DeepSeek: A low-cost, open-source LLM dark horse hatched by quantitative private equity firm High-Flyer with GPU reserves

Origin

Born in Zhanjiang, Guangdong in 1985, Liang Wenfeng earned a master's degree in Information and Electronic Engineering from Zhejiang University. In 2015, he founded the quantitative private equity firm High-Flyer Quant, starting out with high-frequency trading powered by machine learning. To enhance its quantitative strategies, High-Flyer began hoarding tens of thousands of NVIDIA GPUs (such as A100s) starting in 2019, eventually becoming known at its peak as one of the few domestic institutions with a 10,000-card cluster. Amid the LLM wave in 2023, Liang Wenfeng carved out a portion of the GPUs, talent, and capital to establish DeepSeek, positioning it for general artificial intelligence basic research with a 'three no's' starting principle: no external financing and no rush to commercialize.

Milestones

Turning Points

• In 2021, enduring mockery, High-Flyer hoarded about 10,000 A100s, accidentally becoming DeepSeek's ultimate trump card to bypass chip controls later.
• In January 2025, R1 was open-sourced and topped the US AppStore, erasing approximately $589 billion from NVIDIA's market cap in a single day.
• In 2026, shifting from zero financing to accepting roughly 50 billion RMB in external funding and floating IPO rumors signaled the transition of a pure research lab model yielding to capitalization.

Failures & Pitfalls

• Early on, High-Flyer was mocked in the industry for private equity buying GPUs and neglecting its proper business, with the Glowworm supercomputer investing around 1 billion RMB with no commercial return in sight for years.
• In 2025, to blood-feed R&D, Liang Wenfeng clawed back High-Flyer profits, causing multiple quantitative products under its umbrella to post negative annual returns and damaging the core business's profit-generating capacity.
• The price war initiated by V2 drove API prices down to 1 RMB per million tokens; after the industry collectively cut prices, DeepSeek itself had virtually no profit-generating revenue.
• After skyrocketing in popularity, model services were frequently overwhelmed and API top-ups were temporarily suspended, exposing that inference compute reserves severely lagged behind user growth.

Core Data

• Base model training cost：Approximately $5.576 million (public data basis, independent verification unverified)
• Training GPU configuration：2,048 H800s for about two months (public data basis, independent verification unverified)
• Glowworm-2 GPU count：Approximately 10,000 units (public data basis, independent verification unverified)
• 2026 first-round financing scale：Approximately 50 billion RMB (public data basis, independent verification unverified)
• Valuation discussion range：Approximately 400 billion to 480 billion RMB (public data basis, independent verification unverified)
• Model API pricing：1 RMB per million input tokens (public data basis, independent verification unverified)
• NVIDIA market cap vaporized on release day：Approximately $589 billion (public data basis, independent verification unverified)

Competitors / Peers

Domestically benchmarking against Alibaba Tongyi Qianwen, ByteDance Doubao, Moonshot AI's Kimi, and Zhipu AI: Tongyi follows an open-source family-bucket plus cloud monetization route; Doubao relies on ByteDance traffic for C-end applications; Kimi focuses on long-text and secured major investment from Alibaba; Zhipu pursues a government-and-enterprise ToG route and has rushed toward an IPO. Internationally benchmarking against OpenAI and Anthropic, both of which are sprinting ahead with valuations in the hundreds of billions and annual revenue targets in the tens of billions of dollars, DeepSeek has torn open a gap beneath their pricing systems with open-source and ultra-low costs, though its commercial revenue remains far smaller than all tier-one competitors.