U.S. Commerce Department Loosens AI Chip Export Ban, Permitting Nvidia H200 and AMD MI325X Sales to China

In a sweeping recalibration of global technology policy, the U.S. Commerce Department's Bureau of Industry and Security (BIS) announced on Saturday that it will permit the export of current-generation AI accelerators to China. The ruling specifically clears Nvidia's H200 and AMD's MI325X chips for sale to verified commercial entities, dismantling a blanket ban that had defined U.S.-China tech relations for nearly four years.[1][3]

The H200 and MI325X represent the current workhorses of global artificial intelligence infrastructure. Equipped with advanced High Bandwidth Memory (HBM3e), these processors are the foundational engines required to train and run massive frontier models efficiently. By removing them from the restricted entity list, Washington is effectively normalizing the baseline hardware requirements for global AI development.[4]

The mechanism of this policy shift moves the U.S. away from a rigid "performance density threshold"—which banned chips based purely on their processing speed and interconnect bandwidth—toward a "Verified End-User" (VEU) telemetry model. Under the new framework, Chinese tech giants and cloud providers can purchase the hardware provided they submit to continuous, automated auditing of the workloads running on the chips.[3]

The economic rationale behind the reversal was heavily championed by the U.S. semiconductor industry. Silicon Valley executives have spent the last two years arguing that absolute export controls were starving American firms of crucial revenue in their largest overseas market. That lost capital, they warned, was desperately needed to fund the staggering research and development costs for next-generation architectures like Nvidia's upcoming Rubin platform.[2][5]

Beyond corporate balance sheets, the policy shift reflects a stark geopolitical reality check. Intelligence assessments and industry analysts increasingly concluded that the export controls were inadvertently accelerating China's drive toward semiconductor independence. Rather than halting Chinese AI progress, the bans forced a massive influx of state capital into domestic alternatives.

Evidence of this domestic catch-up became undeniable earlier this year when Huawei's Ascend series and SMIC's advanced manufacturing nodes proved far more resilient than Washington anticipated. Chinese labs were successfully training state-of-the-art models by networking thousands of slightly less efficient domestic chips together, creating a parallel hardware ecosystem that threatened to permanently lock U.S. firms out of the market.[4]

The new VEU telemetry system is designed to thread the needle between commercial dominance and national security. Chips exported under the new rules must be integrated with approved cloud monitoring platforms that log workload signatures in real-time. If the telemetry detects computational patterns associated with military applications, weapons development, or mass surveillance, the hardware can be remotely throttled or cryptographically bricked.[1][3]

Despite these safeguards, the enforcement mechanism introduces significant uncertainty. Cybersecurity experts and national security analysts have questioned whether on-chip telemetry can be reliably secured against sophisticated state-sponsored spoofing. The technical challenge of differentiating between a massive matrix multiplication used for a commercial language model and one used for military logistics remains an open problem.

The immediate market reaction was explosive. Semiconductor stocks surged globally, with Nvidia and AMD seeing significant pre-market bumps. Across the broader tech sector, there was a palpable sigh of relief as the global supply chain—which had been rapidly fracturing into inefficient, regional silos—showed signs of restabilizing.[2][5]

For global AI research, the implications are profoundly positive. Over the past two years, Chinese AI labs have been major contributors to the open-source ecosystem, releasing highly capable open-weight models despite hardware constraints. With access to standard, CUDA-optimized hardware restored, the friction of porting models to alternative silicon is eliminated, likely accelerating the pace of global open-source AI breakthroughs.[4]

The decision also eases diplomatic tensions with allied nations. Countries like Japan and the Netherlands, which manufacture the lithography machines required to build these chips, had faced immense pressure from Washington to match U.S. export controls. The loosening of American restrictions provides these nations with much-needed diplomatic breathing room and economic flexibility.[3][5]

Ultimately, the Commerce Department's move signals a strategic pivot from absolute containment to a "run faster" doctrine. By allowing the sale of current-generation chips while keeping next-generation architectures like Blackwell strictly restricted, the U.S. is betting that maintaining a one-generation technological lead—funded by global market dominance—is a safer long-term strategy than attempting to freeze a rapidly evolving science.[3]