DeepSeek Pivots to Huawei Ascend Chips for Planned 8-Trillion-Parameter Model Amid China Data Probe

DeepSeek's chip pivot is a high-stakes bet on Chinese silicon independence. Liang Wenfeng told investors that using more domestic chips for training has become a major priority, with Huawei expected to begin delivering Ascend 950-series (including 950DT) training chips as early as Q4 2026. Reports emphasize DeepSeek is trying to prove Huawei chips can train — not merely run — frontier models. The company is currently training a 2-trillion-parameter model and plans an 8-trillion-parameter successor that would eclipse Alibaba's Qwen 3.8 Max (2.4T) and Moonshot's Kimi K3 (2.8T). The effort is backed by a funding round targeting a roughly $74 billion valuation.

The move away from Nvidia is both strategic and forced, as US export controls constrain access to top-tier GPUs. It also arrives as OpenAI's cheaper GPT-6 Sol and Luna charge directly into DeepSeek's low-cost home turf, pressuring its core value proposition.

Complicating matters, China's Cyberspace Administration opened an investigation into DeepSeek and Moonshot AI after an Anthropic threat report alleged the labs routed sensitive user data to Claude without authorization — DeepSeek reportedly routed 12.1 million exchanges in 14 days in July, allegedly including state and police data. DeepSeek was separately invited to brief the UN Security Council on AI risks. The convergence of a chip pivot, a mega-round, a regulatory probe, and a UN appearance makes DeepSeek one of the week's most geopolitically loaded AI stories, even as the technical question — whether Ascend silicon can sustain 8-trillion-parameter training — remains open.