What Is the Chip Security Act and the Blackwell Fingerprint Finding?
Researchers found numerical traces of Nvidia Blackwell chips in Chinese open models. How it works, what it can't prove, and what the Chip Security Act does.
For Executives, Security and IT

What is the Chip Security Act, and what did researchers find in Chinese AI models?
The Chip Security Act is a proposal that would require location verification for advanced AI chips covered by US export controls. The goal is that a shipment approved for one destination can't quietly end up somewhere else. It arrived alongside a research finding: Will DePue examined six released open-source models from four Chinese labs, including DeepSeek and Moonshot, and reported numerical fingerprints consistent with Nvidia's Blackwell GB200 and GB300 chips. Those chips are subject to export controls.
TL;DR
- Researcher Will DePue reported numerical patterns consistent with Nvidia GB200 and GB300 chips in six open models from four Chinese labs, including DeepSeek and Moonshot.
- The method works because the starting numbers a model is built from carry traces of the hardware, and some of those traces survive training.
- The finding does not show where the chips were, how they were obtained, or whether a full training run happened on them.
- The Chip Security Act would require location verification for advanced chips covered by export controls.
- Bloomberg reports OpenAI told lawmakers it supports tighter chip controls, and Anthropic had already backed them. The measures have bipartisan support.
- For most workers, the near-term effect is uncertainty about which AI models and providers stay easy to use, not a change in daily tools.
How can a model reveal the chips used to train it?
An AI model is a very large collection of numbers called weights. A modern model can have billions of them. Before training starts, every weight has to be given a starting value. This is called initialization.
Here is the part that makes the finding possible. The hardware helps determine how those starting numbers are arranged, and some traces of that initial state survive training. Researchers can work backward from a released model and ask which hardware could reproduce the patterns they see.
So the model itself can contain evidence about the equipment used to make it. You don't need a leaked invoice or a whistleblower. You need the published weights and a way to test which hardware produces matching patterns.
What did Will DePue find?
According to the reporting on his work, DePue looked at six released open-source models from four different Chinese labs, including DeepSeek and Moonshot. He found numerical fingerprints consistent with Blackwell GB200 and GB300 chips.
Those two chips matter because both are covered by US export controls. The commentary around the finding argues that they should not have shown up in open models from China at all.
Note the wording: "consistent with." A fingerprint is evidence that points toward certain hardware. It is not a signed confession, and it has not been presented here as proof that any specific lab broke any specific rule.
What can't the fingerprints prove?
This is where a lot of headlines will overreach. The fingerprints support the idea that Blackwell access existed somewhere in the process. They leave several big questions open:
- Where were the chips at the time? The patterns say nothing about geography. Chips in a data center outside China, rented by a Chinese lab, would leave the same traces.
- How was access obtained? Purchase, rental, a partner, or diversion are all possible, and the numbers can't distinguish between them.
- Did the full training run happen on those chips? A model could have been partly trained, tested, or fine-tuned on that hardware and mostly built elsewhere.
What the finding does change is who can investigate. Researchers and officials now have a technical check they can run independently of what a company says about its own hardware. That is new, and it is why the finding is feeding the policy debate.
What would the Chip Security Act do?
The core idea is simple. Advanced chips covered by export controls would need location verification. A shipment sent to an approved destination should not be able to disappear from view and end up somewhere else.
The bill sits in a larger argument about routes. One route that comes up is renting compute: Chinese firms using data centers in places like Singapore. If sales through those channels were closed or restricted, it would change how Nvidia's chip sales actually work. That is why people are watching Nvidia's next earnings call for how the company talks about its order book and its exposure.
There are also reported allegations that chips were moved to other places. China denies them, and how that plays out is not settled. Treat it as part of the political climate around the bill, not as established fact.
Why are OpenAI and Anthropic backing tighter controls?
Bloomberg reports that OpenAI told lawmakers it supports a package of tighter chip controls. Anthropic had already backed those measures. The package has support from both parties.
When both leading labs and both parties line up behind the same idea, the odds of action go up. The stated aim is closing the routes through which restricted computing power can still reach Chinese developers. The sources here don't spell out the labs' private reasoning, so it's fair to say what they did (publicly support the controls) and leave motives aside.
How does this connect to the price of AI?
It connects more than you might expect. Many companies are shifting routine work to cheaper models. Spending data from the expense platform Ramp showed very little growth in business token spending in August and September. The commentary also notes that some people using open-source models report costs 80 to 90 percent lower, with equivalent results on certain tasks.
Open models, including those from Chinese labs, are part of why cheaper options exist. If export controls tighten, the pace at which those labs can train new models could change. If scrutiny of how they were trained grows, some organizations may get nervous about using them. Neither outcome is certain. Both are plausible enough to plan around.
What does this mean for you?
If you only use AI tools at work. Nothing changes today. Your assistant still works. The thing to notice is that the cheap-and-good model market you may be benefiting from is tied to geopolitics.
If you pick or approve AI tools (managers, IT, procurement, legal). Ask where the models behind your vendors come from. Are any open models from Chinese labs in the stack? Do you have a policy on that? Check with your legal and security teams rather than assuming. Rules could shift, and you want the question answered before it becomes urgent.
If you build workflows on a single model. Don't. Keep your prompts and processes portable so you can swap models if a provider changes price, availability, or policy. Test a second option now, while nothing is on fire.
If you follow the industry for your job or investments. Watch three things: whether the Chip Security Act advances, how Nvidia describes demand and sales channels on its next earnings call, and whether more researchers replicate the fingerprint approach. A single analysis is a lead, not a verdict.
A simple habit helps: for any AI tool you rely on, write down the model it uses, the vendor, and your backup. Five minutes per tool.
Could you build your own AI model and vendor tracker?
If you're the person who has to answer "which models are we using, and who approved them," a spreadsheet gets stale fast. With Remy, you can describe a small internal app in plain English: a register of AI tools and models, a request form for new ones, an approval status, and a reminder to review each entry. Remy builds and deploys it as a full-stack app with sign-in and a database. It also offers nearly 400 AI models from about 60 vendors, so a team can test a backup model without managing separate API keys.
Frequently Asked Questions
Does the fingerprint finding prove Chinese labs broke export rules?
No. It shows patterns consistent with Blackwell GB200 and GB300 chips. It doesn't establish where the chips were, how access was obtained, or whether the full training run used them. Those questions remain open.
How can hardware leave a trace in a finished model?
Hardware helps determine how a model's starting numbers (its initial weights) are arranged. Some traces of that starting state survive training. Researchers work backward from the released model to find hardware that reproduces the pattern.
Is the Chip Security Act law?
Based on the sources here, it is a proposal. It would require location verification for advanced chips under export controls. The reporting describes bipartisan support and backing from Anthropic, and says OpenAI has told lawmakers it supports tighter controls, but it does not say the bill has passed.
Should my company stop using Chinese open models?
The sources don't support a blanket answer. Review which models your tools rely on, ask your legal and security teams about your risk, and keep a backup model ready so you can switch if rules or vendor policies change.
Why does this matter for Nvidia?
Closing routes such as rented compute in places like Singapore could change how Nvidia's chip sales work. Investors are likely to listen for this on Nvidia's next earnings call, alongside questions about demand and financing.
How we source, write, and correct our articles. Editorial standards
Related articles

Are AI Cyber Models Behind the Bank and Airline Outages?

What Caused the Microsoft 365 Outage? Network Fault, AI Debate

How Did Attackers Abuse Binance's Internal AI Tool?
