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’ by Ritwik Gupta, Leah Walker, and Andrew W. Reddie provides concrete evidence of export control violations.

In May 2024, Tencent released the HunyuanDiT text-to-image diffusion model, which was reportedly run and trained on Nvidia A100 GPUs. In September last year, Tencent introduced the GameGen-O diffusion transformer model, which was also believed to use high-end, export-restricted Nvidia GPUs.

The research paper reverse-engineered these models by analyzing representative code signatures.

Interestingly, the training scripts show Nvidia Collective Communications Library (NCCL), which is only compatible with Nvidia GPUs. This rules out the use of AMD tech or any other third-party GPUs.

Next, both models support bfloat16, which is only available with the Ampere microarchitecture, such as Nvidia 30XX, 40XX, A100, and other GPUs. This also rules out the possibility of using older GPS.

More importantly, the training scripts show Remote Direct Memory Access (RDMA) configurations over InfiniBand. Again, RDMA is only supported on Nvidia’s data center GPUs, such as the H100, H20, and A100. Consumer GPUs, such as the RTX 3090, do not allow for this configuration.

Lastly, the training scripts were also tweaked to include Advanced Network Parameter Tuning and bonded interfaces. Such customization is only possible if the researchers had physical access to the hardware, which points to in-house clusters and not off-the-shelf solutions.

All this research points to the possible use of Nvidia A100 or H100, which directly violates the US export restrictions.

How Is China Circumventing Restrictions?

Whenever an export restriction is placed, it takes around a few weeks (or even months) to come into effect. This gives the parties involved a lot of buffer time to stockpile the restricted goods, i.e., Nvidia AI chips.

So, it’s quite possible that China might have pre-ordered and accumulated a sizable quantity of the export-restricted tech before the restrictions actually came into force.

A second theory is that the country may be accessing these chips through illegal black markets operating both within and outside its domestic borders. There have been cases of individuals caught smuggling electronic parts, but there’s still no concrete evidence for these accusations.

Another source for these chips could be third-party entities and shell companies registered outside of China. For this, the US needs to place strict background checks and due diligence procedures to prevent the chips from falling into the hands of the Chinese.

Nvidia’s Modified Chips Can Be a Boost for China

Another important question for the US to ponder is whether these downgraded chips would stop China from developing advanced AI models. You only have to look back at the recently launched Hunyuan-Large open-source LLM model from Tencent.

This AI model delivers state-of-the-art performance and competes directly with Meta‘s Llama 3.1, DeepSeek V2, and Mixtral-8x22B. The project’s README suggests that it was entirely trained on Nvidia H20 GPUs.

The H20 model complies with all US export controls and only offers 75% of the performance when compared to the Nvidia H100. So, technically, using it should not have resulted in an AI model as powerful as the Hunyuan-Large.

However, China has been using advanced machine learning techniques to overcome the shortcomings of these chips through its whack-a-mole approach. For starters, it uses Mixtures-of-Experts – an architecture that activates only parts of the model per training step. This helps achieve the same accuracy as monolithic models on limited hardware.

Tencent also used mixed-precision training with the help of bfloat16. Mixed precision training can train models up to 2.5x faster than full precision training when using advanced GPUs such as Nvidia A100. Similarly, by using Quantization, these models can be converted to a lower bit representation, which speeds up training with minimal loss of accuracy.

Other techniques include large VRAM usage, sharded training, and efficient GPU communications. This suggests that an inferior version of the H20 chip would make no difference to China. It has already developed advanced architecture to customize the configuration of such chips and squeeze out the best performance.

Nvidia might just be aware of this and playing into the hands of the Chinese to save its revenue. Remember, the company has already been under pressure since the release of DeepSeek. It’s an AI model developed at just a fraction of the cost of premium US-made models like ChatGPT and Gemini.

If the Chinese can figure out large-scale cost optimization of high-performance AI models, the demand for Nvidia’s overly expensive AI chips may fall drastically. Therefore, the company is trying all it can to ensure sales are not impacted.

What remains to be seen is whether Trump would allow such circumvention of export rules or if the US is aware enough of these Chinese techniques. Only time will tell.

Read more: Nvidia plans to establish $500 billion worth of domestic production chain