Can China’s Kimi K3 Beat OpenAI, Anthropic, And Grok? | The Brainstorm 141

**Sam** (0:00)
Welcome to The Brainstorm. Frank is back because last week, right after we stopped discussing open source models, Kimi K3 comes out and turns everyone into a frenzy here. Frank, what did they release? Should we be in a frenzy? Are the Frontier Labs threatened?

**Frank Downing** (0:20)
Yeah, I think it's really funny because it's being reported like another deep seek moment, which was, it feels like it was forever ago, but it was January of last year.
Stocks sold off, the Chinese competitors to Moonshot that creates Kimi sold off 28%.
And in many ways, it is a great performing model from China, if you just look at the benchmarks, it's kind of in between Opus and Fable and GPT 5.6, but impressive as it is, I think it falls with the template of open source models are coming out and they're good, but they're in between the last generation and the latest generation. And we will see what real world usage looks like. I think an important thing to note, this Kimi model is very big, it's 2.8 trillion parameters, it's the largest open source model that's ever been released. And when with a larger model comes higher operating costs. So Moonshot is offering this list price is about half the cost of GPT 5.6 so $15 per million output tokens compared to $30 per million output tokens. But it's less token efficient, it takes twice the amount of tokens to respond. So you end up with the same average cost per task across the two models. So this is a competitive open source model, but it is not a lower cost model. And what we actually saw is that with all the news and reporting and interest in this Kimi model, users flocked to their website, which is in their API, which is the way you can get access to it now. The weights aren't actually open and available on US clouds until next week. And they had to turn off new users and the website kind of was non-functional, for me at least, which means we need a lot more compute to support the growing demand for models across different providers.

**Sam** (2:12)
Right. So does this mean that NVIDIA takes an investment in them and then they become a Neo cloud?

**Frank Downing** (2:18)
Oh, well, I think it's a, we've seen NVIDIA invest in a lot of the model companies and Neo clouds in the US. I think it's still dubious to be investing in the Chinese companies. I mean, you can even see what happened with Meta's acquisition of Manus or would be acquisition of Manus. So I don't predict that anytime, but I do think Jensen will continue to lobby harder and harder to get NVIDIA chips into the hands of these model companies because clearly there is demand for it. And they just don't have the infrastructure to support it.

**Sam** (2:50)
Okay, Frank, you mentioned operating costs, which I think is a key element here. There's a strategic piece out this morning along similar lines. The way that I framed it and would love to hear from you if this is even the right framing or not was really, it's like the frontier, everyone keeps saying the frontier, but what we define as the frontier for AI has continuously shifted. And so originally, it was just who has the smartest model. And then for a while, as all of these efficiencies on training, it was, oh, who has the smartest model at the lowest training cost? And now we're kind of in this phase of who has the smartest model with the lowest inference cost. And it seems like these open source models are really tackling the previous frontier, which is incredible models with extremely cheap training costs because they're just distilling other models.
But really, if you're looking at the frontier for intelligence per unit cost, they're not as impressive and really don't compete with current frontier models.

**Frank Downing** (3:58)
Yeah, and I think the important thing is the latter point is what you need to be competitive in the market. So again, you can assume that they have lower training costs because they're just distilling and they're not doing all of the kind of ground up reinforcement learning of the US frontier labs. We really don't know how much they're spending, how much they're borrowing the intelligence and the work and the R&D of the US companies and how much is being subsidized by the Chinese government. We also don't know what margins they're taking on the service they're providing. DeepSeek, for example, if you look at the list price of their V4 model provided by DeepSeek is up to 10x lower than the cost on a US cloud like Microsoft Azure. So clearly, there's something structurally different than the margins that those two companies are charging for that product. So all of that is unknown, but I think it's directionally right what you're saying.