Z.ai

I don't know what's the right word.

This is like flamethrower... um just um scorched earth war that's going on on like twitter every single day right now right and it is like it is like anthropic kind of posturing in a way that uh seems to be um tilting towards restriction of the use of Chinese models, teams to be leaning in the direction of regulatory capture, and of course, dressing it all in the language of safety.

And then on the other hand, you have the extreme position which is like, hey, we're running this on our own servers, on our own GPUs, with our own guardrails, often having fine-tuned them ourselves for our own purposes, having run our own evals, we have a lot more window into what these models are and ability to like poke and prod and test them for various kinds of alignment than we do with anthropics models.

If you're like an enterprise consumer and like, how dare you take them away from us? And look, I think you can tell you a really great point, Alan, which is like, I, if I would love to live in a state of affairs where the American open source or at least American open white community had caught up with the Chinese and also was knocking on the door and like, I wish them all luck.

Like I hope think he gets there.

I hope reflection gets there because at that point, I think I, right.

If, if we both believe, which I think we're maybe open-minded, but I think tend to believe that some of the, the, the use of xenophobia is like, it's a red herring.

It's a, it's a bad faith argument to take away competition or, It seems they wouldn't be able to direct that same weapon against American open source community.

Right, with ZEI.

Yes, ZEI open sourced their Zcode platform, but they quickly had to remove a specific feature.

Yeah, there were serious privacy concerns raised by the community.

The feature essentially sent encrypted project snapshots over to Alibaba Cloud.

Because what I used to believe a lot in Nvidia, and I still think it's an extremely strong company, but I also think that their valuation has already peaked.

It's interesting to me is that Zed AI or Zipu AI from China released a model, you know, that got everybody astonished because it was super fast, super precise, super high accuracy, and it was 100% trained on Huawei chip and also running on Huawei chip.

So I think with the Nvidia export ban going on in China for the past couple of years, Huawei has managed to catch up.

And even if Huawei is still banned from exporting some of its GPUs or something to the US, I think they are going to become so strong that it's going to enable an entire AI model ecosystem in China, which is already happening.

But I want to get a point of view.

When you look at, say, the China model, say, Deep Seat, Moonshot, Minimax, Z.AI, Ali, Baba, and ByteDance, they all have different agendas, different games in terms of how they think about the large language models, or maybe even image video transformer models now.

What is the cleanest way to segment them from your point of view?

I think it's

The economics is interesting to me because depending on who you talk to and obviously, you know, I have a lot of friends in the investment space.

So I get a chance to understand what, what the unit economics per token that OpenAI and Entropic is really like, if it's $1, how many, how much profit are they taking on the table? And then the bad news now is that the Chinese is low-balling them at one-tenth of what their margins really are.

So the question for me then is, what is the right test for the Chinese model's unit economics? Because as you rightfully pointed out, you saw Z.AI, Minimax, they show explosive revenue growth.

This is not like dot-com bubble where there's no revenue, there's tons of revenues, but it reminds me a lot more like the railways, the railway bubble.

The Chinese government very likely knows they're doing this.

DeepSeek, Moonshot AI, Alibaba, Minimax, StepFun, Z.AI, they've extracted billions of tokens across millions of exchanges and requests from U.S. frontier AI models, including variants of Claude, Chachi, BT, Gemini, and Grok, since at least 2024.

These firms route their distillation requests through several pathways to gain unauthorized access to US AI models, often violating American AI companies' terms of use.

These pathways include native application programming interfaces or APIs, remote cloud providers, and third-party aggregators that automatically obfuscate user metadata to avoid detection.

It was his first ever Twitter post, apparently, was a paper about why we should support the open models.

And you're seeing now, for example, Z.AI is, I'm sorry, Moonshot, the Kimmy K3 people, they're in discussions with Google and Microsoft and Amazon about revenue sharing.

So for people to access their models

through our big cloud providers.

So DeepSeek was a company that came out of effectively a quant trading firm that decided, okay, we're going to start competing in these big AI models.

Z.AI is a different type of organization.

It's a Beijing-based group.

that has a lot of cross-pollination with Tsinghua University and had been working on different types of models and Moonshot's another startup.