Generative AI commanded attention in the 2023 tech scene, fueling speculation about its trajectory in the new year. Sega Cheng, iKala's Co-founder and CEO, sheds light on crucial trends worth monitoring in 2024, ranging from computing power shortages and China's tech ambitions to model miniaturization and the rising costs of AI compliance.
The shortage of computing power
The scarcity of high-end GPUs has affected large and small businesses alike. Especially for many AI startups that are reliant on cloud-based GPU computing, the shortages of H100 have led to their adoption of alternatives like Google TPU and AWS chips.
Cheng believes that the shortage of computing power will remain a significant challenge for AI development in 2024. To address this, businesses are adopting a hybrid approach, combining both cloud-based and physical GPUs, resembling hybrid and multi-cloud models. He observed clients' increasing tendency to reserve computing power in advance, converting operational expenses into capital expenditure for hardware procurement. Moreover, deploying GPUs domestically is also becoming a common strategy to reduce system response delays, Cheng suggests.
China's LLM catch-up
Amid the ongoing US-China tech conflict and the continuous efforts to decouple supply chains. Chinese companies, including key players like Baidu, Alibaba, Tencent, Huawei, and others have pushed forward in the development of LLMs. They aim to potentially bridge the gap with their US counterparts. Reports indicate the adoption of open-source models like Meta Llama by these companies. However, Cheng thinks the global influence of Chinese models as it currently stands, remains constrained by the quality of their training data.
LLM miniaturization
In the realm of LLM, there has been a shift towards miniaturization. Mistral AI, a France-based startup, stands out as a noteworthy player in this space. Cheng suggests that Mistral, championing open-source principles, could be viewed as the new generation of OpenAI, particularly as OpenAI shifts towards a more guarded approach and away from open sourcing. Mistral AI presently features models with 7 billion and 13 billion parameters, employing the Mixture of Experts (MoE) architecture. These models show promising performance, trailing GPT-4 by only 1-2%.
Cheng underscores the pivotal lesson from Mistral's success, illustrating that achieving model reduction without significant performance compromise has been proven feasible but still demands a substantial computing power threshold. With miniaturization and the integration of MoE architecture, developers can utilize various small models to create multimodal models or models covering multiple knowledge domains.
Rising compliance costs
With the passing of the EU's AI laws, Cheng accentuates the escalating importance of safety, drawing parallels to the global impact of the General Data Protection Regulation (GDPR). In the case of high-risk AI systems, the EU might mandate developers to relinquish original models, potentially leading to heightened compliance costs for businesses employing AI.
iKala's role in the cloud services market of Taiwan
iKala's core business involves cloud service brokerage and marketing technology applications. Navigating the cloud landscape, iKala seamlessly integrates with AWS and Google Cloud services, while Microsoft Azure services are facilitated through its affiliate, WiAdvance. The revenue distribution between cloud and marketing hovers around the 7:3 to 6:4 ratio.
When questioned about the potential impact of the ChatGPT phenomenon on the company's growth, Cheng says it is difficult to directly attribute to it. Nevertheless, he highlights an overarching surge in the demand for big data and cloud services due to AI, with many enterprises lately focusing on building data platforms.
While growth rates for US cloud service providers are plateauing, Taiwan's cloud market held on with robust momentum. According to iKala's CEO, the latter half of 2023 witnessed a clear shift in client demands, progressing beyond conceptual proofs-of-concept (PoC) to tangible implementation stages.
Cheng points out that the company's latest significant clients hail from the manufacturing and financial sectors. In manufacturing, cloud adoption is driven by continuous automation needs, coupled with the launch of foreign carbon taxes and domestic carbon credit trading. The financial sector's embrace of cloud services, on the other hand, stems from relaxed regulations, escalating geopolitical risks prompting hedging strategies, and increased demand for backup solutions.
Article translated by Jerry Chen
