Cloud Wars: The Battle Between U.S. and Chinese Cloud Hyperscalers Intensifies as Chip Supply Chain Restrictions See No Signs of Easing

NVIDIA’s in-person GTC event returned after a hiatus of 5 years, bringing together almost the entire Artificial Intelligence (AI) ecosystem. While the world has been dominated by talks about generative AI and Large Language Models (LLMs), NVIDIA sits at the center of these conversations by providing specialized infrastructure support required to build and accelerate AI applications, supporting cloud hyperscalers, data center providers, and large enterprises.

Among the slew of updates and partnerships announced during GTC, NVIDIA also announced several public cloud partnerships that cement NVIDIA’s position as a key accelerated computing partner for generative AI innovation on the public cloud platform. Amazon Web Services (AWS), Microsoft Azure, Google Cloud, and Oracle Cloud will be offering the latest Grace Blackwell Graphics Processing Unit (GPU)-based Elastic Compute Cloud (EC2) instances and NVIDIA’s DGX Cloud for building and running inference on multi-trillion-parameter LLMs.

The introduction of NVIDIA Inference Microservices (NIM), built on the CUDA platform, allows enterprises to create and deploy AI applications, using dozens of popular AI models from NVIDIA and its partner ecosystem. The containerization of AI models represents an opportunity for NVIDIA to strengthen its partnerships with the cloud hyperscalers, with AWS integrating NIM into Amazon SageMaker, Microsoft offering NIM on the Azure Marketplace, and integration with Google Cloud’s Kubernetes engine.

The growth of AI, particularly generative AI relies heavily on cloud infrastructure for model training. The introduction of the next-generation Blackwell GPU chips will bring higher levels of AI training performance and accelerated inference capabilities. With the latest cloud partnership announcements, NVIDIA’s AI platform will be available for customers of AWS, Microsoft Azure, Google Cloud, and Oracle Cloud, which collectively represent close to 70% of the entire cloud infrastructure market. This provides the opportunity for enterprises to leverage AI-specific accelerators with vast computing resources to ultimately drive business innovation.

Conspicuously, there was no mention of partnerships with Chinese cloud providers such as Alibaba Cloud and Tencent Cloud. The tightening of China’s access to chip technology and proposals to curb China’s access to U.S. cloud computing firms have forced Chinese cloud hyperscalers to develop in-house AI solutions as an alternative to NVIDIA’s GPUs. Baidu developed its own AI microchip, Kunlun, now in its second iteration to satisfy customer demand for AI-specific workloads. Huawei’s Ascend GPUs are aimed to replace NVIDIA GPUs, while Tencent has developed Zixiao, a chip used for image and speech recognition AI applications.

In the short term, the lack of access to cutting-edge solutions provided by NVIDIA might hurt the Chinese AI ecosystem. The U.S. and China semiconductor war will force Chinese cloud hyperscaler customers to rethink their cloud platform strategy, especially in terms of accelerated AI workloads given the absence of Blackwell-powered GPUs from NVIDIA. The resiliency of the Chinese technology market cannot be taken lightly, as more Chinese chip designers, backed by the likes of Baidu and Tencent begin to fill the void left by their Western counterparts. Startups such as Enflame and Hygon have been pitching their respective chips as replacements for NVIDIA GPUs.

However, while there is no denying the AI innovation coming out of China, restrictions placed on foundries like Taiwan Semiconductor Manufacturing Company (TSMC) by the U.S. government will cause production capacity challenges. As a result, Huawei is working with Semiconductor Manufacturing International Corporation (SMIC), the largest contract chip maker in China, based in Shanghai, to develop its Ascend 910B chip, which is touted as a comparable replacement to NVIDIA’s A100 chip. Baidu also recently ordered 1,600 Ascend 910B chips to fill the void left by NVIDIA.

We are at an early stage of the AI revolution, led by the tremendous demand for enterprise-grade generative AI solutions. While NVIDIA holds court as the king of AI infrastructure, there is a vibrant ecosystem of startups and technology providers that are accelerating AI capabilities and solutions in China, without relying on U.S. chips. Most of the cloud hyperscalers in China have stockpiled NVIDIA GPUs, but this is only a short-term strategy. Alibaba Cloud, Tencent Cloud, and Baidu Cloud are fortifying in-house AI chip capabilities, while at the same time buying AI chips from Huawei and local Chinese chip designers.

The absence of NVIDIA and the ban on Intel and AMD processors in government servers and workstations further illustrate the technology decoupling between China and the United States. Based on the current geopolitical environment, it is highly unlikely that Intel and AMD will be able to make further inroads into the public cloud AI market in China, even with lower-powered chip processing capabilities that were designed specifically for Chinese customers.

While it is highly unlikely that any Chinese companies will dethrone NVIDIA as the dominant AI platform provider, there is strong demand in the Chinese market for AI-specific processing platforms, and ABI Research believes Huawei will be able to fill this demand in China. The close collaboration between Huawei and SMIC ensures that the production capacity for semiconductor chips can be fulfilled to supplement the in-house deployment of AI chips by local cloud hyperscalers. The isolation of China with the intent of scaling down its AI ambitions will further boost AI innovation in the country, ensuring that China continues to be a technological force in the long term.