. Subscribe to China AI Weekly for weekly deep dives.