Letting Nvidia sell H200s to China is closing the door after the horse has bolted
(2025/12/10)
- Reference: 1765324047
- News link: https://www.theregister.co.uk/2025/12/09/nvidia_h200s_china_ai/
- Source link:
Half a decade of US trade policy aimed at denying China access to America's most potent semiconductor tech has only served to spur China to develop homegrown alternatives.
The Trump administration's [1]decision on Monday to reverse course and allow the sale of Nvidia's H200 – still one of the GPU giant's most potent AI chips – to Chinese customers, in exchange for a 25 percent cut of revenue from sales, is unlikely to change that. That ship has already sailed, and Beijing is charting a course toward technological independence.
As evidence of this, on Tuesday the Financial Times [2]reported that the Chinese government planned to restrict access to imported H200s. Chinese leaders have already been moving in this direction for months.
[3]
After the Trump administration reversed a sales ban on Nvidia's made-for-China H20 accelerators, Beijing accused the chipmaker of conspiring with Uncle Sam to plant backdoors in its chips. Nvidia has vehemently denied the insinuation.
[4]
[5]
In September, officials in Beijing reportedly [6]ordered the nation's top tech companies to suspend testing and cancel orders of Nvidia's accelerators and pressured them to pursue homegrown alternatives instead.
Those efforts have faced challenges. Nvidia’s technical superiority has led to market dominance among developers of AI models, including those in China.
[7]
Following the success of DeepSeek R1 earlier this year, the Chinese AI darling [8]reportedly faced pressure to train its next model using Huawei-designed accelerators. However, unstable chips, glacially slow interconnects, and a dodgy software stack ultimately derailed the effort, and DeepSeek has reportedly re-tasked Huawei's Ascend accelerators for inferencing workloads.
Too little too late
Necessity, however, is the mother of all invention. Chinese AI accelerators are therefore already improving rapidly.
Earlier this year, Huawei [9]revealed its CloudMatrix 384 rack systems, which on paper boast 60 percent higher performance for dense 16-bit floating point operations and deliver roughly twice the bandwidth and 3.5x the HBM of Nvidia's GB200 NVL72 rack systems.
Individually, the Ascend 910C NPUs at the heart of the Huawei system only deliver about 75 percent of the FP16/BF16 performance of an Nvidia H200 and two-thirds the memory bandwidth.
But Huawei doesn't need the fastest accelerator to compete with Nvidia, just more chips and power. As the name suggests, Huawei's CloudMatrix systems stitch 384 of the Ascend accelerators together into a single high-performance system.
[10]
As we've previously [11]discussed , the machine comes with a compromise. While it might be faster than Nvidia's GB200 NVL72 racks, analysts estimate it also consumes three times as much energy.
Huawei isn't the only one that's embraced this approach. In November, Chinese search giant Baidu [12]unveiled two new AI accelerators designed to drive down inference costs and speed model training.
Baidu's inference-optimized Tianchi256 is expected to arrive early next year with 256 of its in-house M100 accelerators. A larger Tianchi512 with double the accelerators – almost certainly two Tianchi256 systems stitched together – is expected before 2027.
Observers believe Baidu's more powerful M300 accelerators will arrive in 2027. These chips are designed to support multi-trillion-parameter model training runs, which have become increasingly common over the past year.
We don't yet know how these chips compare to Nvidia's, but it may not matter in the end. If Beijing doesn’t want Baidu to use H20s or H200s, they'll use what they've got.
Alongside Huawei and Baidu are a slew of other Chinese GPU and XPU vendors, including Biren, Cambricon Technologies, and MetaX. According to the market watchers at TrendForce, Cambricon's next-gen chips will reportedly challenge Nvidia's H100s. For reference, the H200 is just an HBM3e-boosted H100 with more memory.
This was always a losing battle
While US efforts may have succeeded in hindering the Middle Kingdom's technological advancement in the short term, they've come at big cost: China is no longer dependent on Western technologies.
There's actually an argument to be made that US trade policy has only managed to make Chinese developers more ambitious and clever, because most of today’s top open-weight models are [13]from Chinese AI labs .
Because the [14]performance caps on AI exports weren't indexed, it was only a matter of time before Huawei or someone else managed to build a more compelling chip than Nvidia or AMD were allowed to sell into China.
The Trump administration's policy reversal therefore compounds the failure of past initiatives.
[15]China recruiting spies in the UK with fake headhunters and 'sites like LinkedIn'
[16]Trump says Nvidia can sell H200s to China – if Washington gets a 25 percent cut
[17]China's first reusable rocket explodes, but its onboard Ethernet network flew
[18]MAGA cognoscenti warn feds away from shielding AI infringers
As our sibling site The Next Platform [19]pointed out last year, this isn't even the first time the US has made this mistake.
As you may recall, in 2010, China became one of the first nations to field a supercomputer accelerated by Nvidia GPUs. The Tianhe-1A catapulted the nation to the peak of the Top500 ranking of publicly known supers. This, understandably, caught the attention of the US and the Department of Energy, which replicated this strategy with the Titan supercomputer two years later.
It wasn't long after that the US began cracking down on the export of high-end accelerators to China. In 2015, the US Commerce Department moved to [20]block Intel from selling its Xeon Phi accelerators in China. Unsurprisingly, this did little to dissuade the Chinese, and in 2017, the Tianhe-2A made its debut bearing a homegrown accelerator called the Matrix-2000.
Today, the Middle Kingdom's main focus is meeting its own demand, but China’s economic policy demands exports of technology and expertise to the world.
Leaders from several US tech companies – including Nvidia, AMD, OpenAI, Microsoft, and CoreWeave – have previously warned that by making it more difficult for the rest of the world to buy their chips, the US government was [21]handing the AI arms race to China.
Nvidia therefore welcomed the Trump administration’s policy change.
“We applaud President Trump’s decision to allow America’s chip industry to compete,” the company told The Register in an email. “Offering H200 to approved commercial customers, vetted by the Department of Commerce, strikes a thoughtful balance that is great for America.” ®
Get our [22]Tech Resources
[1] https://www.theregister.com/2025/12/09/trump_gpu_export_ban_reversal
[2] https://www.reuters.com/world/china/nvidia-shares-gain-trump-allows-some-ai-chip-sales-china-2025-12-09/
[3] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2aTj-ajnNocGx8l5NdheVlwAAAMA&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[4] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aTj-ajnNocGx8l5NdheVlwAAAMA&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[5] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33aTj-ajnNocGx8l5NdheVlwAAAMA&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[6] https://www.theregister.com/2025/09/18/nvidia_china_ai_ban/
[7] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aTj-ajnNocGx8l5NdheVlwAAAMA&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[8] https://www.theregister.com/2025/08/14/dodgy_huawei_deepseek/
[9] https://www.theregister.com/2025/07/29/huawei_rackscale_boogeyman/
[10] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33aTj-ajnNocGx8l5NdheVlwAAAMA&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[11] https://www.theregister.com/2025/07/29/huawei_rackscale_boogeyman/
[12] https://www.theregister.com/2025/11/13/baidu_inference_training_chips/
[13] https://www.theregister.com/2025/07/19/openai_us_china/
[14] https://www.theregister.com/2025/05/28/nvidia_us_chipmakers_ai_requirements_china/
[15] https://www.theregister.com/2025/11/19/mi5_linkedin_china_spy_warning/
[16] https://www.theregister.com/2025/12/09/trump_gpu_export_ban_reversal/
[17] https://www.theregister.com/2025/12/08/asia_tech_news_roundup/
[18] https://www.theregister.com/2025/12/03/maga_bannon_anti_ai_fair_use/
[19] https://www.nextplatform.com/2024/08/13/huaweis-hisilicon-can-compete-with-nvidia-gpus-in-china/
[20] https://www.theregister.com/2015/04/10/us_intel_china_ban/
[21] https://www.theregister.com/2025/05/09/tech_titans_wanna_secure_us/
[22] https://whitepapers.theregister.com/
The Trump administration's [1]decision on Monday to reverse course and allow the sale of Nvidia's H200 – still one of the GPU giant's most potent AI chips – to Chinese customers, in exchange for a 25 percent cut of revenue from sales, is unlikely to change that. That ship has already sailed, and Beijing is charting a course toward technological independence.
As evidence of this, on Tuesday the Financial Times [2]reported that the Chinese government planned to restrict access to imported H200s. Chinese leaders have already been moving in this direction for months.
[3]
After the Trump administration reversed a sales ban on Nvidia's made-for-China H20 accelerators, Beijing accused the chipmaker of conspiring with Uncle Sam to plant backdoors in its chips. Nvidia has vehemently denied the insinuation.
[4]
[5]
In September, officials in Beijing reportedly [6]ordered the nation's top tech companies to suspend testing and cancel orders of Nvidia's accelerators and pressured them to pursue homegrown alternatives instead.
Those efforts have faced challenges. Nvidia’s technical superiority has led to market dominance among developers of AI models, including those in China.
[7]
Following the success of DeepSeek R1 earlier this year, the Chinese AI darling [8]reportedly faced pressure to train its next model using Huawei-designed accelerators. However, unstable chips, glacially slow interconnects, and a dodgy software stack ultimately derailed the effort, and DeepSeek has reportedly re-tasked Huawei's Ascend accelerators for inferencing workloads.
Too little too late
Necessity, however, is the mother of all invention. Chinese AI accelerators are therefore already improving rapidly.
Earlier this year, Huawei [9]revealed its CloudMatrix 384 rack systems, which on paper boast 60 percent higher performance for dense 16-bit floating point operations and deliver roughly twice the bandwidth and 3.5x the HBM of Nvidia's GB200 NVL72 rack systems.
Individually, the Ascend 910C NPUs at the heart of the Huawei system only deliver about 75 percent of the FP16/BF16 performance of an Nvidia H200 and two-thirds the memory bandwidth.
But Huawei doesn't need the fastest accelerator to compete with Nvidia, just more chips and power. As the name suggests, Huawei's CloudMatrix systems stitch 384 of the Ascend accelerators together into a single high-performance system.
[10]
As we've previously [11]discussed , the machine comes with a compromise. While it might be faster than Nvidia's GB200 NVL72 racks, analysts estimate it also consumes three times as much energy.
Huawei isn't the only one that's embraced this approach. In November, Chinese search giant Baidu [12]unveiled two new AI accelerators designed to drive down inference costs and speed model training.
Baidu's inference-optimized Tianchi256 is expected to arrive early next year with 256 of its in-house M100 accelerators. A larger Tianchi512 with double the accelerators – almost certainly two Tianchi256 systems stitched together – is expected before 2027.
Observers believe Baidu's more powerful M300 accelerators will arrive in 2027. These chips are designed to support multi-trillion-parameter model training runs, which have become increasingly common over the past year.
We don't yet know how these chips compare to Nvidia's, but it may not matter in the end. If Beijing doesn’t want Baidu to use H20s or H200s, they'll use what they've got.
Alongside Huawei and Baidu are a slew of other Chinese GPU and XPU vendors, including Biren, Cambricon Technologies, and MetaX. According to the market watchers at TrendForce, Cambricon's next-gen chips will reportedly challenge Nvidia's H100s. For reference, the H200 is just an HBM3e-boosted H100 with more memory.
This was always a losing battle
While US efforts may have succeeded in hindering the Middle Kingdom's technological advancement in the short term, they've come at big cost: China is no longer dependent on Western technologies.
There's actually an argument to be made that US trade policy has only managed to make Chinese developers more ambitious and clever, because most of today’s top open-weight models are [13]from Chinese AI labs .
Because the [14]performance caps on AI exports weren't indexed, it was only a matter of time before Huawei or someone else managed to build a more compelling chip than Nvidia or AMD were allowed to sell into China.
The Trump administration's policy reversal therefore compounds the failure of past initiatives.
[15]China recruiting spies in the UK with fake headhunters and 'sites like LinkedIn'
[16]Trump says Nvidia can sell H200s to China – if Washington gets a 25 percent cut
[17]China's first reusable rocket explodes, but its onboard Ethernet network flew
[18]MAGA cognoscenti warn feds away from shielding AI infringers
As our sibling site The Next Platform [19]pointed out last year, this isn't even the first time the US has made this mistake.
As you may recall, in 2010, China became one of the first nations to field a supercomputer accelerated by Nvidia GPUs. The Tianhe-1A catapulted the nation to the peak of the Top500 ranking of publicly known supers. This, understandably, caught the attention of the US and the Department of Energy, which replicated this strategy with the Titan supercomputer two years later.
It wasn't long after that the US began cracking down on the export of high-end accelerators to China. In 2015, the US Commerce Department moved to [20]block Intel from selling its Xeon Phi accelerators in China. Unsurprisingly, this did little to dissuade the Chinese, and in 2017, the Tianhe-2A made its debut bearing a homegrown accelerator called the Matrix-2000.
Today, the Middle Kingdom's main focus is meeting its own demand, but China’s economic policy demands exports of technology and expertise to the world.
Leaders from several US tech companies – including Nvidia, AMD, OpenAI, Microsoft, and CoreWeave – have previously warned that by making it more difficult for the rest of the world to buy their chips, the US government was [21]handing the AI arms race to China.
Nvidia therefore welcomed the Trump administration’s policy change.
“We applaud President Trump’s decision to allow America’s chip industry to compete,” the company told The Register in an email. “Offering H200 to approved commercial customers, vetted by the Department of Commerce, strikes a thoughtful balance that is great for America.” ®
Get our [22]Tech Resources
[1] https://www.theregister.com/2025/12/09/trump_gpu_export_ban_reversal
[2] https://www.reuters.com/world/china/nvidia-shares-gain-trump-allows-some-ai-chip-sales-china-2025-12-09/
[3] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2aTj-ajnNocGx8l5NdheVlwAAAMA&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[4] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aTj-ajnNocGx8l5NdheVlwAAAMA&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[5] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33aTj-ajnNocGx8l5NdheVlwAAAMA&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[6] https://www.theregister.com/2025/09/18/nvidia_china_ai_ban/
[7] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aTj-ajnNocGx8l5NdheVlwAAAMA&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[8] https://www.theregister.com/2025/08/14/dodgy_huawei_deepseek/
[9] https://www.theregister.com/2025/07/29/huawei_rackscale_boogeyman/
[10] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33aTj-ajnNocGx8l5NdheVlwAAAMA&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[11] https://www.theregister.com/2025/07/29/huawei_rackscale_boogeyman/
[12] https://www.theregister.com/2025/11/13/baidu_inference_training_chips/
[13] https://www.theregister.com/2025/07/19/openai_us_china/
[14] https://www.theregister.com/2025/05/28/nvidia_us_chipmakers_ai_requirements_china/
[15] https://www.theregister.com/2025/11/19/mi5_linkedin_china_spy_warning/
[16] https://www.theregister.com/2025/12/09/trump_gpu_export_ban_reversal/
[17] https://www.theregister.com/2025/12/08/asia_tech_news_roundup/
[18] https://www.theregister.com/2025/12/03/maga_bannon_anti_ai_fair_use/
[19] https://www.nextplatform.com/2024/08/13/huaweis-hisilicon-can-compete-with-nvidia-gpus-in-china/
[20] https://www.theregister.com/2015/04/10/us_intel_china_ban/
[21] https://www.theregister.com/2025/05/09/tech_titans_wanna_secure_us/
[22] https://whitepapers.theregister.com/