AI Semiconductor Market Formula: the Winner Takes All
Jul. 21, 2025 –
The top 5% of companies (based on annual sales) in the semiconductor industry, including NVIDIA, TSMC, SK Hynix, and Broadcom, swept up all the profits generated by the industry last year, according to a report released by global consulting firm McKinsey & Company on July 20.
The economic profit captured by the top 5% companies amounted to $159 billion, while the middle 90% of companies’ profits were limited to $5 billion. The bottom 5% of companies actually incurred losses of $37 billion. In effect, the top 5% companies received a report card exceeding the economic profit ($147 billion) created by the entire semiconductor market.
This market shift occurred in just 2-3 years. During the COVID-19 pandemic period (2021-2022), the economic profit taken by the middle 90% of companies exceeded $30 billion annually. When converted to average profit per company, this is about $130 million. However, as the AI semiconductor boom began in 2023, these companies’ average profit sharply dropped to $38 million. Last year, it further declined to $17 million, representing an 88% reduction in profit over two years.
McKinsey predicted that semiconductor companies within the AI industry group will grow at an average annual rate of 18-29% until 2030, while traditional semiconductor companies not directly related to AI will only grow at an average annual rate of 2-3%. McKinsey analyzed, “While a few companies are recording unprecedented levels of profit riding the AI value creation boom, most companies face a completely different reality.”
The formation of a winner-takes-all structure is possible because leading companies can dominate the standards for new semiconductor products. For existing products, the Joint Electron Device Engineering Council (JEDEC) first creates standards, and companies develop products accordingly. However, for semiconductors with entirely different specifications, the entering company takes the lead in establishing standards. This gives them the privilege of being the "rule maker" in newly opening markets, blocking the entry of latecomers.
For example, when SK Hynix first developed the first generation of HBM in 2013, development and standard establishment proceeded simultaneously. The special DRAM module SOCAMM (scalable compute acceleration memory module), which NVIDIA is currently promoting with the goal of popularizing personal AI supercomputers, is also a representative case where a specific company created an independent memory standard. As the paradigm of the semiconductor industry has shifted to producing customized chips reflecting customer requirements, this phenomenon is expected to intensify further.
An industry insider said, “For products that didn’t exist before, manufacturers’ opinions inevitably have to be reflected significantly in establishing standards.”
The problem is that the Korean semiconductor industry is standing aside from the center of these changes. While Korea occupies more than 50% of the global memory semiconductor market, its competitiveness in the GPU and ASIC markets, which are core to AI semiconductors, is significantly lower than that of the United States and Taiwan. Currently, apart from SK Hynix’s high bandwidth memory (HBM), there are virtually no domestic companies that can be included in NVIDIA’s AI value chain. Some AI semiconductor startups are emerging in Korea, but due to lack of funding and manpower, they are only targeting niche markets where NVIDIA has not entered.