Beth Lyons

So Minimax ran 13 million exchanges targeting agentic coding.

Moonshot AI ran 3.4 million exchanges targeting reasoning traces.

And DeepSeek ran 150,000 exchanges specifically prompting cloud to reconstruct its own chain of thought.

All three rooted traffic through 20,000 fake account, more potentially fake account proxy networks and and did it in a way to dodge detection.

Right.

So there was whatever that's called legally, like you tried to hide it, which meant you knew it was wrong.

Right.

So Minimax specifically redirected 50% of traffic to the newest cloud model 24 hours after its release.

Like all of those things are used to go into constructing the knowledge that they did this with full knowledge.

And, separately, a senior U.S.

official confirmed DeepSeq trained its latest model on NVIDIA Blackwell chips that were banned from export to China.

And they did so at a data center in Inner Mongolia.

So a bunch of what this says is we had some knowledge of this.

We had some suspicion of this.

And this is what's been confirmed.

I think they use paid accounts.

Which is, frankly, boilerplate for any sauce you use, right?

Somewhere in the terms of service is you can't use our service to totally build a competitor to what we do because you use the service and understand what it does.

But the other piece of this is China is the

is that China can do this cheaper and potentially faster and is not beholden to the agreements that we're making in the U.S.