AI Without Nvidia: ByteDance Seeks Alternatives Amid US Chip Sanctions

Key Takeaways

• In the face of U.S. sanctions, Chinese companies are looking to reduce their dependency on Nvidia GPUs.
• Leading the charge is Huawei, whose upcoming Ascend 910C AI chip reportedly performs as well as Nvidia’s H100.
• Huawei customer ByteDance will use the firm’s AI chips to develop a new Large Language Model (LLM).

The AI industry’s reliance on Nvidia GPUs has become so entrenched that it can sometimes feel as if there are no alternatives. 

But in China, where U.S. firms are prohibited from selling high-end AI chips, companies like ByteDance and Huawei are reimagining AI without Nvidia.

Impact of Sanctions on Nvidia Exports

When the U.S. government introduced new semiconductor export restrictions in 2022, Nvidia’s A100 was among the first chips to be affected.

Sanctions targeted the A100 due to concerns that its advanced processing power could be used by China for military applications or to boost the country’s AI capabilities in ways that could pose national security risks.

Since then, not only has the A100’s successor, the H100, been added to the list of controlled devices, but so too have a series of less powerful, mid-range chips. 

In response to the initial export restrictions, Nvidia developed new GPUs like the A800 and H800–modified versions of the A100 and H100 designed to comply with U.S. sanctions while meeting the needs of Chinese AI firms.

However, the latest sanctions have targeted the A800 too, further restricting Chinese firms’ ability to access AI chips.

Alternatives to Nvidia

While there have been reports of Nvidia chips being smuggled into China, if the Chinese AI industry is to thrive, it will likely need to procure an alternative source of powerful GPUs.

To that end, Chinese firms are developing homegrown alternatives to Nvidia chips.

Leading the charge, Huawei’s Ascend 910B offers comparable performance to an Nvidia A100. Meanwhile, the firm has reportedly  marketed its next-generation Ascend 910C as on par with the H100. 

While Huawei may still be a few years behind Nvidia, its chips are now good enough that Bytedance intends to use 910Bs to train a new large language model (LLM), Reuters reported  on Monday, Sept. 30.

In addition to the partnership with Huawei, ByteDance has also tapped Broadcom  to design two new custom GPUs expected to enter production in 2026. 

Large Language Models

In the world of AI development, LLMs function as a barometer of technological prowess and GPU capacity. And up until now, the prospect of training one without Nvidia hardware has been inconceivable.

Major players like OpenAI run superclusters consisting of tens of thousands of H100s around the clock. Even Google, which is heavily invested in the alternative TPU-based AI paradigm, still bought 50,000  H100s in 2023.

According to a company insider cited by Reuters, ByteDance’s new AI will be less powerful than its existing Doubao model. Nevertheless, forging ahead without Nvidia will represent a symbolic victory for Chinese AI.

About the Author

James Morales

Although his background is in crypto and FinTech news, these days, James likes to roam across CCN’s editorial breadth, focusing mostly on digital technology. Having always been fascinated by the latest innovations, he uses his platform as a journalist to explore how new technologies work, why they matter and how they might shape our future.

See more

[email protected] LinkedIn Twitter