Export control expert Jimmy Goodrich: H20 is the world's best inference chip by cost-per-token and China can't replicate it yet
Aug 15, 2025
Key Points
- Export control expert Jimmy Goodrich argues the H20 is the world's cheapest inference chip by cost-per-token, but export to China poses genuine national security risks beyond economic competition.
- Autonomous AI agents running on H20s could enable Chinese cyber warfare and real-time drone networks at scale, while state-owned firms already deploy DeepSeek for disinformation against Taiwan and the US.
- US export controls leak months in advance, allowing Beijing to stockpile chips and observe American debate before half-measures take effect, signaling Washington lacks strategic focus versus China's multi-decade consistency.
Summary
Jimmy Goodrich, a semiconductor export control expert, argues that the H20 is the world's best inference chip by cost-per-token. He also believes exporting it to China poses genuine national security risks, though not primarily economic ones.
Goodrich sees the H20's value in two dimensions. It excels at inference, where memory bandwidth matters most, and delivers that capability at lower cost than alternatives. Nvidia can supply millions of units, a scale Chinese manufacturers cannot match in the near term due to export controls and the complexity of advanced node fabrication. The installed base of CUDA expertise among Chinese developers trained on Nvidia's ecosystem since college amplifies this advantage. For Beijing, the H20 becomes a near-term force multiplier for AI deployment.
Goodrich separates this from economic competition. He points to two concrete national security concerns. The first is cyber and information warfare. Autonomous AI agents scaled across dozens of inference data centers could conduct vulnerability scanning and offensive cyber operations at new speeds. State-owned Chinese firms are already using DeepSeek, which will run on chips like the H20, to conduct disinformation campaigns against Taiwan and the US, according to New York Times reporting from ten days prior. The second concern is military application. Real-time federated drone networks could collect battlefield communications, transcribe them autonomously, and feed analysis to commanders in real time. That capability scales with inference compute, not training compute.
China's AI industry faces internal tension. Companies like DeepSeek, Moonshot, Baidu, and Tencent want the best chips regardless of origin. But Xi Jinping has set a national self-sufficiency target, and a dozen indigenous GPU suppliers centered around Huawei want to capture market share. Beijing is balancing both impulses. It rolls out the red carpet for Nvidia when Jensen Huang visits but channels state contracts to domestic players. Even senior figures inside China criticize Huawei's dominance, viewing it as too aggressive for a single firm to monopolize the AI stack.
Goodrich sees a broader pattern in US export controls. The process leaks months in advance, giving Beijing time to stockpile before Americans debate openly in public and half-impose restrictions. The result is that Beijing views the US as not serious. The US has military capability but lacks focus across semiconductors, rare earths, and supply chain strategy, while China executes multi-decade plans with consistency.