🧠 Food for thought

Implications, context, and why it matters.

A 75,000-chip limit barely matches China’s AI appetite

• A proposed limit of 75,000 H200 chips per Chinese customer looks small next to reports that Chinese tech firms have ordered more than 2 million H200s 1.
• Nvidia reportedly has 700,000 H200s in inventory, which leaves demand far ahead of supply even without new limits 1.
• U.S. proposals also include a ceiling that would keep total H200 shipments to China at 50% of the volume sold to American customers 1.
• China could still receive a large number of chips, though it remains unclear whether Beijing formally approved an initial batch of 400,000 H200s despite some reporting that described it as approval 2.

The limit speeds up a separate AI ecosystem in China

• Restrictions on H200 access add pressure to a ‘twin-track’ approach, where China buys some imported chips while speeding up domestic alternatives to support its self-reliance goals 2.
• Regulators are reportedly weighing rules that would require H200 importers to buy a set amount of local chips from companies such as Huawei 3.
• This splits workloads, with Nvidia chips used for top-end model training while more inference work and government use shifts to domestic silicon 4.
• State-controlled cloud providers such as Baidu and Huawei are building vertically integrated systems that favor domestic supply security over access to the newest Nvidia hardware 5.