Kimi K3, the new flagship model from Chinese AI startup Moonshot AI, just ran out of room. Less than a week after launching on 16 July 2026, the company quietly stopped accepting new subscribers because demand had overwhelmed its available computing capacity. Existing users can still access the service, but if you were thinking about signing up after seeing the benchmarks, you’re now on the outside looking in.
And judging by the reaction online, a lot of people were thinking exactly that.
According to Tech Startups, Kimi K3 attracted enormous attention for its performance on coding and long-document tasks – and its price. The API launched at $3 per million input tokens and $15 per million output tokens, which puts it in the same ballpark as Claude Sonnet 5, but well below Claude Opus 4.8 at $5/$25. Independent benchmarks from Artificial Analysis scored K3 marginally ahead of Opus 4.8 on their intelligence index.
The result was a wave of developers and power users considering, or making the switch. The demand spike hit faster than Moonshot’s infrastructure could handle.
The deeper problem, though, isn’t server capacity. It’s chips. As Tech Startups reports, Chinese AI labs can’t simply scale up by buying more Nvidia hardware. US export controls have made the most advanced accelerators hard to source. So when demand spikes, there’s no quick fix. You can’t just throw more compute at it the way a US-based lab could.
I’ve been using Claude Code daily for coding, so I have a horse in this race. When Kimi K3 launched with benchmark numbers sitting above Opus 4.8 and pricing that undercuts it, I noticed and I wasn’t the only one.
The “fraction of the cost” angle that spread across forums is worth unpacking, though. K3 isn’t the bargain it first appears. As the pricing breakdowns show, K3 always runs in reasoning mode, meaning every query generates reasoning tokens billed at output rates. That $15/M output price can quietly add up in ways the headline sticker doesn’t suggest. It’s not cheap; it’s competitive. There’s a difference.
That said, I was definitely thinking about trying it. My Claude Pro subscription rarely hits limits, so the cost argument doesn’t hit me the way it does someone running heavy API workloads. But if I were billing tokens at volume, I’d absolutely be paying attention to this.
Moonshot built a model impressive enough to trigger a rush of users and then immediately couldn’t serve them, partly because of geopolitical factors outside their control. That’s a structural disadvantage that doesn’t disappear when you fix your server count. Every major scaling moment is going to run into the same chip ceiling.
The irony is that the capacity crunch might actually be the best marketing Kimi K3 could ask for. Nothing creates demand like being told you can’t have something.
Moonshot has said it plans to fully open-source the K3 weights by 27 July 2026. If that lands on time, developers will be able to self-host which sidesteps both the subscription wall and the data sovereignty questions that come with routing sensitive queries through a Beijing-based company’s servers. That’s when the real test begins: whether K3 holds up when independent teams start stress-testing it properly, without Moonshot controlling the evaluation environment.
