
DeepSeek builds open-source large language models and reasoning systems in Hangzhou, China.
DeepSeek was founded in 2023 by Liang Wenfeng as a spin-off from High-Flyer Quant, a quantitative trading hedge fund. The lab has operated without external venture capital until 2026, funded instead by High-Flyer trading profits.
The company employs approximately 200 people, a lean workforce compared to OpenAI roughly 3,500 staff. DeepSeek research culture emphasizes reinforcement learning and architectural efficiency over brute-force scaling, attracting predominantly recent graduates who have produced over 68 research papers on arXiv.
DeepSeek develops open-source large language models using a Mixture-of-Experts architecture. Its flagship V4-Pro model features 1.6 trillion total parameters with 49 billion active per forward pass, supporting a 1 million token context window.
The company offers both a consumer chat application and an API platform, with models released under the MIT license. DeepSeek-V4-Flash provides a lighter 284 billion parameter variant optimized for cost-efficient inference at $0.14 per million input tokens.
DeepSeek January 2025 R1 release triggered a global tech selloff exceeding $1 trillion in market capitalization, primarily impacting NVIDIA and related semiconductor equities. The event demonstrated that efficient training methodologies can challenge the hardware-centric AI investment thesis.
The company is now reportedly in talks for a first external funding round at a $45 billion valuation, led by China state-backed semiconductor investment fund. If completed, this would mark one of the largest private capital infusions into a Chinese AI company and signal intensified state support for domestic AI infrastructure.
DeepSeek training efficiency breakthroughs enable frontier-class model performance at a fraction of the computational cost incurred by Western counterparts. The lab V4-Pro ranks competitively on coding and reasoning benchmarks while undercutting GPT-5.5 pricing by roughly 35 times on the Flash tier.
The company open-weight strategy removes deployment barriers for enterprises requiring on-premises or customized inference. Optimization for Huawei Ascend chips alongside NVIDIA hardware reduces supply-chain concentration risk relative to labs dependent solely on US semiconductors.
DeepSeek operates under US export controls limiting access to the highest-end NVIDIA training accelerators, which constrains its ability to scale dense training runs to match labs with full H100/Blackwell allocations. Domestic Chinese chip alternatives like Huawei Ascend cover inference well but introduce supply and tooling concentration risk on training.
The company lacks the enterprise sales motion, paid-tier monetization base, and global compliance footprint of OpenAI, Anthropic, and Google, leaving it dependent on open-weight distribution and the China API market. Reliance on High-Flyer trading profits historically and on a single 2026 external round limits funding diversification compared to Western frontier labs.
DeepSeek prices API access on a per-token basis through api-docs.deepseek.com, with discounted cache reads and off-peak pricing windows that target high-throughput batch workloads. The V4-Flash tier is positioned roughly 35 times cheaper than peer Western reasoning APIs, while V4-Pro offers a premium tier for long-context and frontier reasoning.
The consumer chat application at chat.deepseek.com remains free at public access, monetized indirectly through brand and developer pull-through to the paid API. Open-weight MIT-license releases let enterprises self-host on NVIDIA or Huawei Ascend hardware, removing API spend for on-premises deployments.