The open-weight frontier lab that trains state-of-the-art models at a fraction of the cost
HQ: Hangzhou, China | Founded: 2023 | Employees: ~300 (2025); ~550-600 by mid-2026 per press | Stage: Bootstrapped subsidiary of High-Flyer (no external VC) | Website: https://deepseek.com
DeepSeek (深度求索, literally deep pursuit) is an artificial-intelligence research company founded on July 17 2023 in Hangzhou by Liang Wenfeng, who also co-founded the quantitative hedge fund High-Flyer in 2015. It is financed exclusively by High-Flyer and its AI-compute subsidiary, with no external venture capital (an arrangement Liang actively keeps). The lab builds open-weight foundation models and runs its own app, API, and research line: DeepSeek-Coder, DeepSeek-Math, DeepSeek-VL and Janus-Pro multimodal models, the DeepSeek-V3 general assistant (Dec 26 2024), the DeepSeek-R1 reasoning model (Jan 20 2025), DeepResearch and DeepSearch tools, DeepSeek-OCR, API, and the 2026 R2 line. V3 and R1.open and R2 regulations. Its training fleet is High-Flyer's Vero supercomputer cluster, one of the largest private GPU parks in the world (~80K+ GPUs concentrated around Hangzhou). Western export-control restrictions removed DeepSeek from US app stores and cut US API access in early 2025; Italy's Garante ordered data-localization changes, South Korea and Taiwan restricted downloads, and the EU's AI Act chapter opened on GPAI transparency. Despite - or because of - the bans, DeepSeek weights became the most widely self-hosted open frontier models of the era, hosted on Hugging Face and mirrored across Chinese cloud marketplaces.
DeepSeek proved the frontier can be reached on commodity budgets and then gave the results away. Its core moat is its people and its dollar efficiency: a small, star-studded research team inside High-Flyer's sovereign-scale compute park (Vero, ~80K+ GPUs) that produced V3 (~$5.6M training cost), the R1 reasoning chain (MIT license, GRPO-style RL), and - post-cleanup - a still-cheaper R2 in 2026. The bull case: open weights as global distribution, an API priced at 1/40th of OpenAI-class models, and relentless algorithmic innovation only possible at this headcount and culture. The bear case: near-zero known revenue, total dependence on one funding source (High-Flyer) and on Chinese compute supply under US export controls, app-store and API bans across the West, and an open-weights model that rivals can copy as fast as DeepSeek ships it.
Browse: Home | Blog | About | All Startups A-Z | View full profile on StartupWiki