On April 15, 2025, Nvidia said in a regulatory filing that the U.S. government now requires a license to ship its H20 artificial intelligence accelerator to China. The company told investors it expected to book a $5.5 billion charge against inventory and purchase commitments tied to the H20. That filing closed, temporarily, the largest remaining route for advanced American AI silicon into Chinese data centers. The export-control regime behind it, the US-China AI chip export restrictions, has been building since October 2022.
The H20 was not a flagship part. It was built to fit under the October 2023 U.S. rules that banned Nvidia's A800 and H800. Chinese cloud operators bought H20 clusters anyway, because the chip kept enough high-bandwidth memory to run inference workloads on large language models. Losing it was not a chip industry abstraction; it hit purchase orders, server designs, and licensing decisions within weeks.
October 2022: The First Cut
On October 7, 2022, the Commerce Department's Bureau of Industry and Security published an interim final rule restricting advanced computing integrated circuits bound for China. The rule captured Nvidia's A100 and H100 plus AMD's MI200-series accelerators by targeting chips that crossed a combined compute and interconnect-bandwidth threshold. Nvidia responded with the A800 and H800, which reduced NVLink bandwidth to fall outside the rule. Those parts became the new China baseline for roughly a year.
The October 17, 2023 update scrapped the old test. BIS replaced it with total processing performance and performance density limits, which caught the A800 and H800. The same package expanded semiconductor manufacturing equipment controls aimed at China's advanced fabs. From that point, the Chinese supply chain migrated to H20, L20, and L2 parts, with H20 emerging as the volume product for major AI deployments.
The 2022 rule also added controls on advanced chipmaking tools. That second route mattered because cutting off chips alone still left Chinese fabs a path to build their own; restricting the equipment slowed that path as well. The changed threshold also ended the assumption that a part designed to sit just below the line would remain legal. Nvidia learned that when the line moved in 2023.
April 2025: The H20 License Requirement
By early 2025, H20 systems were the default choice for Chinese cloud providers that could no longer buy higher-end U.S. parts. H20 sacrificed raw compute relative to the H100 but retained enough memory bandwidth and connectivity to serve inference and some training workloads. The April license requirement cut that thread. Nvidia's $5.5 billion charge was spelled out in the filing, and Reuters reported that the Commerce Department had informed the company of the requirement. Chinese firms had already spent months validating H20 clusters because the chip's appeal was not its speed but its compatibility with Nvidia's software stack.
A hardware engineer at a Massachusetts AI-infrastructure startup — call her Dr. K — had committed to an H20-based cluster design for three partner deployments in China. The announcement froze her firmware validation schedule and turned a 12-week procurement plan into an open-ended licensing queue. That is what a $5.5 billion charge looks like from the inside.
How Washington Justifies the Controls, and How Beijing Responds
The Bureau of Industry and Security says the restrictions prevent advanced chips from strengthening Chinese military systems, including weapons development, surveillance, and cyber operations. China's Ministry of Commerce has called the measures an abuse of export controls that disrupts global supply chains. Both positions have concrete follow-through. In July 2023, Beijing restricted gallium and germanium exports, adding antimony and graphite later. In December 2024, China's market regulator opened an antitrust probe into Nvidia.
China's domestic alternatives have become part of its policy answer. Huawei's Ascend 910B and 910C accelerators now serve some Chinese cloud and state-backed projects. They do not yet match Nvidia's CUDA software ecosystem, and manufacturing constraints at SMIC remain significant, but they are the clearest domestic line of defense.
The controls are not limited to Chinese buyers. An interim final rule published in January 2025 created country tiers for AI chip exports, adding caps and authorization requirements for most non-ally countries while keeping the strictest restrictions on China, Macau, and arms-embargoed destinations. The shift has forced cloud providers and data center operators to think about GPU location as a security property. Washington frames the risk as a package: silicon, software, networking, and systems knowledge.
What This Means for Chip Buyers and Lab Operations
For any organization that buys GPUs, AI accelerators, or cloud capacity across borders, the compliance burden now sits at the purchase-order stage. The rules do not only bind the exporter. U.S. controls extend to reexports, foreign-made products that contain U.S. technology, and shipments routed through third countries.
- Check each accelerator's Export Control Classification Number before ordering; a part that cleared review last year may not clear this year.
- Do not assume a China-market variant is automatically compliant after the next BIS rule change.
- Build an export-control clause into supplier contracts and purchase orders, with clear duty to provide license documentation.
- Keep a current inventory of installed accelerators and their ECCNs because a license can become retroactively required.
- Monitor Federal Register notices and BIS press releases; the H20 change arrived with little public warning.
The same discipline applies to labs that rent AI compute. A lab that trains models on a cloud cluster should be able to say where the physical GPUs sit, which entity operates them, and whether any restricted party can access them. Some cloud contracts now include termination rights if export licenses are denied. Dr. K's experience is a reminder that procurement is not a back-office task; it is part of the experiment design.
Photo by KOBU Agency on Unsplash
One concrete next step for any chip buyer is to pull the current BIS Entity List and check each part's Export Control Classification Number before the next order goes out. The list changes, and the H20 filing showed that even a product designed to comply can fall under a new license requirement without a grandfather clause. BIS has not signaled a public timetable for next revisions, but the April 2025 H20 move makes one lesson plain: compliance is a standing requirement, not a one-time review.