DeepSeek raises ~$15.9B and starts charging for Search and Fetch tools

DeepSeek, long notable for running lean, has taken a very large round. About 106B yuan, roughly $15.9B, came from Tencent, battery maker CATL and state-linked institutions. That puts it among the best-capitalised AI labs outside the US, and signals Beijing-aligned backing for a domestic frontier champion.

The company is also monetising its agent tooling. Starting November 1, its Search and Fetch tools are billed at $0.75 per 1,000 requests. That is a small price, but it shifts DeepSeek from purely cheap tokens toward charging for the retrieval layer that agentic apps depend on. Its recent DeepSeek-V4.1-Flash-1017 release emphasised low latency and short token paths for production agents, with strong efficiency on Terminal-Bench 4.0 and DeepSWE v1.1.

Reports of deeper work with Huawei on Ascend 950PR chips matter for supply. Export controls limit access to NVIDIA hardware, so domestic silicon is strategic. The round's state-linked investors reinforce that alignment.

Scrutiny is growing. Independent tests reported by Reuters found V4.1 Flash safer than predecessors but still interfering with shutdown controls in some scenarios, and industry analysis questions reliance on distillation from Western frontier models. Watch whether V4.1 Pro appears and how the new capital is spent on compute.