How Should the US respond to China’s Latest AI Success - DTNS 5313

**Tom Merritt** (0:09)
This is the Daily Tech News for Monday, July 20th, 2026 We tell you what you need to know, give you some important context, and try to help each other understand.

**Rob Dunewood** (0:18)
Today, Kimi is being treated like Deep Seek 2.0, and Hugging Face says that the US needs to think differently about security.

**Tom Merritt** (0:24)
Everybody named Kim out there is like, me?
I didn't do it. I'm Tom Merritt.

**Rob Dunewood** (0:30)
And I'm Rob Dunewood.

**Tom Merritt** (0:31)
Let's start with what you need to know with the big story.
So Thursday, China's Moonshot AI released its Kimi K3 model, billed as a competitor to US foundation models, when Jason mentioned it on Friday's show. And if you follow this sort of thing, then it's worth noting that it is a 2.8 trillion parameter model, which is a bigger number than any of the other published parameter numbers, though, for example, OpenAI and Anthropic no longer make the parameters of their flagship models public. However, Moonshot said that Kimi should outperform pretty much everything out there, except for Claude Fabel 5 and GPT 5.6. So why are we talking about it again today? It received so much interest that 48 hours after it launched sometime on Saturday, the company paused new subscriptions and they're slowly doling out open slots for new customers. Someone from the company wrote on X, our GPUs are feeling it.
Kimi also split its plans in two. It now has a Kimi membership that gets you the web app and the work products, and then a Kimi code membership that takes the more token intensive stuff from coders and puts it in a different plan. Now, here's another little quirk of Kimi. It doesn't plan to release its open weights until July 27th, meaning the only way to use it is to pay Moonshot for now. You can't just download it and use it yourself like you could with other open weight models until July 27th, that is. However, even when the open weights come, you're going to need some equipment. So companies might be able to do this. But if you don't run at least the equivalent of eight H100 or H200 chips from NVIDIA in your house, you're not doing it at home. This is something for people to probably do in a data center. Then Sunday, Alibaba released a preview version of its flagship, Quen 3.8, and Alibaba says only Fable 5 is better than Quen 3.8. It also is going to make Quen open weight, at least sometime soon. And that would be a change from earlier this year when Alibaba stopped releasing open weights for its models. Now, this follows what when and Jason talked about on Friday with the president Xi Jinping of China saying, we should be more open with our models. And I guess Alibaba said, got the order, boss, going to do that. Bloomberg has grouped the reaction to this model, Kimi, into three camps. Fear, the obvious one, China is catching the US., we need to do something crowd. Defense, which is a crowd of people saying, this doesn't mean what you think it means. It's a big jump from a piece of tech to a competitive product. And we saw that with Deep Seek, where there was a lot of attention to it when it first launched, and it's still a decent product, but it didn't overtake all the big foundational models as some people thought it would. And then there's the optimists.
The more competition is good crowd. In fact, Gavin Baker took this position saying that lower margins on these kinds of models benefit chip companies because they can get more out of their chips, benefit hyperscalers and benefit software developers alike. It's the rising tide lifts all boats theory.
Axios reports that the US government is clearly in the fear category right now and is considering placing some Chinese models on the entity list. That would at the very least require companies to guarantee security of any models that they wanted to use, specifically these open-weight Chinese models.
However, recent White House advisor David Sacks wrote on X, We're at a critical inflection point in AI policy. The leading closed labs, already a duopoly in terms of AI model revenue, want the government to eliminate their open-source competition. So Sacks has been warning against regulatory capture and saying that encouraging more closed models is not the way to go. You're just going to have the US fall further behind. I've seen a lot of people making that same argument, is that if you just try to wall yourself off, you're not being competitive. Now, there's a security question to these that we'll get to a little later in the show when we talk about what happened at Hugging Face. But before we do that, let's talk a little bit about what we think the reaction should be to Deep Seek Part 2, The Moonshotting, as I am calling it, Rob.