China's Kimi K3 'erased' America's AI lead; but has same problem as ChatGPT & Claude

China's Moonshot AI spent the past week being the most talked-about name in artificial intelligence. Its new open model, Kimi K3, beat Anthropic's Fable 5 in front-end coding tests, rattled the Nasdaq, and had Silicon Valley executives arguing about whether China had just erased America's lead overnight. For a Beijing startup backed by Alibaba and Tencent, it was the kind of moment most AI labs spend years chasing.Then reality arrived in the form of a bill nobody could pay fast enough—compute. In a post on X, the company admitted its servers simply couldn't keep up. "Kimi K3 has received far more love than we expected, and our GPUs are feeling it," Moonshot wrote, announcing it was hitting pause on new sign-ups. The irony wasn't lost on anyone: the model built to challenge ChatGPT and Claude had just run headfirst into the exact same problem those American models have been wrestling with for years.Kimi K3's viral demand outran Moonshot AI's GPU capacity in 48 hoursThe numbers moved quickly. Within 48 hours of launch, demand pushed Moonshot's infrastructure "close to the limits of our current capacity," the company said. Its fix was blunt—freeze new subscriptions, protect existing subscribers, and buy time. Current paying users won't be affected, and Moonshot promised to reopen spots "in batches" as it adds capacity.This is familiar territory. OpenAI paused ChatGPT Plus sign-ups in late 2023 when a product launch overwhelmed its systems. Anthropic has rolled out usage caps on Claude for the same reason. Serving a frontier model at scale is expensive and hardware-constrained, and Kimi K3 is heavier than most—a 2.8-trillion-parameter system that Moonshot recommends running on setups with 64 or more accelerators. Popularity, for an AI company, is a cost centre before it's a revenue line.Why Moonshot is splitting Kimi memberships into web and coding plansMoonshot is also using the crunch to restructure how it sells access. Going forward, membership splits into two tracks—Kimi Membership covering the web app, mobile app and Kimi Work, and a separate Kimi Code Membership aimed at coding workflows. The logic is operational rather than cosmetic. Coding sessions are compute-hungry and run long, so separating them lets Moonshot match GPUs to demand instead of spreading capacity thin across every use case.Kimi K3 open-weight release on July 27 could ease the subscription crunchThe bottleneck may be short-lived. Moonshot plans to release Kimi K3's full model weights by July 27, letting companies and governments download, customise and run the model on their own hardware. That open-weight approach is the whole pitch—and it doubles as a pressure valve, since heavy users can self-host instead of leaning on Moonshot's strained API.For now, though, the takeaway is grounded. Kimi K3 proved a Chinese lab can build near-frontier AI at roughly 40% less cost than its US rivals. What this week also proved is that closing the capability gap doesn't close the compute gap. Whether you're in Beijing or San Francisco, the GPUs run out the same way—and the bill comes due just as fast.Get the latest technology news and updates. Download the TOI App.