The House China chair just accused Nvidia (NVDA) of doing the one thing U.S. export controls are supposed to stop: Helping Chinese AI labs get better at using restricted chips.

Representative John Moolenaar, who runs the House Select Committee on the Chinese Communist Party, sent a letter to the Commerce Department alleging Nvidia engineers provided extensive technical support to DeepSeek.

That support helped DeepSeek reach frontier level chatbot performance, despite U.S. limits on advanced AI hardware exports.

Moolenaar says committee documents show Nvidia staff helped DeepSeek improve training efficiency on export controlled H800 chips.

The result was big enough performance gains that DeepSeek models are now reportedly integrated into systems used by the People’s Liberation Army. He also labeled DeepSeek a cybersecurity risk.

Nvidia says this was normal commercial support meant to strengthen the AI ecosystem and improve its own products. Moolenaar says that misses the point. He wants stricter enforcement of the H200 export rule and a Commerce Department briefing by Feb. 13. China also recently blocked Nvidia H200 chips.

The disruption behind the news: Chip bans alone won’t stop China’s AI progress.

This blows a hole straight through the assumption that chip bans alone can slow China’s AI progress.

The core constraint on advanced AI is no longer just raw silicon.

Training tricks, compiler tuning, memory optimization, and model architecture choices now matter as much as owning the very best chip.

Nvidia sits at the center of that stack. When its engineers help a customer squeeze more performance per watt and per dollar, that help compounds fast.

DeepSeek reportedly matched leading U.S. models while using a fraction of the compute budget. Estimates circulating in AI circles put the training cost reduction at 5x to 10x compared with earlier baselines.

That effectively bypasses hardware-based export controls.

Export rules assume a linear relationship between chip capability and AI capability. That assumption is broken. If a Chinese lab can get frontier results from H800 class hardware with vendor level optimization, then the marginal value of blocking H100 or H200 chips collapses.

For Nvidia, this is a business risk. If Washington decides that technical support is functionally equivalent to exporting performance, then software, drivers, and engineering services become regulated assets.

That hits margins, slows deployments, and injects compliance risk into every overseas enterprise deal. Nvidia’s stock price is built on scale and speed. Regulation targets both.

For the U.S. government, the lesson is harsher. You cannot outsource enforcement to intent statements and customer certifications. The letter hints that H200 chips are already being approved for major Chinese firms like Alibaba, Tencent, and ByteDance. If those firms get top tier hardware plus vendor tuning, the gap closes faster than policymakers admit.

What to watch next

First, watch whether Commerce expands controls from hardware to support.

Training assistance, kernel optimization, and model level guidance are the real accelerants now.

Second, watch Nvidia’s behavior.

If it pulls back engineering help in China, performance per chip there will stall. If it does not, expect enforcement actions that spook customers globally.

Third, expect Chinese AI labs to double down on efficiency.

The incentive is clear. Every 10 percent gain in training efficiency weakens export controls more than any single chip ban.

Nvidia’s assistance to DeepSeek underscores a structural change in AI competition. Software optimization and training efficiency now matter as much as access to advanced chips. But U.S. export controls, for now, remain focused primarily on hardware. Nvidia (NVDA) has a Disruption Score of 4. Click here to learn how we calculate the Disruption Score. 

Nvidia is also part of the Disruption Aristocrats, our quarterly list of the world’s top disruptive stocks.