How 5 Billion RMB Funding Powers DeepSeek’s Pure Research Model Without External Financing

Fantasia Quant reported more than 5 billion RMB in revenue for 2025, derived from a 56.6% average fund return and a fee structure of 1% management fee plus 20% performance share. This cash pool can finance the training of 125 V3 models or 2,380 R1 models, and its sustainability improves as DeepSeek’s training efficiency rises.

Unlike many large‑model companies that chase commercialisation, DeepSeek operates a purely research‑focused model: it accepts no external financing, is not owned by any major tech firm, and concentrates on core technology development. Its R1 model is openly released, commercial efforts are limited to an API service, and client‑side maintenance receives little attention. In 2025 DeepSeek produced notable research outputs such as OCR advances and the V3.2 model, added over 60 pages of material to the R1 paper, and recently open‑sourced a memory module.

The funding stability stems from Fantasia Quant’s business, which is entirely separate from DeepSeek’s AI work. The quant fund’s profits are reinvested internally, freeing DeepSeek from equity‑driven pressures and short‑term ROI concerns. Consequently, DeepSeek can allocate resources fully to foundational training rather than high‑throughput inference applications.

Team stability is another strength: all 18 core contributors to the R1 paper remain, and only five of more than 100 authors have left, with one key member returning. Robust financial backing enables competitive salaries and top‑tier resources, allowing researchers to focus wholly on AGI objectives.

DeepSeek’s technical releases also unintentionally boost the domestic chip sector. After the V3.2 launch, Cambricon quickly completed framework adaptation, and its stock opened nearly 5% higher the next day, illustrating how DeepSeek’s roadmap can drive hardware ecosystem activity.