The Daily Pulse

Technology

Research Suggests Restricted Nvidia Chips Powered Chinese AI Models

Reported by Aditya Chauhan (Staff Writer) · Indian Express - Technology ·

Source: Indian Express
Worth knowing
TechnologyResearch Suggests Restricted Nvidia Chips Powered Chinese AI Models
Representative image · Wikimedia Commons

Technical Evidence

Advanced Nvidia Blackwell chips, barred from direct export to China, were used to train leading open-weight artificial intelligence models developed by Chinese firms including DeepSeek and Moonshot AI, according to a research study released on Thursday, November 12. Conducted by former OpenAI researcher Will DePue, who leads the nonprofit Unemployment Labs, the study analyzed weights from models released openly by DeepSeek, Moonshot AI, MiniMax, and Ant Group. Six models released since April 2026 contained numerical fingerprints suggesting they were trained on Blackwell hardware. The researchers recovered initial random values generated by an algorithm called Philox and confirmed matches on physical Blackwell chips for models such as DeepSeek-V4-Flash and Kimi K3.

Training Methods and Costs

The study found that most open-weight models used a mix of Hopper and Blackwell chips, with recent upgrades showing Blackwell usage. Models like Kimi K3 and DeepSeek-V4-Pro each required thousands of Blackwell GPUs running for two to three months. Researchers estimated the cost for these training clusters at roughly $200 million to $300 million, based on a price of about $70,000 per GPU. However, not all Chinese labs rely on Blackwell chips; models from Alibaba, Zhipu, Tencent, Xiaomi, and StepFun appeared Hopper-trained, while Meituan used AI ASIC superpods.

Leadership Changes and Policy

The findings emerge amid a spy scandal involving Nvidia and its former CEO Jensen Huang, who stepped down last week following reports of a relationship with Alice Huang, a former executive of Singapore-based Megaspeed whom US authorities suspect of spying for Beijing. US officials previously claimed Chinese firms were using restricted chips, and current export rules permit Nvidia to ship less powerful Hopper H200 and H20 chips to China with a 25 per cent sales cut to the US government, while Blackwell GPUs remain banned. Gaps in controls, such as remote access to data centres in countries like Malaysia where Nvidia can export chips, have facilitated access. Epoch AI reported that over 2,75,000 Nvidia AI chips were imported to Malaysia in 2024 and 2025, and semiconductor research firm SemiAnalysis estimates that more than 1.5 gigawatts of restricted hardware has reached China.

Context

Semiconductor export controls are designed to prevent foreign military competitors from gaining access to cutting-edge computing power that can accelerate artificial intelligence development. However, global supply chains and offshore data centre leasing create challenges for enforcing geographical restrictions on high-end hardware.

ExplainerWhy this matters

What Happened

A research report by Unemployment Labs analyzed the internal numerical weights of open-weight artificial intelligence models released by Chinese companies like DeepSeek and Moonshot AI, discovering specific mathematical patterns that indicate training on US-restricted Nvidia Blackwell graphics processing units.

Why It Matters

This situation highlights major enforcement gaps in international semiconductor trade policy. Even though direct shipments of high-end chips to China are banned by the United States government, loopholes involving remote cloud access, offshore data centers in countries like Malaysia, and indirect supply chains allow rival developers to utilize cutting-edge hardware for large-scale AI model training.

What Happens Next

Regulators and lawmakers are likely to face increased political pressure to tighten export controls, close offshore data center loopholes, and step up investigations into chip smuggling and compliance verification by major hardware manufacturers.

Read more at Indian Express - Technology →