Nvidia's H20 charge makes export policy a line item

Washington now requires licenses for Nvidia's H20 and AMD's MI308 in China. Up to $6.3 billion in charges show that policy, not demand, can reprice AI compute.
Up to about $5.5 billion of charges for inventory, purchase commitments and reserves will land in Nvidia's quarter ending 27 April, because Washington now requires an export license for its H20 chips to China. The requirement, disclosed on 15 April, covers Hong Kong and Macau and any chip with comparable memory or interconnect bandwidth, and on 14 April the government told Nvidia it would stand for the indefinite future.
Export policy has become a line item in semiconductor earnings. The H20 was the China product that the export rules first set in 2022 had left unrestricted, a chip built to the letter of those rules, and the rules moved under it within a week. AMD followed the next day with up to about $800 million for its MI308. Any return that depends on selling compute across borders now carries policy as a recurring risk, not a one-off event.
The exposure is large. China, including Hong Kong, billed $17.1 billion of Nvidia's $130.5 billion revenue in its last fiscal year, and a market that accounted for 13 percent of revenue can no longer be modeled as a reliable customer of US silicon. The charges also show where demand goes when a market closes. Chinese buyers turn to domestic accelerators, which hands local chip makers a protected customer base and a reason to invest, while demand from US data centers absorbs the supply for now.
Compute supply is now a policy question as well as a price question for every company that rents it. Teams serving customers in many countries need to know which chips their models run on and where, because a license rule can change hardware availability in a region overnight. Products that run efficiently on several accelerators, from Nvidia and AMD to the clouds' own chips, carry less risk than products tuned to one. Compliance opens a market too: tools that record where models are trained and served, and on which hardware, gain buyers as the AI diffusion rule approaches its 15 May compliance date.
Strong US demand makes it tempting to read all this as one bad quarter for one supplier. The diffusion rule tests that reading, since whether it takes effect on 15 May as written or is replaced decides how advanced chips reach dozens of other countries. Nvidia's results for the quarter ending 27 April will give the final charge and guidance on China, and with them how much demand has moved elsewhere. The clearest measure of how far export rules slow China at all is whether its labs keep releasing competitive models trained and served on domestic hardware.
AI compute is a strategic good, and its price now includes policy. The resilience Fanaura's thesis asks for means economics that survive a change in which chips a company can buy and where. A business that depends on one supplier and one open border fails that test.
Building an AI-native company? Apply in five minutes. →
More from Fanaura
Building AI products that run efficiently across several chip platforms? Send us the project.
Fanaura invests its own capital and the capital of its limited partners (LPs), usually through a separate SPV for each deal. If we see a fit, we get in touch.
Get the news by email
One short note each weekday. No marketing, and you can unsubscribe at any time.