DeepSeek reportedly seeks $3B and pitches V4 tuned for Huawei Ascend — a direct challenge to Nvidia

What happened

DeepSeek (幻方量化旗下 DeepSeek), the secretive AI arm behind last year’s R1 breakthrough, has reportedly broken its “never finance” stance and is seeking at least $300 million at a valuation of no less than $10 billion. It has been reported that the raise would bankroll more compute and higher pay to stem researcher departures after a year-and-a-half lull since R1. Domestic and some U.S. investors are said to be circling, but it has been reported that some American VCs remain cautious because of DeepSeek’s China-linked identity.

Technical claims and release plans

DeepSeek’s V4 is billed as a trillion‑parameter, Mixture‑of‑Experts model that activates roughly 37 billion parameters per token to keep inference costs similar to its V3 — efficiency‑first engineering, reportedly. V4 is said to introduce an “Engram” conditional memory for constant‑time retrieval over ultra‑long contexts (claims of 97% recall at 1 million tokens have been reported), and to be DeepSeek’s first native multimodal generator supporting text, image and video. Internal benchmarks reportedly show strong code performance (SWE‑bench >80%, HumanEval ~90%), and the company plans an Apache 2.0 release of weights consistent with its open‑source posture.

Hardware strategy and the geopolitical angle

Crucially, DeepSeek has reportedly re‑written core runtime code to migrate from Nvidia’s CUDA ecosystem to Huawei Ascend (昇腾) and the CANN stack, granted early access to domestic chip vendors while withholding early optimization access to Nvidia and AMD. If V4 runs competitively on Huawei (华为) hardware, it would mark a rare case of a cutting‑edge model not dependent on Nvidia — a shift with clear geopolitical ramifications as U.S. export controls and chip trade policy tighten. It has been reported that DeepSeek even posted on‑site server roles in Inner Mongolia, signalling a push to scale domestic compute.

Stakes and next steps

The timing is notable: Stanford’s AI Index reportedly shows the performance gap between U.S. and China’s top models narrowing to a few percentage points, and OpenAI’s recent massive fundraising underscores how capital‑intensive frontier AI has become. Will optimization for Chinese silicon meaningfully erode Nvidia’s moat? DeepSeek’s V4 and its financing plans are a high‑stakes experiment in that question — one the industry will be watching closely.