Tech • IA • Crypto
Kimi K4 Is Bigger Than Anyone Expected
9.2/10TL;DR
Chinese AI firms are scaling massive models using restricted Nvidia chips, exposing gaps in US export controls and intensifying global competition over AI infrastructure.
KEY POINTS
Moonshot AI plans larger K4 model
Moonshot AI is preparing a successor to its recently released Kimi K3, already one of the largest open-weight models at 2.8 trillion parameters. Sources indicate the upcoming K4 will be significantly larger, but its development hinges on securing more advanced Nvidia chips, particularly from the Blackwell line.
Use of restricted Nvidia hardware
Multiple accounts indicate K3 was trained using Nvidia Blackwell GPUs, despite US export restrictions barring Chinese firms from accessing such hardware. This aligns with statements from US officials, reinforcing concerns that enforcement mechanisms are being bypassed.
Conflicting narratives on training location
US officials have suggested K3 was trained in Thailand, where access to certain chips may be permissible. However, other sources say portions of training occurred داخل China, citing strict Chinese data transfer laws that make exporting massive datasets difficult. This creates a regulatory clash between US export rules and Chinese data policies.
Fragmented infrastructure and engineering workaround
Due to limited chip availability, Moonshot reportedly assembled training clusters from multiple Chinese cloud providers, each contributing smaller eight-GPU Blackwell servers. Engineers then optimized networking across providers to enable distributed training at scale, compensating for the lack of large, unified clusters typical in US labs.
Broader pattern across Chinese AI labs
Similar patterns appear across the industry. Alibaba’s Qwen 3.8 Max and DeepSeek V4 were also reportedly trained using Nvidia hardware, including Blackwell chips. This suggests widespread reliance on US technology despite geopolitical tensions and domestic investment in alternatives.
Dependence on Nvidia for training
While China is advancing in inference hardware, training remains heavily dependent on Nvidia. Researchers cite stability and high-speed interconnects as critical advantages not yet matched by domestic chips, making large-scale training extremely difficult without US technology.
Inference bottlenecks and policy tensions
For deployment, Moonshot relies on Nvidia H20 chips, which are permitted under US rules. Running K3 requires at least 64 server-grade GPUs, straining supply. China is discouraging use of H20 in favor of domestic chips, but local alternatives face up to six-month delays, leaving companies caught between policy and practicality.
Workarounds via overseas compute
Chinese firms are increasingly renting compute abroad. For example, Tencent has used data centers in Japan equipped with thousands of Nvidia GPUs. These arrangements remain legal for now, though US regulators are considering tighter controls on remote access to advanced chips.
US debate over Chinese open models
The rise of powerful Chinese open-weight models has triggered division in the US. Some policymakers warn of security risks and accuse firms of training via distillation from US models. Meanwhile, industry leaders including Nvidia support continued access, arguing that both open and closed AI ecosystems are essential.
Export controls under scrutiny
The growing evidence that Chinese firms continue to access restricted hardware has raised questions about the effectiveness of US export controls. Investigations are underway into how companies obtain or access advanced chips despite formal restrictions.
CONCLUSION
China’s rapid progress in large-scale AI models underscores both its engineering ingenuity and continued reliance on US hardware, highlighting a fragile and contested global AI supply chain.