Nvidia just fired up H200 production for China again, which means Beijing blinked first in the AI cold war.

The Summary

  • Nvidia is restarting H200 chip manufacturing for Chinese customers, CEO Jensen Huang confirmed at a Tuesday press conference, signaling Chinese government approval of purchase orders.
  • This reversal comes after months of export controls that turned AI compute into a geopolitical chess piece.
  • China choosing to buy from Nvidia instead of betting entirely on domestic alternatives tells you everything about where homegrown Chinese AI chips actually stand.

The Signal

The H200 restart isn't just a sales win for Nvidia. It's a data point on how the AI arms race actually works when theory meets manufacturing reality. The H200 is Nvidia's previous-generation flagship, a step down from the H100 and miles below the newer Blackwell architecture. But it's still leagues ahead of anything China can produce at scale domestically.

Huang's announcement implies Beijing approved these purchase orders, which means someone in the Chinese government did the math. They looked at SMIC's 7nm process struggles, at the gap between Huawei's Ascend chips and what DeepSeek or ByteDance actually need to compete globally, and decided that buying gimped American chips beats waiting for domestic alternatives that might not arrive in time.

This matters for the agent economy because China isn't sitting out AI development. They're fielding competitive models, they're building applications, and they're racing toward AGI just like everyone else. If they're willing to swallow the geopolitical cost of buying from Nvidia again, it means the compute gap is real and widening. The companies building agents, the ones training foundation models, the infrastructure layer of Web4, they all run on these chips. China paying Nvidia means they couldn't wait.

For Nvidia, this is revenue and leverage. They get to sell older-generation chips at whatever margin the market will bear, while Blackwell and whatever comes next stay stateside. For China, it's a tactical retreat that buys time but doesn't close the gap. And for everyone watching the AI race, it confirms what the hype obscures: chips still matter more than anything else. You can't prompt-engineer your way around bad hardware.

The Implication

Watch how many H200s actually ship and where they end up. If Chinese AI labs suddenly start announcing new model releases six months from now, you'll know the compute arrived. If export controls tighten again in response, this was a brief window. Either way, this tells you that the real bottleneck in AI development isn't ideas or talent, it's silicon. The companies that control chip supply control the pace of the entire agent economy. Plan accordingly.

Source: The Information