What is the issue? Chinese companies are allowed to rent Nvidia hardware from abroad. Why does everyone suddenly seems to forget that?
Most interesting (and imo important) finding of this is that there is no evidence that DeepSeek V4.1 Flash was initialized on Nvidia hardware.
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 unemploymentlabs.com/checkpo….
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.
1
42
The article admits that the GPUs could have been rented. Yet it leads with "which are barred from export to China".
1
11
Petr Pilař retweeted
Banning open-source AI would hurt defenders 10x more than attackers, which would make the world 10x more dangerous and this is a good example why!
Pretty impressive. Model is 1.3B and used only ~1300 hours of data.
Expansion beyond game data should be relatively straightforward, by utilizing existing realistic videos and labelling the camera movement (and possibly other actions) in those to create (action -> outcome) dataset.
Matrix-Game 2.0 — The FIRST open-source, real-time, long-sequence interactive world model
Last week, DeepMind's Genie 3 shook the AI world with real-time interactive world models.
But... it wasn't open-sourced.
Today, Matrix-Game 2.0 changed the game. 🚀
25FPS. Minutes-long interaction. Fully open-source.
1
140
OpenAI is reportedly working on Chinchilla-like scaling laws for the optimal number of CEOs per employee.
1
76
GitHub should have a built-in Kickstarter-like feature for PRs. Often I would be happy to pay a few thousdand for a specific feature to exist, but there isn't a clean way to do that today.
1
76
Stripe Managed Payments as MoR seems to be really low value proposition. They only support SaaS and only support EU, UK and US. UK and EU are the easiest countries to file VAT in -- literally takes less than a minute to upload the right CSV for whole EU into the German tax portal.
1
107
It's incredible how archaic USA is compared to EU in this regard. Why isn't there a centralized portal that lets you report and pay Sales Tax for whole of US?
Stripe Managed Payments as MoR seems to be really low value proposition. They only support SaaS and only support UK and EU. These are the easiest countries to file VAT in -- literally takes less than a minute to upload the right CSV for whole EU into the German tax portal.
There's a new version of this post
1
60
Qwen 3 235B-A22B seem to be actually good. Nice little model. LlamaCon is going to be interesting.
98
Petr Pilař retweeted
see if we can finish the job tonight #Qwen3 👻
The AI vibes took a hard turn in Google's favor (*). Best model is all you need?
Curious how Claude 4 🦁 will shake things up.
(*) At least in the techy forums forums and Discord groups I participate in.
1
77
Did CoreWeave pay for this article?
"""
While we are grateful that they have SOC2 compliance, we cannot place them any higher for now since they only recently got SOC2 compliance.
"""
Why does it matter when they got certified, if they are certified now.
"""
Due to its roots as an ex-Russian cloud provider, it boasts an exceptionally talented team of cracked ex-Russian engineers.
"""
Cool cool cool.
SemiAnalysis, you can do better.
Today we introduce the first ever Neocloud Rating system - ClusterMAX
12 months of technical testing & benchmarking of 30+ Clouds covering >90% of the market
Hyperscalers vs Neoclouds, Coreweave IPO, Neocloud bubble burst
GPU cloud deep dive by provider
semianalysis.com/2025/03/26/…
1
6
3,674









