AI
Moonshot's Kimi K3 hits GPU wall: maker suspends new signups within 48 hours
The Chinese AI startup discovered that releasing a high-performing model is only part of the challenge—keeping its inference infrastructure stable proved harder when demand exceeded available capacity.
The New Stack
Moonshot AI became the latest company in the sector to face a hard lesson: launching a successful model depends critically on having the infrastructure to run it. Within two days of releasing Kimi K3, the startup shuttered new subscription registrations after user demand consumed virtually all available GPU resources. Existing subscribers retain access while the company races to expand its hardware footprint before reopening the sign-up gates in phases.
Why inference suddenly became the bottleneck#
The crunch underscores a structural tension in modern AI: building powerful models is easier than provisioning enough compute to serve them at scale. As models handle longer, more complex coding tasks and autonomous agent workflows—which generate substantially more tokens than typical chat interactions—operators discover they need far more processing power than earlier projections suggested.
Moonshot confirmed the surge in a public statement: the 2.8-trillion-parameter K3 model experienced demand that "pushed close to the limits of our current capacity" in its opening 48 hours. The company acknowledged it would expand infrastructure and restore signups incrementally rather than all at once.
A Citigroup analyst noted that agentic workflows differ fundamentally from one-off queries. Tasks where agents continuously create, consume, and transform token streams for extended periods monopolize GPU memory far longer than a chatbot exchange, making it nearly impossible to maintain responsive latency as user bases grow. The bottleneck, analysts argue, has shifted from raw compute to sustained memory bandwidth.
China's supply chains add extra pressure#
For Moonshot and peer Chinese AI startups, this universal challenge carries a regional dimension. US export rules have throttled access to Nvidia's cutting-edge accelerators, forcing companies to piece together their computing infrastructure from older-generation chips and domestic alternatives. Rather than owning sprawling data centers outright, most rent capacity from cloud platforms like Alibaba, Tencent, and Huawei—giving them less direct control over expansion.
The scarcity has prompted a data-center arms race within China. Alibaba has earmarked over $53 billion for AI and cloud infrastructure across three years, while ByteDance is weighing investments approaching $70 billion for AI compute facilities alone.
