Will China’s Investment in Open Source Help It Catch the United States in the Generative AI Market?

Open source is quickly taking over the Artificial Intelligence (AI) market. Meta stands out as a leader with its commitment to open source (however, licensing restrictions and data availability lead many to question if these models are “open”) and the staggering investment it is making in the Llama models. However, often overlooked is the open-source AI investment in China. Open source is not new to China; since the release of Red Flag Linux in 1999, private entities supported by government investment and the massive growth in software talent have led to a burgeoning open-source ecosystem. Many key players are now focused on investing in open-source Generative Artificial Intelligence (Gen AI):

• Alibaba recently announced Qwen 2.5 and the release of over 100 open-source AI models, including text-to-video. 
• Huawei continues to invest in MindSpore AI, an AI computing framework to support cloud-to-device AI development, execution, and deployment. This platform competes with PyTorch and TensorFlow.
• Tencent continues to develop numerous open-source projects, especially targeting application development.
  • China Telecom releases TeleChat2-115B, claiming that the model was trained with 100% domestically-produced chips.
• 01.ai launches Yi, a open-source code language model with fewer than 10B parameters.

Chinese vendors are not only targeting open software development, but they are also leveraging open standards to accelerate AI system development. Similar to UALink (an open standard approach to scaling up accelerators spearheaded by AMD), Alibaba has initiated the Alink Industrial Alliance. This alliance, established between 18 entities (including AMD) covering accelerators, servers, the cloud, and industrial fields, aims to promote the development of a unified scale-up standard to accelerate large-scale AI system deployment.

However, not all key players in China agree with the open-source AI strategy. Baidu released Ernie Bot and has been very vocal about its closed-source approach to Gen AI, claiming that it offers the most effective monetization opportunity and highlighting that Large Language Models (LLMs) can never be truly open source, as they are still developed and, to an extent, controlled by one entity. This debate is likely to continue, as open-source development in Gen AI is proving very costly, and the Return on Investment (ROI) has not yet been realized.

Drawing a line between open-source investment and monetization is critical for AI vendors, given the costs involved in leading-edge Gen AI development. Most hyperscalers (Chinese and U.S.) have built their core open-source commercial strategy around developer stickiness to drive platform engagement and increase cloud traffic. But Chinese hyperscalers are not thinking purely commercial, in fact, other key geopolitical and social strategies are key drivers behind open-source AI investment:

• Global Competition: With the United States leading in closed- and open-source models, China aims to quickly improve its global position through open, community-based investment. This will enable hyperscalers to build traction with developers globally and increase stickiness to their cloud services within different regions. This will be important as Alibaba and Huawei look to build out their international position.
• Shift toward Self-Reliance: Open-source model development is being used by Chinese hyperscalers to highlight the importance of Huawei and other homegrown leading accelerators (e.g., Cambricon) in developing leading-edge AI models. From a wider economic perspective, building open-source AI models that compete with Llama and others can highlight China’s leading position in the AI market.
• Compliance with AI Regulation: China’s regulation has a major focus on transparency, especially for Gen AI services. It requires the transparent development of applications, including which data are used, how the models are trained, etc. Open-source investment will help enterprises comply with this regulation by offering transparency into the underlying processes within model development.
• Accelerate Adoption of Gen AI: Given the cost and infrastructure requirements to train and deploy AI models, the U.S.-China ban on leading-edge hardware could significantly impact Gen AI adoption. However, open-source pre-trained Gen AI models can reduce new infrastructure demands, supporting the ability for enterprise to bring models into production—enterprises will be able to cheaply access leading models. This will help start building out the application ecosystem and accelerate market development.

With an eye on its position in the global AI market, China’s open-source approach will be the fastest way to grow its international reputation and compete effectively with the United States, both commercially and geopolitically. However, back of mind will be the U.S. restrictions on leading chips (i.e., NVIDIA H100s) and the impact this has on the development of competitive, leading-edge models.

China’s open-source investment is having the desired impact with leading open-source models being released—Qwen, Yi, DeepSeek, and InternLM. However, China’s Gen AI growth will face challenges from both internal and external sources moving forward:

• Hardware Concerns: Even with significant strides in homegrown solutions (Huawei, Cambricon), satiating increasing demand for AI accelerators will not be easy, and without significant growth in supply, it will be challenging to keep developing the ecosystem.
• U.S. Restrictions on Open-Source AI Models and Technologies: On top of developing their own open-source models, Alibaba has been commercializing Llama 3.1 by providing access across its platform. Similar to building their own models, this aims to drive developer engagement. However, this commercial activity is at risk of further restrictions from the United States. OpenAI currently restricts access to its GPT models in China, and if this restriction were extended to open-source models, open-source innovation could be stifled in China.
• Scale-up Infrastructure Will Take Time to Implement, Which Will Hinder Leading-Edge Development: Although leading-edge chips are now being developed in China (Huawei and Cambricon), building competitive leading-edge models (quickly and cost-effectively) requires huge AI systems made up of thousands of interconnected Graphics Processing Units (GPUs). These AI systems are one of the key factors driving NVIDIA’s growth. Although Alink will support this scale-out architecture, it will take time to develop, and hence China’s model development will face challenges compared to those with access to NVIDIA AI systems.