Did China just close the AI gap?

Did your AI-obsessed friend suddenly switch from gushing over Anthropic’s Fable 5 to frothing at China’s new Kimmy? Or was it Kim K? Jimmy Kimmel?

We got you. 

The tech world’s latest earthquake comes from a Beijing-based startup, Moonshot AI. That name is not just a nod to any ‘impossible mission’ vibes but also Pink Floyd’s legendary 1973 album (the firm’s Chinese name translates literally to The Dark Side of the Moon). To boot, the startup’s rock-mad founding trio (Yang, Zhou, and Wu) launched Moonshot on the album’s exact 50th anniversary.

So name trivia aside…. why so special? 

First, this new Kimi K3 model is powerful, performing up there with Anthropic’s Fable 5 and OpenAI’s GPT5.6 Sol in many metrics, vaulting past not just China’s original AI earthquake (DeepSeek) or even the sequel (Z.ai) but also big US names like Google, Musk, and Meta.

Second, it’s efficient, pioneering new tricks to only use <2% of its massive brain (2.8 trillion parameters) at any one moment. That means it’s roughly half the cost of some top US models, despite matching or beating them on many tasks — like Cape Verde matching Argentina with half the players.

And third, it’s open-weight^* , meaning (subject to a 27 July release) anyone can download the model’s trained underlying parameters (weights).

So Kimi K3 is not just powerful and efficient, but yours to (say) modify and tweak, rather than route constantly via your pricey tech overlords.

Now that hopefully gives a sense why the AI world has gone bonkers: Kimi K3 itself has already had to pause its subscriptions amid a flood of user demand, and is now prepping to ride the IPO wave within months.

As for everyone else…?

• Shares in local rival Z.ai have just crashed 30% (!) in a week

• US-based ‘EDA’ stocks (Cadence, Synopsys) are down 9% because Kimi K3 apparently designed real chips without having to rely on their pricey tools, and

• US-based chip stocks (Nvidia, AMD) are down 5% amid investor worries that China’s open models could (again) reduce demand for high-end chips.

It’s all giving DeepSeek 2025 release energy, which brings us to our next question: China’s AI hype faded a little last year, so surely this year’s buzz will fade too..?

And sure, there are similarities in the way Wall Street is again now questioning America’s frontier pricing power + massive chip spend…

• Last time, DeepSeek didn’t kill demand for premium US models and chips, which quickly re-established their dominance, all while…

• Our world’s overall AI spending pie just kept growing faster than any single new product could disrupt (à la Jevons Paradox).

Plus there are familiar debates around Moonshot’s US dependencies:

• Efficiency aside, Moonshot will still need top chips (though maybe more of the memory variety sold out of Korea), and…

• US pioneer Anthropic already accused Moonshot (way back in Feb) of distilling top US models to train and improve Moonshot’s own versions.

There’s now a fierce and nerdy debate raging over that second point, given Kimi’s betters (Fable 5 and GPT 5.6) only appeared a few weeks earlier — not a lot of time to ctrl-c + ctrl-v everything, though others argue you only need a few well-chosen reasoning traces to effectively reverse-engineer.

The other factor is data, with lesser-known but critical firms like Mercor charging those US AI labs huge premiums for early dataset access to train their models, but then opening it all up to the rest at bargain-basement prices within weeks.

Still, details aside, the bigger picture is getting clearer — maybe DeepSeek was less a one-off, and more a pattern now stabilising via Moonshot: China can develop near-frontier AI at a fraction of the US cost, with America’s lead now measured in weeks.

Maybe that’s why President Xi just seemed so upbeat (by Marxist-Leninist standards at least?) at Shanghai’s World AI Conference — especially after 29 nations just agreed to form a new global AI cooperation body hosted in… Shanghai.

Sound even smarter:

• ^* Sticking with car analogies for AI…

  • Open-weight (Kimi K3) basically means you own the car

  • Closed-model (many big US names) means you just rent it, and

  • Open-source (Meta’s Llama) means you own the car and blueprints and factory tools.

• But how do open models make money? The details differ, but they basically hope to win market share, then profit from the broader ecosystem that emerges.