How China's AI companies are outmaneuvering US chip sanctions
How China's AI companies are outmaneuvering US chip sanctions
From efficient code to edge models, new strategies redefine the AI race.
The chip embargo imposed by the Biden administration on China nearly two years ago had one clear goal: to curb efforts by Chinese companies to develop advanced artificial intelligence (AI) systems by denying access to high-performance chips. The success has been partial at best and has already caused Chinese AI companies to form new methods which enable them to extract more from weaker chips or focus on smaller, more specialized AI models that require fewer resources.
As a result, these companies are managing to produce AI models that are in demand in the market while reducing China's dependence on Western chips, allowing for greater reliance on locally produced chips, even though they are relatively weaker. These working methods could also trickle back to AI companies in the West, helping them address two of the significant challenges of the modern AI era: a persistent shortage of the most advanced chips, which mainly affects small, resource-limited companies, and increasing energy consumption by advanced models that create problems even for large companies like Google and OpenAI.
The world of Generative AI is based on a basic paradigm: to create more advanced models, stronger computing power and more powerful processors are needed. With this understanding, in October 2022, the Biden administration decided to ban the export of high-performance chips, particularly AI chips, to China to limit the country's ability to develop advanced AI models and systems based on these chips, such as weapon systems.
Since then, restrictions have been tightened several times, and China has responded with various restrictions of its own. Although an underground network has succeeded in smuggling advanced NVIDIA chips into the country, the supply is limited. This means that local companies are forced to purchase lower-performance chips developed by NVIDIA to comply with export restrictions or use locally manufactured chips which can’t reach the performance level of NVIDIA’s more advanced chips.
Writing more efficient code
AI companies in China don’t have access to the most advanced computing capabilities and are forced to find shortcuts to bring GenAI models to market that can compete with models from Western companies. One method is to write more efficient code that allows for the development of large language models (the technology behind services like ChatGPT) using the limited processing cycles provided by less advanced processors.
For example, 01.AI, a startup backed by Alibaba and Xiaomi, uses this method to reduce the time and energy required to train AI models. "We don’t have many AI processors in China, and that forces us to develop very efficient AI infrastructure," said Kai-Fu Lee, the company's founder, to the Wall Street Journal.
Other local companies are focusing on developing smaller, more focused models and applications instead of large, general models. These models are called "edge models" because they can run locally on edge devices such as smartphones and laptops without transferring data to the cloud. Such models are at the heart of Apple's GenAI system, which will be publicly available starting this month. For example, the startup Baichuan is working with Qualcomm to integrate a relatively small LLM into AI computers in China. Samsung has also integrated small models from Baidu and ByteDance (the parent company of TikTok) into devices it sells in China.
The year of small models
China is currently at the forefront of the next AI wave. "The coming year is the year of small models," Professor Winston Ma, a law professor at New York University and an advisor to the AI startup Dragon Global, told the Wall Street Journal. "Small models need less training data, they are faster, and they allow for quicker responses."
Larger companies, such as Alibaba and Tencent, are also trying to cope with the shortage of advanced chips by improving engineering capabilities and algorithms and producing self-developed chips. "We shouldn't think that a shortage of the most advanced AI chips means we cannot lead in AI," said Zhang Ping’an, a senior Huawei executive, at an AI conference in July. "We need to abandon this mindset."
The solutions that Chinese AI companies are developing in response to the U.S. chip embargo demonstrate how efforts to prevent technological developments can be bypassed and outmaneuvered, especially when it comes to rapidly evolving technologies. Despite their efforts, the U.S. government has not succeeded in crippling AI development in China; instead, it has pushed the local market to promote alternative solutions, whether in local chip manufacturing or development methods that rely on less advanced chips. These solutions allow Chinese companies to progress in the field while reducing the market's dependence on the West, and forcing the U.S. government to consider more creative solutions to limiting AI development in China.
Chinese solutions may also seep into the West, benefiting both small and large companies alike. Small companies struggling to purchase the most advanced chips could draw inspiration from shortcuts found by Chinese companies. Even companies like Google, Microsoft, and OpenAI, face challenges due to significant investment in AI model development, including energy consumption of data centers, which hampers efforts to meet greenhouse gas reduction targets. More efficient training methods focusing on smaller, specialized models could help mitigate, and even reduce, the demand for more powerful computing resources.
