The generative artificial intelligence (AI) industry has not yet fully matured, and it is currently in a transitional phase just before the emergence of a 'dominant design.' As transformer-based large language models (LLMs) rapidly converge to become the de facto standard, the competition for securing graphics processing units (GPUs), high-bandwidth memory (HBM), and power infrastructure is becoming increasingly important.
On May 19, the Software Policy Research Institute (SPRI) released a report titled "Software-Centric Society," which assessed the current state of the AI industry as entering a transitional phase in the late fluid period.
The term 'dominant design' refers to a phenomenon where a specific technological structure becomes the market standard, altering the competitive landscape of the industry. The report cites the Ford Model T in the automotive industry and the Apple iPhone in the smartphone market as prime examples. Once a particular design establishes itself as the industry standard, the focus of competition shifts from new features to production efficiency and ecosystem competition.
The report indicates that the structure of transformer-based LLMs is converging to become the de facto standard in the AI industry. It states, "Almost all major AI companies have adopted transformer-based large language model structures," noting that the industry is currently converging on a specific design at the architectural level.
However, it has not yet reached an official standardization phase. There is currently no official API or architectural specifications from international organizations such as the Organization for Economic Cooperation and Development (OECD), the International Organization for Standardization (ISO), or the International Telecommunication Union (ITU), and the revenue models of AI companies have not yet stabilized. In fact, the report highlights that Anthropic spent approximately $4.1 billion solely on research and development computing costs in 2025.
The report particularly identifies infrastructure bottlenecks as a key variable in the competition within the AI industry. Currently, HBM supply is concentrated among a few companies, including SK Hynix, Samsung Electronics, and Micron, while systems based on NVIDIA's H100 and H200 have effectively become the standard for AI computing.
Power supply issues are also emerging as a critical factor. A single AI training cluster requires hundreds of megawatts (MW) to gigawatts (GW) of power, making transmission capacity, cooling facilities, and power supply capabilities significant constraints on model expansion. The report predicts, "The future dominant design will likely be influenced not only by the excellence of model design but also by the accessibility of infrastructure resources."
Furthermore, the report forecasts that if computing efficiency, official standardization, infrastructure supply stabilization, and revenue model stabilization occur simultaneously within the next 2 to 3 years, the transformer-based AI model paradigm is likely to establish itself as the dominant design across the industry.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.