US Closes Nvidia Blackwell Loophole: China HQ Firms Abroad Need Licenses

The US Department of Commerce posted weekend guidance on May 31, 2026 closing a loophole that let China-headquartered firms buy Nvidia Blackwell, Rubin, and AMD MI350x class chips through subsidiaries outside China — without the export licenses Washington already required for shipments into mainland China.

Industry sources quoted by Reuters estimated hundreds of thousands of advanced processors may have flowed through the gap over roughly the past year. BIS said it was clarifying requirements in place since 2023, not inventing new law overnight.

What Changed on May 31?

The Bureau of Industry and Security (BIS) now enforces license requirements based on ultimate purchaser headquarters, not only the shipping address.

Covered processors include Nvidia Blackwell and Rubin families and AMD MI350x, per Reuters and trade press.

Nvidia told reporters the guidance does not change its own licensing obligations — Commerce had already imposed letter requirements on the company. The fix targets who could buy without a license, not Nvidia's existing compliance posture.

Why Washington Moved on a Sunday

A Washington policy paper on the loophole circulated before the guidance dropped, according to people familiar with the process quoted by Reuters and CNA.

The political trigger: reports that China's most capable AI labs were still accessing top-tier US silicon by routing purchases through offshore entities while broader AI Diffusion rules sat in limbo. The Biden-era diffusion framework had global licensing requirements; the Trump administration paused enforcement in May 2025, reopening ambiguity BIS now tried to narrow.

Chris McGuire, a trade lawyer quoted in coverage, called the prior gap a "HUGE problem" but noted another loophole remains: foundries like TSMC may still face uneven due-diligence obligations on whether chips are bound for Chinese front companies.

What Is Not Affected

The guidance does not:

• Force data centers to power down already-deployed clusters
• Cut off servicing of installed advanced servers
• Retroactively void every shipment already in the field

This is forward-looking export enforcement, not a recall order.

Developer and Infrastructure Impact

GPU procurement in Asia just got legally noisier. If you run training clusters for a China-HQ parent with a Singapore SPV, legal and compliance teams need a license path or a non-US chip stack — see DeepSeek V4 on Huawei Chips.

Cloud region planning: Hyperscalers with Malaysia and UAE buildouts marketed to global AI tenants will face sharper KYC and end-user checks on who consumes H100/Blackwell-class capacity.

Price and lead time: Tighter enforcement usually shows up as longer allocation queues and higher spot GPU rents before it shows up in blog posts — pair with LLM API Pricing when modeling inference budgets.

For the wider tariff and control stack, cross-read Trump 145% China Tariffs and Nvidia/TSMC Hardware and China Cheap Energy + SMIC Record Revenue.

Key Takeaways

• May 31, 2026: BIS guidance — China-headquartered buyers need export licenses for Blackwell, Rubin, MI350x even outside China
• Scale: Industry source cited hundreds of thousands of chips potentially exported via the loophole
• Nvidia: says its existing license duties unchanged; rule closes buyer-side gap
• Not retroactive on deployed data centers; foundry due-diligence gap may remain
• For developers: offshore SPV routing no longer a compliance shortcut; expect stricter tenant screening in Gulf and ASEAN regions

Sources

• US takes step to halt Nvidia AI chip shipments to Chinese firms outside China — CNA/Reuters, May 31 2026
• US Commerce Department closes loophole on Nvidia chip exports — CryptoBriefing/Reuters summary