Foxconn Chairman Young Liu recently coined the term "sovereign server," introducing the topic of sovereign AI into the hardware domain. This statement has been seen as another significant remark on sovereign AI following Nvidia CEO Jensen Huang's earlier comments.
Additionally, software companies suggest that sovereign AI could serve as a barrier for enterprises entering certain markets or provide them with competitive advantages where optimism is in the air regarding the potential surge of local small model applications in this wave.
sovereign AI is not a new concept. Its roots can be traced back to long-discussed ideas such as sovereign cloud, data sovereignty, and localized AI. It implies that organizations or nations wish to control sensitive data and its storage locations. Companies like AWS, Microsoft, Google Cloud, and VMware all offer sovereign cloud service providers.
Nvidia's master stroke
At the World Government Summit held in Dubai in February 2024, Huang clarified the meaning of sovereign AI. The firm further explained that a nation's sovereign AI means being able to produce its own AI using its infrastructure, data, labor force, and business networks.
Since the rise of generative AI, countries have significantly increased their investment in sovereign AI development. In August 2024, Nvidia announced the launch of microservices for genAI in Japan and Taiwan, allowing enterprises, government agencies, and universities to host native LLMs in their environments.
Their data indicated that Pegatron adopted the Llama 3-Taiwan 70B NIM microservice for application development, integrating it with its PEGAAi Agentic AI system to achieve process automation and enhance manufacturing and operational efficiency. Chang Chun Group, Unimicron Technology, TechOrange, LegalSign, and the generative AI startup APMIC are also utilizing Llama-3-Taiwan 70B NIM.
Foxconn's new term
During Foxconn's annual technology day, Liu focused his discussion on sovereign AI in relation to server hardware and local manufacturing. He mentioned that while many people have heard of sovereign AI, he expected there will be more discussions related to "sovereign servers" in the future.
To ensure confidentiality and security, where servers are manufactured will determine their usage in that respective country. This will ensure data safety for both hardware and software.
However, the server industry features characteristics such as foreign design outsourcing and Taiwanese contract manufacturing, which means that not every country has production lines. This prompts the issue regarding the logistics of implementing sovereign servers.
The degree of foreign involvement in design may vary, which means it may be hard to separate whether the sovereignty belongs to the brand or the manufacturer. Furthermore, there also lies the potential of sovereign servers giving rise to new unique business models. Many of these issues remain up in the air.
The surge in hard and software collaboration
APMIC, which provides model-as-a-service solutions, noted that it has established partnerships with Dell and HPE, continuing to develop sovereign computing server solutions.
Founded in 2017, APMIC is an independent software vendor in the field of generative AI under Nvidia. Originally named Asia-Pacific Intelligent Machines, it recently rebranded to strengthen its international presence. The English abbreviation was changed to Accelerate Private Machine Intelligence Company.
APMIC founder and CEO Jerry Wu stated that sovereign AI refers to the government's or organization's control over AI technologies and associated data, while data sovereignty encompasses factors such as storage location, usage methods, who can use it, and cross-border transfer restrictions.
To comply with data sovereignty regulations, businesses may need to adopt decentralized cloud strategies, deploying data and AI processing capabilities across different geographical regions. Data sovereignty could pose barriers for enterprises entering certain markets but might also become a competitive advantage.
He believes that from a technical perspective, the entry must begin with software to fully grasp the data utilization processes, including distributed fine-tuning and inference of LLMs, data governance tools, data preprocessing, and encryption masking technologies.
Wu emphasized that opportunities for chip and server manufacturers lie in the increasing demand for local model applications, with a growing shift towards small language models (SLMs), leading to heightened demand for chips supporting edge computing. Software vendors can develop software services that drive the computational power of local servers and create integrated systems for complete LLM frameworks, ensuring that all AI developments occur in a controlled environment.
