US requires licences for Nvidia H20 exports to China

Background

The H20 went on sale at the end of 2023 as an AI accelerator built specifically for the Chinese market^[1]. Before the restrictions, it was the most advanced Nvidia chip that Chinese buyers could legally buy for training and inference^[1]. It was not a top-end part: compared with the H100, its performance was heavily cut down^[1]. Even so, demand was large. The market research firm Omdia estimated that ByteDance and Tencent each ordered about 230,000 Nvidia AI accelerators in 2024, mostly H20s, which put them second and third among buyers worldwide, behind Microsoft at 485,000^[1]. Sina Finance reported that the H20 generated an estimated US$12 billion to US$15 billion of revenue in 2024^[2].

What happened

Nvidia disclosed on 15 April 2025 that the US government had told it on 9 April that exports of the H20 to China, including Hong Kong and Macao, and to D:5 arms-embargoed countries, would require a licence^[1][2]. On 14 April the government told the company that the requirement would remain in effect for the indefinite future^[2].

The control was drawn around capability rather than a single product name. The licence requirement also covered any integrated circuit that reached the H20's memory bandwidth, its interconnect bandwidth, or a combination of the two^[1]. A different chip that moved data at similar rates would therefore fall under the same rule.

The cost

Nvidia said it expected to record about US$5.5 billion in its fiscal first quarter, which ended on 27 April 2025, for H20 inventory, purchase commitments and related reserves^[1][2]. Its shares fell more than 6% in after-hours trading following the disclosure^[1].

Why it matters

An export licence is not, by itself, a ban. It means a shipment cannot go ahead until the government approves it; where approvals are not granted, the practical effect is that the product stops moving. For Nvidia, the measure turned a product line designed for one national market into goods the company may not be able to sell. The size of the charge — about US$5.5 billion, booked against inventory, purchase commitments and related reserves — reflects how much stock and forward ordering had built up around the part^[1][2].

The revenue figures show the stake. The H20 was estimated to have brought in US$12 billion to US$15 billion in 2024^[2], and its two largest known customers, ByteDance and Tencent, were among the biggest buyers of Nvidia accelerators anywhere, ranked only behind Microsoft^[1]. That made the chip a rare case of a leading AI supplier depending on a single export destination for a significant block of business.

The design of the rule also matters beyond this one product. Because the thresholds were written in terms of memory bandwidth and interconnect bandwidth rather than a model number, a future part aimed at the same customers would be measured against the same limits^[1]. The H20's own history shows the same logic in reverse: it was a China-specific design, and it was the most capable Nvidia chip Chinese buyers could legally obtain, up to the point when it needed a licence^[1].

For the buyers, the immediate consequence was uncertainty about a supply line they had ordered in six-figure volumes^[1]. For Nvidia, the quarter that ended on 27 April 2025 carried a charge roughly a third the size of the revenue the chip was estimated to have produced in a good year^[2] — a measure of how quickly a legally sold product can become a liability on the balance sheet.