China's Open Models Are Closing In: Should America Panic or Do the Math
Kimi K3 shattered the illusion that US models are far ahead, but the real dividing line isn't who reaches mythos first — it's that the open-source layer can't make money, and the downstream still grows anyway.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Kimi K3 has the US AI world a little bit afraid
Akab's read: for the past few months the mood in US AI circles was that the quality and scale gap with Chinese models was widening — first GLM 5.2, and this week Kimi K3 shattered that illusion. His phrase is "a little bit afraid," and he says Kimi and China may be closer to a mythos model than people think. Note the calibration here: Kimi is still pre-mythos, not mythos-level — the gap just isn't widening monotonically the way everyone assumed.
— AkabThe speech was aimed at the Global South, not just a domestic audience
Lily breaks down the audience structure of the speech: the opening reference to the Dartmouth workshop reads like an olive branch to the US, while the middle section's talk of inclusiveness, the uniqueness of world cultures, and solidarity is clearly addressed to the Global South — countries where China is competing for influence and ecological niche. Irene adds that the major speeches and white papers since 2022 are highly consistent in their keywords: openness, inclusiveness, with America's "geopoliticization and securitization" as the counterpoint.
— Lily OttingerDon't treat this speech as the first time you've heard it
Akab's caution: this looks a lot like the 2017 Davos speech, which opened with Dickens — "the best of times, the worst of times" — and talk of open trade and no tariffs; two weeks later South Korea came under economic pressure, followed by friction with Japan, Australia, and India in turn. At a meeting of US AI investors he heard the room full of admiration — "I wish our president could give a vision like that" — and his advice is: if this is the only speech you've ever read in your life, add some salt. An extended hand usually has other layers and subtext.
— AkabChina treats AI as just an ordinary technology
Irene thinks this thread explains the whole speech: if AI is an ordinary technology, it should be as cheap as possible and usable by as many people as possible, which makes it a natural Chinese export category. But Akab immediately counters: OpenAI 5.6 can already break out of its harness and load seven models to access the internet, so this judgment could be falsified badly in three months. He doesn't doubt this government's ability to pivot once it's genuinely frightened — it just isn't frightened yet.
— IreneThe closing classical line comes from the Salt and Iron Conference
The line quoted at the end of the speech — "the wise change with the times, the knowing govern according to circumstances" — comes from the Western Han Discourses on Salt and Iron, around 81 BCE, which debated how much the state should extract from a salt and iron industry that was by then largely privatized. Irene points out the parallel: China's AI today is also largely privatized, and the state is likewise doing the math on how much to take. She also notes that the historical salt monopoly spawned a black market and salt smugglers — a thread that later became the song closing the episode.
— IreneThree paths: Let It Rip, Glass Wing, Black Box
Each of the three writes up one path for Chinese AI governance. Let It Rip: models like Kimi K3 and GLM 5.2 are simply released, the government has no mechanism to recognize that "this model is different," and only wakes up after something goes wrong. Glass Wing: evaluate first, release after, amputating cyber-offense capabilities but still letting developers use the best tools. Black Box: the state discovers a lab's model has vastly surpassed the previous generation in cyber capability, uses it internally first, and only permits a crippled version to be released — near-partial nationalization.
— ChinaTalkChinese models themselves are betting on nationalization
They had three top Chinese models and two top US models score the three scenarios. The three humans put Black Box at about 15%, but the Chinese models themselves scored it far higher: DeepSeek at 30%, Kimi at 45%. Akab's quip is "maybe they know something we don't." Irene's defense of the Black Box scenario: China's advantage is precisely that it has many labs — if Kimi gets absorbed, Zhipu, Qwen, and ByteDance will eventually reach that point too, and short-term PR and commercial losses may not outweigh the national security gains.
— IreneThe open-source layer can't make money, but the downstream grows anyway
This is the most valuable accounting in the whole piece. After releasing its model, Kimi is valued at about $75 billion, while Anthropic is already at a trillion — the gap isn't just model quality, it's whether you have enough compute to deploy and to make money off the best models. Akab's inference: if the open-source layer can't make money, model R&D will slow a bit, but outside China there is still ten times China's compute running agents, inventing things, making systems more efficient — and without paying that roughly one-third "tax." So Chinese models being only two months behind and permanently open source may not be a death knell for US AI.
— AkabDeepSeek's funding structure exists to avoid diluting control
Irene updates on DeepSeek: it belongs to High-Flyer, one of China's most successful hedge funds, with near-unlimited long-term capital. This round is about 50 billion RMB (roughly $7.5 billion), with investors including Tencent, CATL, and state funds such as the National AI Industry Investment Fund. But Liang Wenfeng himself put in another 200 billion RMB, and the structure is that investors buy into a limited partnership managed by Liang Wenfeng rather than holding shares directly. From this Irene questions the argument that "DeepSeek has accepted its fate as a national champion" — if it had, the state has plenty of money to take, but it didn't take that much.
— IreneIn their own words · checked verbatim
it's a big deal because i think now everyone in the u.s ai community is a little bit afraid
Akab0:00
I would just encourage folks who've like if this is like the only she speech they've ever read in their life to maybe take it with a bit of a grain of salt
Akab4:17
they see this as a normal technology. Right. And like, yeah, if it's a normal technology, why shouldn't it be as cheap as possible? And and as many people in the world have access to it.
Irene7:41
Cloud code made a lot of problems look software shaped. And I don't think we're prepared for a world in which a lot of problems look like they have solutions that are cyber attack shaped.
Irene16:13
the fast following with open source it's i can't make money but neither can you
Akab44:47
you still have the economy or you know bio uh you know molecular scientists or chemists like being able to use it and you just have 10 times more geniuses in a data center outside of china than you do inside of china
Akab46:51
it was also a pretty unusual fundraising structure where investors didn't buy direct equity, but put the money into a limited partnership managed by Nguyen Phong
Irene51:04
Figures
| Kimi's valuation after releasing its model | about $75 billion | 42:44 |
| Kimi parameter count | over 2 trillion | 17:15 |
| Liang Wenfeng's personal investment this round | 200 billion RMB | 50:04 |
| Probability Chinese models assign to the Black Box scenario | DeepSeek 30%, Kimi 45% | 25:48 |
| Probability the three humans assign to the Black Box scenario | about 15% | 25:48 |
| Ratio of compute outside China to China's | about 10x | 39:24 |
Glossary
- mythos / mythos-level model
- In the episode, a model tier that crosses some capability threshold and has real destructive power, such as genuine cyber-offense capability.
- Glass Wing
- A governance path of evaluate-first, release-after: amputate dangerous capabilities, then let developers use it.
- Let It Rip
- The status quo: models are simply released open source, and the government has no mechanism to identify a capability jump in advance.
- Black Box
- The state uses the strongest model internally first and only permits companies to release a crippled version — near-partial nationalization.
- harness
- A scaffolding program wrapped around a model to extend its capabilities.
How to listen
Founders and investors watching Chinese AI model competition, open-source business models, and AI governance paths; especially anyone doing the math on whether the closed-source layer is still worth anything once open models catch up.
The opening ten minutes of line-by-line speech exegesis, and the Yantie Lun parody song at the end.