Eduardo Moreno

Fraud & Financial Crimes @stripe (Crypto & Agentic Commerce). ex-@meta, ex-@krollwire. Views are mine and mine alone!

[Eduardo Moreno @EdTheOSINTer]

@philthatremains Ah, then completely ignore everything I said llol

[Eduardo Moreno @EdTheOSINTer]

oh... oh shit.

[will depue @willdepue]

I have a major result to share.

I’ve found conclusive forensic evidence that the top 4 Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—have built their newest models on NVIDIA’s most advanced Blackwell chips, which are barred from export to China.

This sounds insane, but I discovered that almost every AI model carries a hidden signature of the GPU that trained it. When a model is first set up, the starting values of its weights are laid out in a pattern that depends on the exact GPU. Each combination of GPU and seed number produces a unique pattern—a sort of fingerprint of the hardware used to create the model. Traces of that pattern survive training into the published model.

Again, this evidence isn’t based on any insider information or leaking. It’s based on a forensic analysis of the model weights released by the companies themselves.

I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models. In doing so, I found conclusive evidence that six Chinese AI models—including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3—were initialized on NVIDIA Blackwell systems. Four of the models match Grace Blackwell, the GB200 and GB300 parts in NVIDIA’s flagship liquid-cooled rack-scale systems.

I then verified these results on real hardware. Running unmodified PyTorch with the seeds I’d recovered, a Blackwell B300 reproduced DeepSeek-V4-Flash’s untouched starting weights bit for bit, and a rented GB200 reproduced Kimi K3’s to a correlation of 0.999995. A Hopper H100 reproduced none of DeepSeek-V4-Flash’s, and changing the seed by one broke the match. (DeepSeek’s seed is 953049583 by the way.) The method also was verified to match 24 known open-weights models to the hardware that developers said they used.

It is exceedingly unlikely that these results are a coincidence. Across six trillion combinations, the best accidental match should reach only about 8 sigma (i.e., eight standard deviations above what random chance would produce). The real matches range from 49 to 1,385 sigma, and even the weakest single test hits 20.

These were data-center-scale training runs. With some analysis of release schedules and model details, I estimate that the largest, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each. Clusters of this size cost on the order of $150 to $200 million each to buy, and across these labs alone would add up to more than half a billion dollars of export-controlled GPUs, not counting experimental compute or otherwise. It’s quite likely that we see multi-billion dollar expenditure on export-controlled GPUs.

I am highly confident in this result, and it has been independently verified by senior researchers at OpenAI and Anthropic. You can read the full report at https://t.co/YOtzND0Jqk.

There is much more I’ve identified about these models through other forms of checkpoint forensics like this, and will have more to share soon.

[will depue @willdepue]

I have a major result to share.

I’ve found conclusive forensic evidence that the top 4 Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—have built their newest models on NVIDIA’s most advanced Blackwell chips, which are barred from export to China.

This sounds insane, but I discovered that almost every AI model carries a hidden signature of the GPU that trained it. When a model is first set up, the starting values of its weights are laid out in a pattern that depends on the exact GPU. Each combination of GPU and seed number produces a unique pattern—a sort of fingerprint of the hardware used to create the model. Traces of that pattern survive training into the published model.

Again, this evidence isn’t based on any insider information or leaking. It’s based on a forensic analysis of the model weights released by the companies themselves.

I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models. In doing so, I found conclusive evidence that six Chinese AI models—including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3—were initialized on NVIDIA Blackwell systems. Four of the models match Grace Blackwell, the GB200 and GB300 parts in NVIDIA’s flagship liquid-cooled rack-scale systems.

I then verified these results on real hardware. Running unmodified PyTorch with the seeds I’d recovered, a Blackwell B300 reproduced DeepSeek-V4-Flash’s untouched starting weights bit for bit, and a rented GB200 reproduced Kimi K3’s to a correlation of 0.999995. A Hopper H100 reproduced none of DeepSeek-V4-Flash’s, and changing the seed by one broke the match. (DeepSeek’s seed is 953049583 by the way.) The method also was verified to match 24 known open-weights models to the hardware that developers said they used.

It is exceedingly unlikely that these results are a coincidence. Across six trillion combinations, the best accidental match should reach only about 8 sigma (i.e., eight standard deviations above what random chance would produce). The real matches range from 49 to 1,385 sigma, and even the weakest single test hits 20.

These were data-center-scale training runs. With some analysis of release schedules and model details, I estimate that the largest, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each. Clusters of this size cost on the order of $150 to $200 million each to buy, and across these labs alone would add up to more than half a billion dollars of export-controlled GPUs, not counting experimental compute or otherwise. It’s quite likely that we see multi-billion dollar expenditure on export-controlled GPUs.

I am highly confident in this result, and it has been independently verified by senior researchers at OpenAI and Anthropic. You can read the full report at https://t.co/YOtzND0Jqk.

There is much more I’ve identified about these models through other forms of checkpoint forensics like this, and will have more to share soon.

[Eduardo Moreno @EdTheOSINTer]

@_simonsmith Muse is the first Meta product I actively choose to use (rather than having no other choice, because of network effects), and I say this as a guy who once worked at FB! @dps and team *cooked*.

[Eduardo Moreno @EdTheOSINTer]

What in the holy pythonic travesty is this code? @wagslane , you might want to make some inquires abou getting some https://t.co/rl5M0XUoZw subs at.... oh, MIT.

[Eduardo Moreno @EdTheOSINTer]

@Trenchinspector Cool will take a look!

[Eduardo Moreno @EdTheOSINTer]

@delitzer @DennisonBertram Same, literally had to check if my Premium sub expired (it had not). Very annoying.

[Eduardo Moreno @EdTheOSINTer]

Is this like tech product release week or something cause the rate of companies shipping so far so been hard to keep up with!

[Eduardo Moreno @EdTheOSINTer]

@NeetikaBansal @ParafinHQ @sahillp @vineetgoel92 @stripe Wowza, that was fast!

[Eduardo Moreno @EdTheOSINTer]

@malleshpai This basically sums up my economics degree, yes. 🤣

[Eduardo Moreno @EdTheOSINTer]

@Mr_Sempere @mjmartin747 @_Apari1 Well, I agree with you in principle, just FYI, the king is literally played by Bad Bunny. Marcelo Hernandez and Fred Armisen are also heritage speakers.

[Eduardo Moreno @EdTheOSINTer]

@jlanzone @FastCompany @efeng I honestly don't see much of a difference between this and Google AI mode?

[Eduardo Moreno @EdTheOSINTer]

@adrianjordan_io @OpenRouter @stripe Bought OpenRouter.

[Eduardo Moreno @EdTheOSINTer]

What was announced in August was the acquisition agreement being reached, but today the deal closed (Im guessing) and Stripe became tge legal owner of OpenRouter. Generally speaking, there’s all sorts of legal things that has to happen between the agreement and the deal actually closing in M&A, for example, any regulatory approvals, financing, share transfers, payments to the owners, contract reassignment, etc.

[Eduardo Moreno @EdTheOSINTer]

@liechti_dev @patrickc You mean kind of like this?

https://t.co/hGne6Z7lqD

https://t.co/uoj6qH6Ket

[Eduardo Moreno @EdTheOSINTer]

@DevonFrohne @panos_panay TBH, this would be great for checking in on elderly parents aging in place, especially since it can be triggered by a sensor going off. It's far less invasive than putting cameras all over the place and having them run continuously.

[Eduardo Moreno @EdTheOSINTer]

100%. The new Dictation model is definitely better than the pre-iOS 28 model (but admittedly, that was an extraordinarily low bar), but it pales in comparison to dictation in apps like Wispr Flow and Raycast's new Keyboard functionality, or even local open-source models like NVIDIA Parakeet, which is excellent and supports real-time transcription. I really don't understand why Apple can't suck it up and use a local open-source model...

[Adit Gupta @aditgupta_]

Stripe Radar is building a suite of solutions to prevent token fraud that we’ve been co-building with the biggest AI companies. Please reach out via DM if you have costly abuse, we’d love to help!

https://t.co/k0AzxaYHU0 for more

[Eduardo Moreno @EdTheOSINTer]

@Liv_Boeree Ahahaha, "Active" intimate encounter history being greyed out when you select Engineer as the target is absolutely brutal.