Staffing at the U.S. Commerce Department office that vets exports of Nvidia and AMD AI accelerators has collapsed over the past year, and approval times for chipmakers are now running into months, Bloomberg has reported.
Citing more than 20 people familiar with the situation, it’s understood that the Bureau of Industry and Security is struggling to process a workload expanded by the Trump administration’s tariff probes and AI chip export reviews, while Under Secretary of Commerce Jeffrey Kessler is insisting on personally examining nearly every license application and telling companies to “simply call him to have a license approved.”
A Bloomberg analysis of Office of Personnel Management figures, LinkedIn profile changes, and agency records found that BIS has shed 101 employees, a 19% reduction, since 2024. Turnover among rulemaking and licensing staff specifically has run at nearly 20%, the outlet said.
Latest Videos From
Since late February, top officials at the bureau have been focused on the Iran war, pulling energy away from the technology-export push that shaped the administration's first year, Bloomberg said. The same conflict has pushed Trump's planned meeting with Chinese President Xi Jinping to next month, with AI chip access and rare earth supplies both expected to come up.
Those delays fall on the very products that the Trump administration has spent the past year working to clear for export. Nvidia's licenses to supply Saudi Arabia and the United Arab Emirates, the 25% import-duty framework for H200 shipments to "approved" Chinese customers, and AMD's MI308 approvals all route through the Bureau of Industry and Security. Despite what seemed like good progress, Nvidia hasn’t sold a single H200 to China months after the White House cleared the deal, despite having received orders, and BIS processing delays now appear to be the reason why.
Middle East licensing is adding a further layer of difficulty. The export permissions that Cerebras and Nvidia received last year for shipments to the UAE and Saudi Arabia came with dollar-for-dollar U.S. investment matching requirements, meaning each case must be individually negotiated rather than stamped against a standardized template. But given that the BIS has lost a fifth of its licensing staff and is now handling a much more complex caseload, it’s easy to see why turnaround times ballooned to 76 days in the first half of 2025, well beyond the average 2023 turnaround time of 38 days.
Fiscal 2023 is the most recent year for which the Bureau of Industry and Security has published a formal annual report, and at that point, the agency was processing nearly 38,000 applications a year with an 85% approval rate. Its reports for 2024 and 2025 have yet to be released, leaving chipmakers reliant on anecdotal turnaround data to work out how long their own applications are likely to sit in the queue.
Stay On the Cutting Edge: Get the Tom's Hardware Newsletter
Get Tom's Hardware's best news and in-depth reviews, straight to your inbox.
Luke James is a freelance writer and journalist. Although his background is in legal, he has a personal interest in all things tech, especially hardware and microelectronics, and anything regulatory.
I honestly feel bad for Chinese AI researchers. Both governments are trying to keep them from buying Nvidia but nobody else’s chips have the software ecosystem to be good at training models. Inference is fine on any brand but AI companies are generally more worried about training in this phase of the game.
I honestly feel bad for Chinese AI researchers. Both governments are trying to keep them from buying Nvidia but nobody else’s chips have the software ecosystem to be good at training models. Inference is fine on any brand but AI companies are generally more worried about training in this phase of the game.
Their pain could be everyone else's gain, if it helps break the "CUDA moat".
Their pain could be everyone else's gain, if it helps break the "CUDA moat".
Well that’s true. AMD supporting the ZLuda translation layer was also looking good until Nvidia went to court to try and prevent anyone from using it commercially.
I honestly feel bad for Chinese AI researchers. Both governments are trying to keep them from buying Nvidia but nobody else’s chips have the software ecosystem to be good at training models. Inference is fine on any brand but AI companies are generally more worried about training in this phase of the game.
Don't feel too bad. Gamer's Nexus investigation THE NVIDIA AI GPU BLACK MARKET pretty much shows that the cards can be found no problem in China and in many cases with price bumps smaller than U.S. resellers.
I honestly feel bad for Chinese AI researchers. Both governments are trying to keep them from buying Nvidia but nobody else’s chips have the software ecosystem to be good at training models. Inference is fine on any brand but AI companies are generally more worried about training in this phase of the game.
There are good AI models that have been trained on chinese hardware, you don't need an ecosystem for training a neural network. It's in reality just the library PyTorch that is generally used and was Cuda centric (support now AMD ROCm). Chinese made a translation layer for PyTorch Cuda for their hardware.
There are good AI models that have been trained on chinese hardware, you don't need an ecosystem for training a neural network. It's in reality just the library PyTorch that is generally used and was Cuda centric (support now AMD ROCm). Chinese made a translation layer for PyTorch Cuda for their hardware.
Yeah and Deepseek is on record saying they couldn’t finish a single training run on Huawei hardware. Of course it’s possible to train outside of CUDA BUT It’s less performant and less reliable because the vast majority of GPGPU knowledge and code is centered around CUDA.
Don't feel too bad. Gamer's Nexus investigation THE NVIDIA AI GPU BLACK MARKET pretty much shows that the cards can be found no problem in China and in many cases with price bumps smaller than U.S. resellers.
I watched that video when it was new. They were definitely paying more than buying directly from Super Micro or other approved Nvidia partners. They only got decent deals on consumer cards like 4090s converted to 48GB. The best most had access to was the A100 which is Ampere but at least it was made on TSMC n7 instead of Samsung 8nm.
