In 2026, a $725 billion (about 1,064 trillion won) artificial intelligence (AI) investment boom unleashed by the United States' four major hyperscalers (large-scale data center operating corporations)—Amazon Web Services (AWS), Microsoft (MS), Alphabet (Google), and Meta Platforms—is delivering a massive "AI windfall" to the Korean economy. Because of the resulting shortage in memory semiconductors (hereafter memory), securities firms project that the combined 2026 operating profit of Samsung Electronics and SK hynix will reach 600 trillion won.
According to a May 15 report by Bloomberg Economics (local time), Samsung Electronics and SK hynix will generate about 17% of Korea's gross domestic product (GDP) in 2026 and will drive 2.7 percentage points of GDP growth, which is forecast to be around 3%.
In particular, over the next year (the second half of 2026 to the first half of 2027), the two companies' corporate tax payments will reach 80 trillion won, comparable to Korea's total corporate tax revenue in 2025, and are expected to rise to 100 trillion won in the following year. As Kim Yong-beom, the presidential chief of policy at the Blue House, recently argued on social media (SNS) for a "national dividend" system funded by AI excess tax revenue, debate over the AI windfall is spreading.
But before such discussions, it is necessary to recognize that this memory boom is operating in a way completely different from the past. In the AI era, as the semiconductor industry shifts to "customer order-based production," manufacturers' market leadership is not what it used to be despite price surges. To gauge the sustainability of excess profits, we must examine how production, inventory management, and pricing structures are changing.
Memory prices have soared, but utilization has slumped
According to the Ministry of Data and Statistics (MODS) Korean Statistical Information Service (KOSIS), the semiconductor manufacturing capacity index rose from 55.3 in March 2016 to 191.5 in March 2026, a 3.46-fold increase. In contrast, the semiconductor production index expanded only 2.81 times over the same period, from 57.3 to 161.3. While fabs (factories) and equipment—semiconductor manufacturing infrastructure—expanded more than threefold, production growth failed to keep pace. Actual production relative to capacity fell from about 110% in 2016 to around 90% over the past year. In past booms, fabs ran above "rated capacity" through overtime and line conversions, but recently "full-throttle equipment runs" have diminished.
This trend also appears in semiconductor manufacturing utilization rates. In past booms, fab utilization peaked at 120%. However, even as prices for the latest server DRAM, "DDR5 16Gb," surged more than sixfold in a year, utilization stayed in the 80%–90% range over the past six months. It only exceeded 100% in February, when mass production of Samsung Electronics' sixth-generation High Bandwidth Memory (HBM4) kicked into full gear. It is notable that manufacturers have not indiscriminately ramped up utilization.
The semiconductor market is shifting to "made-to-order" production
This change stems from the altered production paradigm in the AI era. As competition intensifies over ultra-fine processes to support high-performance computing (HPC) for AI training and inference, semiconductor equipment prices and fab construction costs are rising exponentially. The latest extreme ultraviolet (EUV) lithography tool "High-NA" from Netherlands-based ASML costs about $400 million (about 596.5 billion won) per unit. The surge in semiconductor capacity over the past decade reflects rising capital intensity from the adoption of ultra-expensive equipment.
Semiconductor manufacturers are pivoting to a focus on high–value-added products. In the era when commodity DRAM and NAND dominated the market, a strategy of mass production to cut unit costs and create demand worked during booms. But now that made-to-order semiconductor production tailored to big tech (large information technology corporations) is the norm, a strategy of churning out commodity memory alone cannot secure profitability. For HBM, on which SK hynix and Samsung Electronics have staked their future, success hinges less on simple output and more on optimization with graphics processing units (GPUs), advanced packaging, high yields, and customer certification. In the AI era, competitiveness boils down to the ability to reliably supply chips optimized for the systems designed by customers.
In the AI era, market power is shifting to design and platforms
The changed production mode suggests that in the AI era, power in the semiconductor industry has shifted from manufacturing to design and platforms. Whereas past memory price increases arose from manufacturers' market power through supply control, the recent price spike stems from the explosive expansion of the AI ecosystem led by big tech such as Nvidia. Korean memory manufacturers are merely enjoying excess profits from their pivotal position at the supply chain bottleneck.
The shift in market power also shows up in inventory management. In past booms, manufacturers raised utilization and proactively built inventory to meet surging demand. Recently, however, they have been reluctant to overproduce. Even as DRAM prices soared up to tenfold after June 2025, the semiconductor inventory index stayed below the baseline of 100. For customized memory such as HBM produced with high-cost equipment and advanced processes, excess inventory can translate into massive losses. As "produce after preorders" has strengthened, utilization and inventory are kept from exceeding certain thresholds. The more made-to-order production takes hold, the weaker inventory becomes as a buffer across the supply chain, and bottlenecks intensify. In booms, supply bottlenecks generate huge excess profits, but when the investment cycle turns, order cancellations can lead to inventory losses and spill over into systemwide risk for the economy.
Where should the AI windfall be used?
The AI-driven memory supercycle has delivered a massive windfall to the Korean economy, but it also carries risks as the economy's dependence on semiconductors rises. The current excess profits do not come from the market dominance of Samsung Electronics and SK hynix, but from the spillover gains of supply chain bottlenecks created by hyperscalers' astronomical capital expenditure. If the AI data center bubble thesis materializes and memory prices fall sharply, today's windfall could turn into an "AI shock."
How to distribute excess profits from the AI windfall is not the pressing issue now. The focus should be on where to reinvest a windfall created by bottlenecks. This means using it not for income support to boost consumption and asset markets, but as funding for industrial restructuring and future investment.
Otherwise, the Korean economy could become a victim of an "AI Dutch disease." Above all, excess tax revenue from the AI boom should be directed not to populist redistribution but to funding restructuring in low-productivity industries. At the same time, it should serve as seed money for investments in infrastructure, technology, and talent. It is worth heeding the advice that "this boom is the last chance to change an economic structure dependent on memory price cycles."