Moonshot Kimi K4 Scale Expected to Exceed K3 Amid NVIDIA Chip Supply Constraints

Moonshot plans larger Kimi K4 model but faces delays due to NVIDIA Blackwell chip shortages and US export bans. Current K3 inference relies on H20s, causing subscription pauses.

Woofun AI reports that Moonshot AI is developing the Kimi K4 model, with parameter scale expected to significantly surpass the 2.8 trillion-parameter K3.

However, development timelines remain undefined as the company seeks additional NVIDIA Blackwell chips to meet computing demands. Data compiled by Woofun AI shows that portions of K3 training occurred in China, utilizing Blackwell chips sourced from at least two Chinese cloud providers. To compensate for insufficient single-provider capacity, Moonshot integrated multiple eight-card servers and optimized cross-data-center communication. US export restrictions on Blackwell chips to China and ongoing investigations into procurement channels complicate future scaling.

Meanwhile, K3 inference operations face H20 chip shortages, forcing a suspension of new subscriptions within 48 hours of demand spikes. Domestic chip alternatives remain inadequate due to limited server availability and six-month lead times.