Chinese artificial intelligence lab DeepSeek has reportedly raised more than 50 billion yuan, or $7.4 billion, in new funding.

The Information and the Wall Street Journal today cited sources as saying that the company is now worth over $50 billion. That reportedly makes it China’s most valuable AI startup.

DeepSeek founder and Chief Executive Officer Liang Wenfeng reportedly contributed $3 billion to the raise. Earlier this month, Reuters reported that Tencent Holdings Ltd was considering investing about $1.48 billion. Most of the participants in the round reportedly deposited the capital into a limited partnership managed by Liang.

DeepSeek, officially Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co. Ltd., is a subsidiary of a hedge fund called High-Flyer. The AI lab rose to prominence in January 2025 when it open-sourced an advanced reasoning model known as R1. The launch sparked a broad selloff in chip stocks that briefly reduced Nvidia Corp.’s market capitalization by 15% at the time.

R1 nearly matches the output quality of o1, a reasoning model that OpenAI Group PBC released in December 2024. Furthermore, DeepSeek’s model can perform inference using a fraction of the hardware. The model’s launch weighed on chip stocks because investors were concerned its hardware-efficiency may reduce the need for AI accelerators.

DeepSeek released a successor to R1 called DeepSeek-V4-Pro in April. The latter model includes 1.6 trillion parameters, more than twice as many as its predecessor. It features a mixture-of-experts architecture that only activates 284 billion parameters when answering prompts. DeepSeek trained the model on a dataset that comprised more than 32 trillion tokens.

Neural networks use a data structure called a KV cache to perform inference. V4-Pro can process prompts that contain 1 million tokens with a KV cache one-10th the size of DeepSeek-V3.2, an earlier DeepSeek model. The result is a significant reduction in memory usage, which lowers inference expenses.

The cost-efficiency of DeepSeek’s algorithms has caught the attention of Microsoft Corp. Axios reported today that the tech giant may integrate a customized DeepSeek model into its Cowork Copilot application, which launched into general availability today. The goal is to provide a lower-cost alternative to the OpenAI and Anthropic Group PBC algorithms that currently power the software.

Microsoft reportedly plans to use a fine-tuned version of DeepSeek V4 or another open-source model. Fine-tuning is a process through which developers tailor an AI to a specific workload, which increases its output quality and speed. Microsoft is expected to roll out the customized model in the coming weeks.

Image: Unsplash