Following restrictions on exporting NVIDIA chips to China, DeepSeek has reportedly finalized an order of at least 160,000 Huawei AI accelerators for a new data center in Inner Mongolia — a gigawatt-scale project.
According to a Bloomberg report, the Chinese AI lab DeepSeek is finalizing an order of at least 160,000 Huawei Ascend 950DT AI accelerators for a new data center in Inner Mongolia. The power requirement is estimated to equal the consumption of about 750,000 homes, placing it among gigawatt-scale data centers.
A notable technical point is that DeepSeek, at least for now, does not plan to use these chips for training its models and still depends on NVIDIA chips for that sensitive stage; the Huawei chips are intended mainly for large-scale inference.
According to the report, Ascend 950DT deliveries begin in the fourth quarter of 2026, but because of Huawei's production capacity limits — especially a shortage of advanced memory — fulfilling the whole order could take more than a year. Partial commissioning of the data center is expected in late 2027 or early 2028.
The news reflects a larger trend: with US export restrictions continuing, the AI hardware ecosystem is gradually splitting into two fairly separate supply chains — one built on NVIDIA, the other on Huawei. For IT leaders planning long-term infrastructure, it is a reminder that dependence on a single hardware supplier is no longer just a technical risk but also a geopolitical one.
This note is a summary and rewrite based on the report below; see the original source for the full report.
BloombergThe 2026 report by the Cambridge Centre for Alternative Finance (CCAF), with BIS, IMF, WEF and others, shows 81% of financial firms use AI and software engineering is the most mature use — yet concern over lost human oversight and weak explainability is serious.
According to a SiliconANGLE report on September 18, 2026, a week that saw multi-billion-dollar AI infrastructure investments alongside safety tools to rein in rogue agents — the AI-agent industry's paradox.
Research by Medrano, Verma and Chhabra shows that in a production-like RAG system, RAG Fusion raises recall, but after re-ranking and context truncation much of the gain disappears and precision even dips slightly.
