Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
InfrastructureInsight · On RadarTechCrunch→ original

Kimi K3 and DeepSeek V4 Weights Show Traces of Banned Nvidia Blackwell Chips

The open Chinese models Kimi K3 (Moonshot AI) and DeepSeek V4 were apparently trained on Nvidia Blackwell chips, which the US has banned from shipping to China, according to new analysis. Will DePue, a former OpenAI researcher, found in the models' weights, the files that hold what they have "learned," a trace of the chip configuration, like the mark a machine tool leaves on a part. But this is not proof of a violation: the chips could have been sitting in a data center in Malaysia. Companies renting GPUs abroad should check who owns the provider.

AI-processed from TechCrunch; edited by Hamidun News
Kimi K3 and DeepSeek V4 Weights Show Traces of Banned Nvidia Blackwell Chips
Source: TechCrunch. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceTechCrunch ↗
Editor-in-Chief
Epoch AIaround 2 million chips were smuggled into China in 2024–2025
Key metric #2Around 275,000 Nvidia accelerators were shipped to Malaysia over the same years
Key metric #3Nvidia CEO Jensen Huang stepped down from the company amid the Megaspeed scandal

The most prominent open models from China, Moonshot AI's Kimi K3 and the DeepSeek V4 family, were apparently trained on Nvidia Blackwell chips, according to a new study. These are the company's most powerful accelerators, and the US bans shipping them to China and Chinese firms. The problem can be framed this way: the country has closed its border to the most advanced hardware, yet the output of that hardware still ended up in the open. For anyone building products on open weights, tracking the GPU market, or renting compute overseas, this is an important story. Hardware prices, model availability, and regulatory risk all depend on it.

How to tell from the weights what a model was trained on

Will DePue previously worked as a researcher at OpenAI and now heads the nonprofit Unemployment Labs. He analyzed the weights of open models from DeepSeek, Moonshot AI, MiniMax, and Ant Group. The weights are a file of numbers that stores everything the model has learned. According to the researcher, they also retain a trace of the configuration of the chips the training ran on. It is like the mark a machine tool leaves on a part: from it, you can infer what equipment made the part.

According to DePue, the most popular models, the Kimi K3 and DeepSeek V4 releases, show traces pointing to Blackwell. The TechCrunch summary does not include the detailed methodology, so the reliability of the method cannot yet be assessed from the outside. But the conclusion matches public statements. Earlier this year, the White House technology adviser Michael Kratsios said, citing US intelligence, that Moonshot had used banned Nvidia chips.

What this proves and what it does not

The finding does not mean the models were trained only on Nvidia hardware. Nor does it mean the law was necessarily broken. Chinese companies often rent compute in data centers abroad, a popular way to get access to capacity. According to Epoch AI, around 275,000 Nvidia accelerators were shipped to Malaysia in 2024–2025, and an entire industrial sector has grown up around them there.

Critics argue that this is where the loophole lies. The restriction works by geography: chips cannot be obtained in mainland China, Hong Kong, or Macau, but they can be obtained in Malaysia. Sam Bresnick of Georgetown University's Center for Security and Emerging Technology told TechCrunch that this logic makes no sense if the goal is to preserve the US technological edge. Hardliners have long called for controlling not only who owns the chips and where they are installed, but also who is actually running computations on them.

The scale of smuggling is also large. Epoch AI estimates that around two million advanced chips reached China in 2024–2025 through schemes that concealed the end recipient.

The Nvidia scandal

The topic has become hot because of Megaspeed, which is linked to chip shipments to China. The company has long operated in this market through its subsidiary Speedmatrix, and US and Singapore authorities are looking into it. Bloomberg reported that a firm connected to Megaspeed executive Alice Huang took part in smuggling banned chips into a data center in Shanghai.

According to TechCrunch, Nvidia CEO Jensen Huang stepped down from the company this week after admitting to a romance with Alice Huang. The outlet notes that personal details draw the most public attention, but the decision was probably also shaped by Nvidia's possible liability for export-control violations. Nvidia and Huang had previously vouched for Megaspeed's data centers and said company staff had visited and inspected them.

The work of US agencies is under scrutiny as well. Reportedly, Commerce Secretary Howard Lutnick may lose his post: his departments failed to identify suspicious buyers and were too closely engaged with Nvidia representatives.

What comes next for the market

There is a flip side. The short seller Culper argued in May that Nvidia has a "China problem," but he was not betting on US crackdowns. He argued that China would respond by developing its own chips and that Nvidia's billion-dollar demand from the country would disappear.

Meanwhile, about 600,000 H200 chips, a less powerful model that the Trump administration allowed to be sold on condition of a workaround route and a 25% fee paid to the US, are sitting in warehouses in Hong Kong awaiting shipment to China. Observers have no single explanation for why they are not moving. The theories include preparation of a trade deal, concerns about tracking features built into the chips, or an effort to support domestic production.

Graham Webster of Stanford does not believe the last one because of the enormous demand.

For practitioners, the takeaway is this. Open Chinese models are unlikely to get weaker in the near term because of the bans. But regulatory risk is growing around hardware suppliers and cloud providers: if investigations confirm smuggling, intermediaries and Nvidia itself could be hurt. Teams that depend on renting GPUs from overseas providers should find out in advance who their ultimate owners are.

Story

Key Questions Answered

How to tell from the weights what a model was trained on?

Will DePue previously worked as a researcher at OpenAI and now heads the nonprofit Unemployment Labs. He analyzed the weights of open models from DeepSeek, Moonshot AI, MiniMax, and Ant Group. The weights are a file of numbers that stores everything the model has learned. According to the researcher, they also retain a trace of the configuration of the chips the training ran on. It is like the mark a machine tool leaves on a part: from it, you can infer what equipment made the part.

What this proves and what it does not?

The finding does not mean the models were trained only on Nvidia hardware. Nor does it mean the law was necessarily broken. Chinese companies often rent compute in data centers abroad, a popular way to get access to capacity. According to Epoch AI, around 275,000 Nvidia accelerators were shipped to Malaysia in 2024–2025, and an entire industrial sector has grown up around them there. Critics argue that this is where the loophole lies.

What are the key details on The Nvidia scandal?

The topic has become hot because of Megaspeed, which is linked to chip shipments to China. The company has long operated in this market through its subsidiary Speedmatrix, and US and Singapore authorities are looking into it. Bloomberg reported that a firm connected to Megaspeed executive Alice Huang took part in smuggling banned chips into a data center in Shanghai. According to TechCrunch, Nvidia CEO Jensen Huang stepped down from the company this week after admitting to a romance with Alice Huang.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…