Moonshot AI hit a wall. Or rather, the servers hit a wall.
Just days after launching Kimi K3, the Beijing-based firm had to stop new signups. Not because the code broke. Because it worked too well.
Demand crushed capacity.
The company admitted on X late Sunday that the traffic had pushed systems to their absolute limit within 48 hours. If you already have an account? You’re good. Everything stays the same.
But if you’re trying to join the party now? You’re locked out.
Moonshot said signups will resume gradually. “In batches,” they called it. A polite way of saying: wait your turn. We need more chips. We can’t get them fast enough.
This is the new normal.
Chinese AI labs build models that grab global headlines. Then they discover they don’t have the servers to handle the rush. The scramble is real. And it tells a story about who is winning which war.
The Kimi K3 signup halt and the compute shortage
Why did Kimi K3 halt new signups? Simple economics meets hard hardware.
Kimi K3 is a 2.8 trillion-parameter open-weight model. Open weight means developers can look under the hood. Adapt it. Build on it. Not just pay for an API.
It’s large. Unusually so.
And that size demands massive compute resources. Lian Jye Su, chief analyst at Omdia, put it bluntly. New model releases cause spikes. Spikes strain infrastructure. Moonshot simply didn’t have enough compute chips for this particular surge.
It might be an underestimation of popularity. It might be the sheer weight of the model. Either way, capacity is tight. Allocating it is expensive.
“The demand was ‘close to the limits’,” Moonshot said. Translation: we ran out of room to grow.
How Kimi K3 compares to US models in benchmarks
People are comparing Kimi K3 to US giants. Specifically, GPT-5.6 Sol and the rumored Claude Fable 5.
According to the South China Morning Post, Kimi K3 outperformed both on some benchmarks.
It topped Arena’s leaderboard for front-end coding capability right after launch. Public release. Immediate traction.
Does that mean it beats GPT-4? Or whatever the latest OpenAI model is called next week? Probably not yet. But it’s in the conversation.
Chinese frontier models are open. American models are mostly walled gardens. That difference changes everything for developers. They can tinker. They can optimize. They don’t just have to accept what the big tech firms serve up.
US tech stocks face pressure from Chinese AI competitors
The stock market doesn’t care about open weights. It cares about margins.
And Chinese alternatives are cheap. Efficient. Getting better fast.
Shares of major US tech firms have dipped since Kimi K3 dropped. Why? Because if cheaper models do the job, US pricing power weakens. Even with export restrictions limiting China’s access to top-tier chips, the gap is narrowing.
China is proving it can compete.
Not with brute force alone. With efficiency. With open access. With a workforce that ships fast.
What other Chinese AI models are releasing now
Kimi K3 isn’t alone.
Over the weekend, Alibaba previewed Qwen3.8 Max. 2.4 trillion parameters. They claim it ranks second only to Anthropic’s Fable 5 among frontier systems.
Second? To a model that may or may not exist in public benchmarks yet?
Alibaba said it. No independent verification. Yet.
Then there’s Zhipu, aka Z.ai. They released GLM-5.2 last month. It’s gaining swift uptake internationally. The trend is clear: Chinese labs are flooding the market.
DeepSeek V4 started it. Early 2025. The market shock. Many saw China as a serious rival then. Kimi K3 just proves they’re not slowing down.
The broader US vs China AI rivalry
This isn’t just about Kimi K3. It’s about the architecture of the future AI economy.
The US controls the chips. The supply chains. The foundational layers.
China controls the scale. The speed. The willingness to open-source.
The US has export restrictions. Hard limits. They block access to the most advanced hardware. But innovation finds a way. Or builds its own way.
Moonshot running out of compute is a bottleneck. But it’s also a signal. Demand is real. It’s global. And it’s exceeding what Western-centric infrastructure assumed it could handle.
A crowded field. More players. Faster releases. Lower costs.
Who wins?
Maybe it’s not one winner. Maybe it’s fragmentation. Different models for different needs. Open source for developers. Closed APIs for enterprise.
The stampede is happening. The servers are sweating.
And Moonshot? They’re just trying to keep the lights on for everyone else.
You can wait. Or you can look at the alternatives. They’re multiplying. Fast.
