Nvidia’s downgraded H20 chips might not be enough to stop China’s AI ambitions

Nvidia is planning to introduce a downgraded version of its H20 AI chip to cater to the Chinese market. Recently, Trump had imposed a licensing requirement for the export of these chips to China. However, Nvidia doesn’t seem to be in the mood to let go of its huge Chinese market.

After all, China accounts for 13% of the company’s total sales, amounting to $17 billion in revenue as of January 2025. And, in only five months, the company is sitting on $18 billion worth of H20 orders – a considerable part of which can go down the drain because of the export regulations.

However, no company would want a huge chunk of its revenue to be taken away for domestic production. Hence, Nvidia seems to have found a workaround for these export restrictions. The exact details of what capabilities the company plans to downgrade haven’t been made public yet.

The chips are set for a July launch, and Nvidia has already intimated major customers like Tencent. However, with Trump keeping a keen eye on which AI tech gets into the hands of the Chinese, the US government might block this move.

China Bypassing US Circumventions

The question we’re asking is, do export restrictions even work? There are a lot of assumptions about China circumventing export controls introduced by the US.

However, a report titled ‘ Whack-a Chip: The Futility of Hardware Centric Export Controls (19459044)’ by Ritwik, Leah, and Andrew W. Reddie, provides concrete evidence of violations of export control.

Tencent released HunyuanDiT, a text-to image diffusion model in May 2024. It was reportedly trained and run on Nvidia A100 graphics cards. Tencent released the GameGen-O diffuser model in September of last year. It was believed that this model also used high-end Nvidia GPUs with export restrictions.

This research paper reverse-engineered the models by analyzing representative codes signatures.

• It is interesting to note that the training scripts use Nvidia Collective Communication Library (NCCL), a library only compatible with Nvidia graphics cards. This eliminates the use of AMD technology or any third-party GPUs.
• Next both models support bfloat16 which is only available on Ampere microarchitecture GPUs such as Nvidia 40XX and A100. This also eliminates the possibility of using older GPS.
• The training scripts also show Remote Direct Memory Access configurations using InfiniBand. RDMA is supported only on Nvidia data center GPUs such as the A100, H100, and H20. Consumer GPUs such as the RTX3090 do not support this configuration.
• Finally, the training scripts have been tweaked so that they include Advanced Network Parameter tuning and bonded interfaces. This customization is only possible when the researchers have physical access to the hardware. This indicates that in-house clusters are the best solution, not off-the shelf solutions.

This research indicates that Nvidia A100 and H100 could be used, which would directly violate the US export restrictions.

What is China doing to circumvent the restrictions?

It takes a few weeks or even months for an export restriction to take effect. This gives the parties involved plenty of time to stockpile restricted goods such as Nvidia AI chip.

It’s possible that China may have pre-ordered a large quantity of the export restricted tech before the restrictions came into effect.

Another theory is that China may be able to access these chips via illegal black markets, both inside and outside of its borders. Although there have been some cases of people being caught smuggling electronics parts, there is no concrete evidence to support these allegations.

Shell companies and third-party entities registered outside China could also be a source of these chips. In order to prevent these chips from falling into Chinese hands, the US must implement strict background checks and due-diligence procedures.

Nvidia’s Modified Chips Could Be a Boost For China

A second important question that the US should consider is whether these downgraded Nvidia chips would prevent China from developing advanced AI model. Tencent’s Hunyuan Large open-source LLM is a good example.

The AI model is a state-of-the art AI that competes with Meta’s Llama, DeepSeek, and Mixtral-8x22B. The README of the project indicates that it was trained exclusively on Nvidia’s H20 GPUs.

According to the H20 model, it complies with US export controls but only offers 75% the performance of the NvidiaH100. Technically, it shouldn’t have produced an AI model that was as powerful as Hunyuan-Large.

China has used advanced machine learning techniques through its whack a mole approach to overcome the limitations of these chips. It uses Mixtures-of-Experts, an architecture that activates parts of the model at a time. This allows for the same accuracy of monolithic models with limited hardware. Tencent used mixed-precision with bfloat16. Mixed precision training is up to 2.5x faster when using advanced GPUs like the Nvidia A100. Quantization can also be used to convert these models into a lower bit representation. This speeds up training without compromising accuracy.

Other methods include large VRAM use, sharded learning, and efficient GPU communication. This means that a lesser version of the H20 would not make a difference to China. It has developed advanced architectures to customize the configurations of such chips, and squeeze the best performance. Nvidia may be aware of this, and is playing into the Chinese’s hands to save revenue. Remember that the company is already under pressure after the release of DeepSeek. It’s a model of AI developed at a fraction of what premium US-made models such as ChatGPT and Gemini cost.

The demand for Nvidia’s expensive AI chips could fall dramatically if the Chinese are able to optimize the cost of high-performance AI models on a large scale. The company is doing everything it can to avoid any negative impact on sales.

It remains to be seen whether Trump would allow this circumvention of export regulations or if the US has enough knowledge of these Chinese technologies. Only time will tell.

Read More: Nvidia plans $500 billion worth of domestic manufacturing chain