DeepSeek Plans Fresh Funding Round After $7.4B Close

NVIDIA claims its inference software stack has reduced token costs by up to 5x on the DeepSeek V4 model within one month on its Blackwell platform. The company argues that as AI moves from pilots to production, software optimization across serving, acceleration, and infrastructure layers becomes critical to cost efficiency. Leading inference providers including Baseten, Cognition, Deep Infra, and Together AI is already deploying these tools to improve throughput and reduce latency on Blackwell GPUs.

DeepSeek has open-sourced DSpark, an MIT-licensed framework that accelerates large language model inference by up to 85% without altering model outputs. The system uses speculative decoding, where a smaller draft model predicts likely token sequences that a larger model then validates, reducing computational overhead. DeepSeek has released technical papers, model checkpoints, and training code via GitHub and Hugging Face, making the technique available to researchers and enterprises running open-weight models.

DeepSeek, a three-year-old Chinese AI lab that had never raised outside funding, completed a $7.4 billion Series A in mid-June, valuing the company at over $50 billion. The fundraising marks the largest first-time raise by a Chinese startup. According to three people familiar with CEO Liang Wenfeng's thinking, the decision to seek external capital was prompted by Anthropic's April release of Mythos, a model preview that Anthropic claimed could find and exploit software vulnerabilities.

Microsoft is exploring the integration of DeepSeek's V4 model as a cost-effective option for its Copilot Cowork AI assistant, according to reporting from Axios. The company is evaluating either a Microsoft-hosted version of DeepSeek V4 or another open-source alternative to reduce expenses associated with powering the assistant. This move reflects Microsoft's effort to balance capability with cost efficiency in its AI product offerings.