U.S. Moves to Block Chinese AI Firms from Circumventing Nvidia Chip Restrictions

nashnova research
今天发布阅读约 11 分钟

New U.S. Commerce Department rules close the loophole that let Chinese AI labs remotely tap Nvidia's top chips through Southeast Asian data centers. DeepSeek and Moonshot AI have already throttled services — whether Huawei's domestic chips can fill the gap is now the pivotal variable for China's AI competitiveness.

01

How was the backdoor shut?

Chinese AI labs including DeepSeek and Moonshot AI had been training models by remotely accessing restricted Nvidia chips housed in data centers in Malaysia and Indonesia.
A key target of the new rules is Megaspeed, which built Southeast Asian data centers and arranged for billions of dollars' worth of top-tier Nvidia chips to reach China. This means → Washington is no longer just controlling who *buys* chips; it is controlling who gets to *borrow* them remotely.
The exposure of a personal relationship between Megaspeed executive Alice Huang and former Nvidia CEO Jensen Huang has triggered an SEC investigation, compounding the compliance risks around this channel.
02

Are knock-on effects already visible?

Aolani Cloud, a Singapore-based cloud provider, had been sourcing Nvidia chips to supply compute to ByteDance. Banks involved in its financing are now reassessing the deal. This means → lenders are proactively screening "who is the end user," turning the financing chain itself into an enforcement lever.
Aolani says it is working to ensure compliance; China's Ministry of Commerce has labeled the rules "an abuse of export controls."
In plain terms = it is not just that chips are harder to buy — the money to *rent* compute is harder to borrow, too.
03

How hard is the compute squeeze hitting China's AI sector?

DeepSeek and Moonshot AI have capped user access to their chatbots and slowed model research — a direct signal that the supply-side shortage is reaching the product layer.
SemiAnalysis research director Daniel Nishball says: "The new regulations will shut down remote access for most Chinese labs."
Industry sources note that Chinese firms are trying to obtain compute through locally registered entities in the U.S. and Europe with no obvious Chinese ties, but such arrangements are high-risk and hard to scale to the cluster sizes needed for frontier model training.
04

Can smuggled chips make up the difference?

China has reportedly built several large domestic compute clusters using smuggled hardware, including Nvidia Blackwell chips.
But industry sources say this is far from sufficient for AI training demands. This means → smuggling can patch small gaps, but it cannot sustain the massive clusters frontier model training requires.
05

Can Huawei fill the gap?

Leveraging networking expertise from its telecom business, Huawei can now integrate hundreds of thousands of AI chips into clusters and claims performance close to America's top chips.
Huawei plans to launch two new AI chips next year and follow-on products in 2028–2029, though their compute power will still trail Nvidia's latest generation.
DeepSeek is training its next flagship model, V5, on Huawei Ascend chips. This reflects a shift: domestic substitution has moved from "backup option" to "live proving ground."
06

Why is V5 the make-or-break test?

If V5 is successfully trained primarily on Huawei hardware, it will demonstrate that Chinese domestic chips can support large-scale frontier AI training.
In plain terms = V5 is an open exam — whether Huawei chips can independently power a top-tier model will be answered in this training cycle.
This is the pivotal validation point for whether China's AI sector can stay competitive as the Nvidia channel narrows.

市场有风险,内容仅供研究参考,不构成投资建议。