DeepSeek

Headquarters

Hangzhou, China

Website

www.deepseek.com

Employees

200 (2025, approximate)

Founded

2023

Listing

private

Dependency

• Huawei: DeepSeek runs its own V4 model infrastructure on Huawei's Ascend hardware, a strategic departure from its earlier models which relied on NVIDIA's H800 GPUs.
• CUDA: DeepSeek releases V4's open weights in CUDA-compatible formats even as its own production infrastructure runs on Huawei's Ascend hardware, keeping the model usable by the much larger installed base of NVIDIA GPUs.

Development

• Large Language Models: DeepSeek develops the DeepSeek V/R family of large language models, notable for competitive performance at a fraction of the cost of proprietary frontier models.
• AI Accelerators: DeepSeek is reportedly developing its own in-house inference chip to reduce dependence on both NVIDIA and Huawei hardware, joining other Chinese AI labs like Zhipu in exploring custom silicon.

Competition / Other

• OpenAI: DeepSeek's open-weight models compete with OpenAI's GPT models, offering comparable capability at a fraction of the training and inference cost.
• Anthropic: DeepSeek's open-weight models compete with Anthropic's Claude models, offering comparable capability at a fraction of the training and inference cost.
• Baidu: DeepSeek competes with Baidu (Ernie) and other Chinese AI labs for domestic model adoption and enterprise API customers.
• Mistral AI: Mistral AI competes with DeepSeek and other cost-efficient open-weight model providers for developer and enterprise adoption.
• Tencent: Tencent's in-house Hunyuan AI model family competes with DeepSeek and other Chinese open-source model developers.