Market View

Market Size

The AI industry market is projected to continue its rapid growth, bringing innovative changes to numerous sectors. By 2030, it is expected to increase the productivity of professional jobs, including white-collar professions such as software developers and legal professionals, by over 140%. As of 2023, the labor productivity value of professional jobs is $41 trillion, and by 2030, it could increase to approximately $100 trillion, including the productivity value contributed by AI. Consequently, enterprise spending on AI software is anticipated to grow annually by 42%, with businesses projected to spend around $140 billion on AI software by 2030 to enhance productivity.

According to MIT research, Generative AI can boost productivity for professionals by up to 40%. For instance, using GPT-4 for developing new product ideas, crafting marketing slogans, and writing 2,500-word articles resulted in 42.5% higher performance compared to groups that did not use AI. The most significant performance improvements were observed among less experienced professionals. While the Second Industrial Revolution increased or effectively replaced worker productivity, the AI Industrial Revolution is expected to significantly enhance or even fully replace professional productivity.

As a result, the AI market size, which was approximately $247 billion in 2023, is expected to reach about $1.8 trillion by 2030, growing at a compound annual growth rate (CAGR) of 20.4% to 36.8%. This growth is driven by the adoption of advanced AI models across industries, facilitated by sophisticated algorithms and machine learning. The AI industry is poised for continued rapid expansion due to ongoing technological advancements and increased data utilization, with AI applications accelerating across various sectors.

Supply Imbalance

Despite the industry not yet reaching mainstream adoption, GPU supply remains highly unstable and expensive. According to a report by OpenAI, developing GPT-3 involved using NVIDIA's V100 model, with costs estimated at around $4.6 million. The development utilized between 5,000 to 10,000 NVIDIA V100 GPUs in a parallel model configuration. For GPT-4, it is speculated that thousands of NVIDIA A100 GPUs (the latest devices) were required, with training periods extending over several months. Additionally, GPT-4's development involved a supercomputer with 285,000 CPU cores and 10,000 GPUs, supplemented by Microsoft's Azure cloud services.

Considering the increasing complexity of training data and parameters needed for new AI models, it is estimated that approximately 30,000 to 50,000 NVIDIA A100 GPUs will be required in the next five years. Using Amazon AWS's EC2 P4d instances (A100 GPUs), the cost is $3 per hour, totaling $26,280 annually. Operating 30,000 GPUs for a year would cost about $800 million (₩1.1 trillion).

If the speed of advancement in computing processors outpaces the increase in computing power required for AI model development, these cost issues might gradually be resolved over time. However, according to Moore's Law, computing processor performance doubles every 18 months, but this rate is currently slowing. Intel's CEO, Pat Gelsinger, states that actual performance is doubling every three years due to economic and physical constraints. Economically, the cost of building new semiconductor fabs has more than doubled, complicating expansion efforts. Physically, as transistors shrink to the nanoscale, there are inherent limitations to improving silicon-based transistor performance.

References indicate that AI computing development can be divided into three major eras. Despite a significant reduction in the rate of computing resource growth required since the advent of deep learning, the trend of doubling computing power annually continues. This rate is much faster than the current speed of performance improvement in computing processors.

By addressing the GPU supply imbalance, GPGPU aims to lead a balanced and fair development of the AI industry by building a global network of idle GPU resources. This approach will help ensure a stable supply and efficient utilization of computing power, ultimately supporting the industry's rapid growth and innovation.

Pre Deep Learning Era (1952-2010)

• Annual growth rate: approximately 41.2%

Deep Learning Era (2010-2022)

• Annual growth rate: approximately 216.7%

Large-Scale Era (2015-2022)

• Annual growth rate: approximately 93.7%

As a result, while the computing power required for AI model and service development is increasing rapidly, the advancement of processor performance is slowing, leading to a significant imbalance between supply and demand. This phenomenon is expected to dramatically accelerate the shortage of supply as AI services become more popular and widespread.

During this period, advancements in AI development software optimization and the release of more efficient GPU processors may reduce the necessary costs. However, if the growth rate of computing resources required for AI model development continues to outpace Moore's Law, the imbalance between GPU demand and supply will widen significantly.

Challenges for Startups

The high cost of GPU cloud services is particularly detrimental to small-scale AI startups. Developing AI models independently incurs costs of at least several billion won, forcing these startups to rely on AI model services like ChatGPT or Hugging Face. While using hosted AI services can effectively reduce initial costs, it does not promote a healthy development environment for the entire AI market in the long term for the following reasons:

• Cost: Although initial costs can be reduced, the long-term usage fees are higher compared to developing proprietary AI models. This becomes a significant issue as service usage scales up.

• Difficulty in Differentiation: When multiple startups use the same hosted models, it becomes challenging to differentiate their products and services. This weakens their competitiveness in the market and makes it difficult to attract new customers.

• Customization Limitations: Hosted AI models are designed to be general-purpose, which limits the ability to tailor them to specific business needs. Startups may struggle to implement the services they ultimately aim to develop.

AI startups relying on hosted AI services due to cost constraints face various challenges, including long-term cost increases, customization limitations, data security issues, reduced competitiveness, and performance and scalability issues. This results in several critical problems for AI startups:

• Unpredictable Demand: During the service development phase, it is difficult to predict the actual market demand for the product. Building a GPU server center to secure GPU computing resources under uncertain demand conditions poses significant economic risks and uncertainties for the company.

• Geographical Limitations: Even if a company secures GPU resources through its GPU server center, geographical issues may arise. Customers far from the server center may experience high latency, affecting service quality. This is why centralized cloud services often establish server centers in multiple regions.

• Security and Management: Building a GPU server center requires significant expenditure on security maintenance, facilities, and management. For services handling large-scale or sensitive personal information, substantial costs and specialized personnel are needed to maintain security.

To address these challenges, it is essential to establish a cost-effective cloud infrastructure. Utilizing a decentralized GPU network to achieve cost efficiency is the most effective alternative.