Generative AI is rapidly gaining traction, attracting many aspiring entrants into the field. However, the lack of computing resources and the difficulty of data collection pose significant barriers for newcomers. Taiwanese cloud gaming service provider Ubitus has revealed its possession of tens of thousands of datacenter-grade GPUs, placing it among the few companies in the "GPU-rich" group. This revelation may potentially position major gaming companies as a new force in advancing GenAI research.
In August, SemiAnalysis published an article that divided computing power demand into two groups based on GPU availability: GPU-Rich and GPU-Poor. The GPU-Rich group comprises a small number of companies that each own more than 20,000 A100 or H100 GPUs, including OpenAI, Google, Anthropic, X, Inflection, and Meta. Some of these companies may possess over 100,000 GPUs. Meanwhile, in San Francisco Bay Area of the US, many teams are actively boasting their GPU holdings, signaling a burgeoning trend.
On November 8, Ubitus took a significant step by collaborating with the National Taiwan University's Department of Computer Science and Information Engineering. The company's founder and CEO, Wesley Kuo, disclosed that Ubitus is probably the largest holder of GPUs in Taiwan, with its GPUs originally intended to support cloud gaming computations.
Kuo recalled a visit by former Google Taiwan managing director Chien Lee-Feng and others, who suggested that Ubitus, with its abundant resources, should consider contributing to Taiwan's technological landscape. This suggestion arose around the beginning of 2023, shortly after OpenAI's chatbot, ChatGPT, gained widespread popularity.
Developing localized, traditional Chinese LLM for Taiwan
Kuo emphasized his team's expertise in cloud gaming software development and deployment for both PC and mobile applications. They are also channeling their surplus resources into large language model (LLM) applications, with a particular focus on developing localized and traditional Chinese LLM for Taiwan and extending support for other local languages in the future.
Currently, Nvidia chips primarily serve data center servers and the gaming industry. In the second quarter of 2023, its datacenter revenue reached US$10.32 billion, jumping 141% on quarter and 171% on year, while gaming revenue experienced only a slight growth.
Taiwanese industry experts familiar with Nvidia highlighted the significant differences in specifications and unit prices between gaming GPUs and datacenter GPUs. For instance, the RTX 4090 chip is a top-of-the-line gaming card, while the H100 and A100 chips are primarily used in datacenter servers. Gaming cards start at NT$10,000, while datacenter GPUs range from NT$100,000 to NT$1 million, resulting in a substantial price gap.
In response, Kuo stated that his team primarily owns datacenter-grade GPUs, including H100, A100, and L40S, rather than general consumer gaming cards. This preference stems from the risk of overheating in gaming cards when faced with excessive computing requirements.
Furthermore, they opted for PCIe specifications instead of servers with sets of 8 GPUs, because it's more convenient to repair individual GPUs. In contrast, servers with eight soldered-together GPUs would require component disassembly for maintenance.
Google Taiwan, following the conclusion of the 2023 Taipei Game Show in early February, pointed out that Taiwan ranks as the fifth largest mobile game market globally. In 2022, only South Korean and Taiwanese markets showed revenue growth, with Taiwan market experiencing the fastest growth rate.
AWS cloud service user and Taiwan-based startup, Wonders.ai, also revealed in February that their virtual character application already includes voice dialogue functionality, with approximately 10 billion model parameters for mobile use. The team started developing this concept approximately three years ago, not following the ChatGPT trend.
