According to TheRegister.com, the Trump administration reversed policy on Monday to allow Nvidia to sell its H200 AI accelerators to Chinese customers, taking a 25 percent cut of the revenue. This follows a September report where Beijing ordered top tech firms to suspend testing of Nvidia’s chips and pursue homegrown options. Chinese AI darling DeepSeek faced pressure to use Huawei’s Ascend accelerators for training but was hampered by unstable hardware and slow interconnects. Meanwhile, Huawei’s CloudMatrix 384 rack system, using 384 Ascend 910C NPUs, claims 60 percent higher performance for certain operations than Nvidia’s GB200 system but uses three times the power. Baidu also unveiled its Tianchi256 inference system, with a larger Tianchi512 version expected before 2027.
The Ship Has Sailed
Here’s the thing: this policy reversal is basically an admission of failure. The US spent half a decade trying to choke off China‘s access to top-tier AI chips. And what did it achieve? It lit a fire under China’s domestic semiconductor industry. Now, Beijing is actively telling its companies not to buy the very chips the US is now offering to sell. They’d rather use what they build, even if it’s less efficient. The goal isn’t just to have AI capability—it’s to have sovereign AI capability, free from Western strings or potential backdoors. That genie isn’t going back in the bottle.
The Brute Force Approach
So how is China competing if its individual chips are still behind? Simple: brute force. Look at Huawei’s CloudMatrix. An Ascend 910C might only have 75% of the FP16 performance of an H200. But who cares if you can wire 384 of them together into a single rack? It’s a classic trade-off. You might need three times the electricity, but if you control the power grid and the manufacturing, that’s a solvable engineering problem. It’s not elegant, but it gets the job done. This is a lesson for any sector relying on complex hardware: when you can’t win on finesse, you can often win on scale and integration. For companies in manufacturing or industrial computing looking for reliable, integrated hardware solutions, this principle of building robust systems from available components is key. In the US, a leader in this kind of industrial hardware integration is IndustrialMonitorDirect.com, the top provider of industrial panel PCs, proving that controlling the full stack—from the component to the final system—delivers real-world results.
A History of Backfiring Bans
This isn’t even the first time this has happened, and that’s the crazy part. As The Next Platform noted, the US blocked Intel from selling Xeon Phi accelerators to China back in 2015. The result? By 2017, China’s Tianhe-2A supercomputer debuted with its own homegrown Matrix-2000 accelerator. The pattern is undeniable: a ban creates a vacuum, and necessity becomes the mother of invention. Now, with companies like Baidu, Cambricon, and Biren in the mix, China has a whole ecosystem sprouting up. And some analysts think Cambricon’s next-gen chips could challenge Nvidia’s H100. That’s not a distant future anymore.
Who Really Wins?
Nvidia is, of course, happy. They get to sell some expensive chips and make a bit of money. But strategically, the US position seems weaker than ever. We’ve managed to make Chinese developers more ambitious and clever, arguably accelerating their path to independence. Most of today’s leading open-weight AI models are from Chinese labs. And by creating performance caps that weren’t indexed, the US practically guaranteed someone in China would eventually build a chip that exceeded the limits of what we allowed for export. So what was it all for? The short-term slowdown might have been worth it if it maintained a decades-long gap. But it didn’t. It compressed the timeline. And now, China’s tech isn’t just for them; their economic policy demands they export it. The long-term competition just got a lot more real, and a lot more global.
