US Closes Remote-Access Loophole in AI Chip Export Controls
Trump administration requires licenses for data centers providing compute to restricted countries, targeting workarounds like Megaspeed's overseas cloud operations.
Last verified:
Remote-Access Loophole Closes Under Trump Administration
The Trump administration is moving to plug a significant gap in US AI chip export enforcement by targeting overseas data centers that have become conduits for restricted countries to access advanced computing power. According to TechCrunch, the Commerce Department’s new interim rule requires exporters and operators to obtain licenses before providing compute capacity to any entity connected to countries in the Department of Commerce’s D:5 embargo group—China, Russia, Iran, North Korea, and 18 additional nations.
The control mechanism is straightforward but expansive: any data center or network infrastructure exceeding 12 million TPP (total processing performance) that could theoretically serve a restricted-country customer now requires a license. According to TechCrunch, this threshold approximates the maximum compute density of roughly 750 Nvidia H100 or H200 processors. Critically, the rule extends to indirect access through intermediaries and brokers, forcing operators to monitor customer ownership structures and activities for any connection to embargoed entities.
Why Existing Controls Were Insufficient
The Biden administration had already prohibited direct chip exports to companies domiciled in restricted countries, but this oversight created an arbitrage opportunity: middleman entities and cloud providers could legally purchase chips abroad and resell compute access to those same restricted buyers. According to TechCrunch, Chinese AI companies have been rushing to secure overseas computing capacity precisely to exploit this gap, while intermediaries like Megaspeed allegedly accelerated the trend by purchasing thousands of Nvidia chips for redistribution.
The timing coincides with upheaval at Nvidia itself. According to TechCrunch, Jensen Huang, Nvidia’s co-founder, resigned following disclosure that a Chinese intelligence operative was linked to one of the company’s customers—a development that reportedly catalyzed the government’s push to accelerate enforcement against remote-access schemes.
Scope and Exceptions
The rule simultaneously broadens and narrows enforcement depending on jurisdiction. According to TechCrunch, Malaysia, Singapore, Indonesia, Thailand, and Brazil face new licensing requirements due to “higher-than-normal risk of diversion,” though chips “destined” for US-based companies remain exempt even in those countries. Notably, the rule excludes commercial recommendation systems (e-commerce, social media), which operate at compute scales well below the threshold.
Existing data center operators have 90 days from the rule’s effective date to comply, while future operators must obtain licenses prospectively. According to TechCrunch, the interim rule justifies the expansion as “a logical extension of existing controls” designed to prevent restricted entities from acquiring advanced semiconductors for training sophisticated AI models.
Why This Matters
This enforcement action signals a shift toward choke-point control rather than border control. By targeting infrastructure rather than chip manufacturing alone, the US is attempting to collapse the economic viability of overseas intermediation—the strategy that rendered the first generation of export restrictions leaky. For data center operators in neutral or allied nations, compliance costs and customer-vetting burdens will rise substantially. For Chinese AI companies, the calculus for overseas cloud procurement grows riskier; legitimate access to frontier compute now hinges on demonstrating separation from any ownership link to embargoed entities, a standard many cannot plausibly meet. The 12M TPP threshold also creates a floor below which small-scale distributed inference or fine-tuning operations may still operate in gray zones—suggesting the next iteration of workarounds will target fragmentation across sub-threshold pools.
Frequently Asked Questions
What is the 12 million TPP threshold, and how does it translate to actual hardware?
According to TechCrunch, 12 million TPP (total processing performance) roughly corresponds to the maximum compute capacity of approximately 750 Nvidia H100 or H200 GPUs. This metric is the Commerce Department's standard for measuring computing power under export-control rules.
Does this rule affect cloud services used by US companies?
No. According to the interim rule, the licensing requirement does not apply to chips destined for companies based in the US, even if they operate data centers in Malaysia, Singapore, Indonesia, Thailand, or Brazil.
What is Megaspeed, and why was it mentioned in the context of these controls?
According to TechCrunch, Megaspeed was allegedly purchasing thousands of Nvidia chips to provide compute access to Chinese businesses, circumventing existing export restrictions. The US government reportedly seeks to place Megaspeed on a special entity list, and the scandal involving Nvidia co-founder Jensen Huang's resignation was linked to a Chinese spy connected to one of Megaspeed's customers.
Are existing data center operators required to comply immediately?
No. According to TechCrunch, existing operators have a 90-day grace period from the rule's effective date to obtain the required licenses.