AI Export Controls: What the Next Round Means for Chinese Buyers and Their Partners
Each round of AI chip and model export controls has been tighter than the last. The next round is coming. Here is what I am telling companies on both sides of the Pacific to prepare for — and the compliance gaps most are missing.
The trajectory is clear — and it is not flattening
I have watched the U.S. AI export control regime evolve from the initial semiconductor restrictions of 2022 through the Entity List additions, the advanced computing chip rules, the cloud computing reporting requirements, and now the emerging model-weight controls. Each round has been broader, tighter, and more technically specific than the last. Anyone who tells you the next administration will relax these does not understand the bipartisan consensus behind them. AI controls are not a partisan issue in Washington — they are one of the few areas where Congress and the executive branch agree across the aisle. I tell clients to plan for tightening, not loosening, regardless of who occupies the White House.
The gap most companies miss: it is not just about chips
When companies think about AI export controls, they think about chips — NVIDIA GPUs, AMD accelerators, the latest fabrication-node restrictions. That is the visible layer. The layer that catches companies off guard is the software and model layer. The emerging controls on AI model weights, on fine-tuning services, on cloud computing access to frontier models — these create obligations that extend far beyond the hardware buyer. If you are a Chinese company using a U.S. cloud provider to train models, if you are a Western company providing AI services to Chinese clients, if you are a multinational with an AI team in China — you have exposure that the chip-focused compliance programs are not catching. I sit with clients and walk through their entire AI stack, not just the hardware procurement. That is where the real compliance gaps live.
What Chinese buyers should be doing now
For Chinese companies that rely on U.S. AI technology — directly or through cloud — the planning horizon has shortened. I advise three things. First, inventory your AI dependencies: which models, which cloud services, which chips, which software libraries. Know exactly what you would lose if access were cut off tomorrow. Second, build redundancy for the most critical dependencies — domestic alternatives where they exist, open-source equivalents where they do not, and qualified non-U.S. alternatives where neither is available. Third, document your compliance posture thoroughly. If you are operating within the current rules, be able to prove it. The companies that get hurt worst in a tightening cycle are not the ones with the most exposure — they are the ones with the least documentation that they were operating in good faith.
What Western companies should be doing now
For U.S. and allied companies providing AI products or services, the compliance obligation runs the other direction. You need to know who your end users are, what they are doing with your technology, and whether your engagement falls within the current rules. I am seeing too many companies that built AI service relationships with Chinese clients in 2023 and 2024 and have not re-screened those relationships against the current control framework. The rules have changed; your screening needs to change with them. End-use screening is not a one-time exercise — it is a continuous obligation, and the enforcement environment is getting stricter, not more forgiving.
Frequently Asked Questions
Will the AI export controls relax after the next election?
I do not believe so. The bipartisan consensus behind AI controls is strong and durable. Companies should plan for continued tightening. Any relaxation would be marginal and tactical, not a fundamental reversal of the trajectory.
Are open-source AI models exempt from these controls?
Not necessarily. The control framework is evolving to address model weights and training services, not just hardware. Open-source does not automatically mean exempt — it depends on the specific model, the specific control, and the specific end use. Do not assume open-source is a safe harbor without specific legal analysis.
Need advisory on Export Controls & Sanctions Compliance?
Brian B. Su works directly with decision-makers.