Remember DeepSeek? The chinese model that matched OpenAI at a fraction of the cost and wiped $600 billion off Nvidia’s market cap in a single day. Well, it just happened again.
At the World Artificial Intelligence Conference in Shanghai, a Beijing startup called Moonshot AI unveiled Kimi K3.
Demand for Kimi K3 was so high, Moonshot had to pause new subscriptions within days of launch. Bloomberg called it “China’s second DeepSeek moment.”
So, what exactly is Kimi K3 and why does it matter this much?
Moonshot AI was founded in March 2023 by Yang Zhilin – an AI and machine learning (ML) researcher. In just three years his team has built something that has put every major AI lab on notice.
Kimi K3 is a 2.8 trillion-parameter AI model – the world’s largest open-weight AI model ever released. In simple language, parameters mean more capacity to understand nuance, reason through complexity and generate accurate, sophisticated responses.
Most current publicly known models operate in the hundreds of billions of parameters. Kimi K3 is built at a scale that would have been considered impossible twelve months ago.
But size alone isn’t the story. The story is what the model can do with that scale.
Two of the most respected third-party AI benchmarking organisations tested Kimi K3 against the world’s leading models and found it delivers performance comparable to Anthropic’s frontier Fable model and substantially outperforms OpenAI’s GPT 5.6 Sol.
On coding, reasoning and agent tasks – the categories that matter most for enterprise and developer adoption – Kimi K3 is at or near the top of every benchmark.
And then there’s the cost. Kimi K3 completes each intelligence index task for $0.95 compared with $1.04 for OpenAI’s flagship model.
Frontier performance. Lower cost. Open-weight, meaning the underlying model weights are published publicly and can be downloaded, modified and deployed by anyone.
The architecture is built on a mixture-of-experts system, which is a design that activates only the relevant parts of the model for each task rather than running the full 2.8 trillion parameters simultaneously. This is why inference costs stay competitive despite the enormous scale: the model is large enough to reason at frontier level but efficient enough to run without the economics of a hyperscaler data centre.
The cloud play hiding inside the AI story
Moonshot AI is a private company. You can’t buy shares in it.
What you can buy is Alibaba (NYSE: BABA), which holds a 36% stake in Moonshot, making it the largest outside shareholder in the startup that just launched the world’s most capable open-weight AI model.
That stake was not a passive bet. When Moonshot raised $2 billion in May 2026 at a valuation of more than $20 billion, Alibaba was at the table.
Now Moonshot is approaching a Hong Kong IPO at more than $30 billion. Alibaba’s 36% stake, carried at the May valuation, is worth over $7 billion and rising.
But that’s not actually the main reason Kimi K3 matters for Alibaba investors.
Bernstein analysts put it plainly in a note published Friday: Kimi’s rapid rise is likely to become “a meaningful catalyst for Alibaba Cloud.”
Here’s why…
Moonshot doesn’t run Kimi on its own infrastructure. It runs on cloud compute. As the most capable open-weight model in the world, Kimi K3 will be downloaded, deployed and integrated by developers, enterprises and governments across China and beyond – all of whom need cloud infrastructure to run it at scale.
Alibaba Cloud is the dominant cloud provider in China. Every enterprise deploying Kimi K3 in production is a potential Alibaba Cloud customer.
That’s before you add the Apple partnership. Alibaba is Apple’s AI provider in China, meaning Apple Intelligence in the world’s largest smartphone market runs on Alibaba infrastructure.
The DeepSeek lesson and opportunity…
When DeepSeek launched in January 2025, the market sold Nvidia.
The reasoning?
If Chinese models are this efficient, the trillion-dollar AI hardware buildout might be unnecessary. Nvidia recovered. The AI infrastructure buildout accelerated.
So, what’s the lesson?
Efficient models don’t reduce compute demand. They democratise access to AI, expand the universe of users and increase total compute demand.
DeepSeek didn’t kill the AI buildout. It accelerated it.
Kimi K3 follows the same logic. A 2.8 trillion-parameter open-weight model available for free means every company that couldn’t afford frontier AI last year can now experiment with a model that rivals OpenAI. That experimentation becomes deployment. That deployment becomes cloud spend. The companies best positioned for that wave are the ones with the infrastructure, distribution and cloud relationships to capture it. In China, that’s Alibaba.
That gap between what’s been priced and what the sum of these developments implies is the opportunity.
Not a subscriber to Money Morning?
You can get free daily recommendations like these with Money Morning eletter. Just sign up here.
