U.S. Restricts Nvidia’s H20, Threatening $5.5 Billion

The United States has tightened its restrictions on artificial intelligence chips destined for China once again. Nvidia will need an export license to sell its H20 processor there, a chip specifically designed to comply with earlier U.S. restrictions. The move puts a significant business line at risk and points to a charge of up to approximately $5.5 billion.

Nvidia said the U.S. government notified it of the new requirement on April 9. The rule applies to H20 exports to China—including Hong Kong and Macao—as well as to other destinations subject to equivalent controls. The licensing requirement will remain in effect indefinitely.

A Chip Built for the Chinese Market

The H20 is not one of Nvidia’s most powerful accelerators. It is a version adapted for China after Washington expanded export controls on advanced processors in October 2023. The earlier controls already covered products such as the A100 and H100, while the 2023 expansion also blocked variants such as the A800 and H800, which had been designed to comply with the previous restrictions.

The idea was to leave room for a lower-capability chip that could still be sold legally. The H20 emerged from that balancing act: it offers less computing power than Nvidia’s flagship chips but is useful for inference, the stage at which an already-trained model answers questions, generates text or analyzes data. It is a core task in deploying AI assistants and services at scale.

That path has just narrowed. The Commerce Department has justified the licensing requirement on the grounds that the H20 could be used in, or diverted to, supercomputers in China. The United States is seeking to prevent the computing capacity needed to develop advanced AI from strengthening military or strategic applications in the Asian country.

An Immediate Impact on the Bottom Line

The charge Nvidia expects is not a fine. It is an accounting adjustment tied to H20 inventory, purchase commitments and reserves related to products the company had expected to sell in the Chinese market. Nvidia would record it in the first quarter of its 2026 fiscal year, which ends April 27.

The figure shows how commercially important the H20 had become. Although Nvidia sells its processors to server manufacturers and technology companies, China remains a critical market for AI infrastructure. The country’s major internet and cloud companies need chips to meet demand for generative models, recommendation systems and enterprise services.

A license is not, on paper, a total ban: it allows authorities to review individual transactions. But it introduces uncertainty that will be difficult for Nvidia, its customers and the manufacturers integrating these chips into servers to manage. Orders could be put on hold while approvals are processed, and distributors will have to reassess which inventory they can receive and deliver.

Competition Extends Beyond Models

The H20 case shows that the rivalry between the United States and China is also being fought over components that may appear highly technical. An AI model needs data, talent, electricity and software, but its scale depends on the accelerators that perform enormous numbers of mathematical operations. Controlling those chips means controlling a significant part of the capacity to create and deploy advanced AI.

For Nvidia, the problem is twofold. The company is losing predictability in China and must prevent products intended for other markets from being re-exported to restricted destinations. For Chinese companies, the new barrier increases the incentive to turn to local suppliers, although matching Nvidia’s performance, energy efficiency and software ecosystem will not happen overnight.

The key question will be how Washington handles the licenses and whether the H20 retains any commercial path into China. For now, the chip that was supposed to keep a regulatory door open has become proof that the door can close even after the product is redesigned.