Scramble for AI Supremacy Chips Banned, Software on the Road to Success: Two New Chapters in Nvidia's China Thriller

Related Vendor

Wuxi InfiMotion Technology Co., Ltd.

Beijing may want to get rid of Nvidia, but that's harder than planned. Just as the Chinese government has banned Nvidia chips, the Chinese cloud provider Alibaba announces a global strategic partnership with... Nvidia.

The leading e-commerce and AI company Alibaba, founded by Jack Ma, announced on the sidelines of the Apsara Conference in its hometown Hangzhou that its subsidiary Alibaba Cloud will integrate "the complete Nvidia Physical AI software suite" into its AI platform PAI, reports the Chinese business portal Diyi Caijing.

This Nvidia software stack, which will be made available to enterprise customers of Alibaba Cloud through this partnership, includes tools such as Isaac Sim, Isaac Lab, Cosmos, and physical artificial intelligence databases.

The "physical AI," as it is also called in Germany, is the link between the software with its AI models and the hardware in the "physical world," such as in the control of robots, autonomous vehicles, drones, or other applications of "embodied AI."

Alibaba's AI Strategy

Alibaba is currently expanding its AI business globally, including data centers in dozens of countries. While politicians in Washington and Beijing become increasingly entrenched in their mutually damaging chip and technology war, two of the global leading "enablers" of practical AI applications are joining forces here.

Nvidia's GPUs are considered the best in the world for training AI models. By integrating the software stack developed by Nvidia alongside it, a powerful combination of big data and machine learning on the Alibaba platform and Nvidia's simulation tools, used by many users worldwide, is created.

The partnership with Nvidia will "further shorten the development cycle for applications such as embodied AI and assisted driving," said a spokesperson for Alibaba Cloud. In China, the speed at which AI is currently spreading across all industries has exceeded the expectations of even its biggest advocates.

"Isaac Sim" is Nvidia's simulation platform for robotics. It allows robots to be tested in virtual 3D environments before being deployed in the real world, such as the production lines of car factories. "Isaac Lab" is a related environment for reinforcement learning for robots. And Nvidia's "Cosmos" links simulation, models, and AI training to make robots, cars, or drones smarter.

How this announcement by the two multinational private companies will be received by the Cold Warriors in Washington and Beijing remains to be seen. Just days earlier, the Chinese government had banned domestic technology companies from purchasing Nvidia's H20 and RTX Pro 6000D chips.

However, Alibaba is, firstly, a private company, and secondly, it did not mention in the announcement of the partnership with Nvidia whether it would continue to use chips from the Americans. The partnership is about Nvidia's software, not hardware.

Tame Hardware

However, Beijing's ban on Nvidia chips marks an important turning point in the "chip war." While the U.S. government has attempted to curb China's rise as an equal AI power since 2018 through gradually intensified boycotts of semiconductors and equipment for their production, companies in both countries had tried to adapt as best they could.

Nvidia had developed the H20 and the RTX Pro 6000D specifically for the Chinese market after Washington banned the sale of top chips to the People's Republic. Both have reduced computing power compared to the H100 and RTX Pro 6000 GPUs available outside China. The H20 has approximately a quarter less power. This way, both GPUs did not fall under the U.S. government boycott—only to be banned by China itself.

Many in the U.S. had hoped to get "the best of both worlds" with slimmed-down chip variants, say Chinese analysts. They wanted to hinder rival China in the global AI race while still allowing Nvidia to conduct billion-dollar business in the Chinese market, the largest application market on earth for semiconductors of all kinds.

Instead, it now appears that in China the chips were seen less as "slimmed-down" or "throttled" and more as something less flattering. "China's rejection of the 'crippled Nvidia chips' dismisses the notion that Beijing should be satisfied with second-rate products deliberately designed to limit its capabilities," writes Professor Jinghang Zeng of the University of Hong Kong in an opinion piece for the South China Morning Post.

China's communist state and party leadership would now rather impose a withdrawal from Nvidia on its domestic industry than remain dependent on American chips and the accompanying Nvidia software CUDA. At the same time, it did not want to let the Trump administration dictate how much computing power is available in China's AI labs.

At a Politburo meeting in April this year, China's strongman Xi Jinping ordered the domestic industry to further accelerate its efforts toward "self-sufficiency" in terms of semiconductors. The goal, according to the party leader addressing the assembled leadership of his country, is to establish a Chinese "independently controllable ecosystem" of hardware and software for artificial intelligence.

Homebrew Computing Power

Chinese chip manufacturers are making rapid progress, and many Chinese AI companies are already working with them. Deepseek, the Chinese AI startup that surprised the world with a strong rival to OpenAI's ChatGPT, is experimenting with Huawei's Ascend-910-C chip for inference. ByteDance and the Ant Group are using Huawei's Ascend-910-B for their model training.

When it comes to another area, the training of large language models, the computing power of GPUs is important but not the only factor for speed and ultimate success. Another equally important factor is the available VRAM (Video Random Access Memory) of the GPUs.

Chinese developers have quickly devised a range of strategies to bypass the limitations of Washington's boycotts with software or by chaining together "throttled" or older chips. This is possible both with swiftly stockpiled H20 chips and with new developments from Huawei, Cambricon, and other Chinese manufacturers. These substitution strategies include "sharding," also known as "splitting," where a multitude of chips collectively provides enough VRAM to train a large language model (LLM).

Software Opens Doors

Other software tricks include "Model Ensembling" or "Mixture-of-Experts (MoE)." Essentially, they all aim for the same goal: to replace the Nvidia chips blacklisted by Washington, but already prohibitively expensive for many users, with a combination of somewhat weaker chips and better software.

"Advances in software are making old hardware increasingly useful," write experts from the "Berkeley AI Research Lab" and the "Berkeley Risk and Security Lab" in California in an essay on the "futility of hardware-centric export controls."

By analyzing software code in Chinese language models, the Berkeley experts, like detectives on a trail, were able to trace one codebase signal after another, showing how Chinese AI companies have developed really good "large models" even without Nvidia's H100.

The U.S. boycotts were based on the assumption that "the only strategy needed would be to restrict a country's access to advanced computing power" to limit its ability to develop strong AI and other innovations, write the American scientists. "In reality," the essay continues, "both the strategy and the underlying assumption have proven to be unreliable." (sb)

Beijing's Tougher Approach

Nvidia Becomes a Pawn in the "Chip War" Between China and the USA