Employees touring Samsung Semiconductor process. [Photo = Samsung Electronics] 

As Nvidia announces record-breaking performance yet sees its stock price adjust, market attention is shifting beyond simple earnings to the sustainability of the AI semiconductor cycle. 

While revenue remains strong, investors are now beginning to ask about the 'growth rate' and 'capital efficiency' rather than the 'absolute level' of growth.

This stock price fluctuation is not yet at a stage that would directly impact orders for Korean semiconductors.

The supply of High Bandwidth Memory (HBM) for AI servers remains tight, with major volumes being shipped based on long-term contracts and pre-orders. SK Hynix is leading the market with HBM3E, and Samsung Electronics is also in a phase of expanding its production capacity.

The issue is not short-term supply and demand, but the direction of 2026.

The Peak of Growth Rate, Not the Peak of Growth

The first act of AI was an infrastructure expansion phase centered on large-scale model training. The simultaneous surge in demand for data center expansion, GPUs, and HBM led to a rapid improvement in the performance of memory manufacturers.

Conceptual diagram of HBM stack structure + Base Die position.

However, the structure may change in 2026. This is because the AI industry is entering a transitional period where the central axis shifts from 'training' to 'inference.'

In the inference stage, server design and cost structures may change, and improvements in computational efficiency will be required simultaneously. This is a variable that could lead to changes in HBM capacity or the pace of price increases.

Therefore, the question the market is raising now is not about 'demand collapse' but about the 'peak of the growth rate.'

The slowing pace of growth is a more critical turning point than a decrease in absolute volume.

Year 3 of CAPEX, Entering a Period of ROI Scrutiny

Over the past 2-3 years, big tech companies have invested trillions of won in facilities. AI data centers are no longer just future investments but tangible assets accompanied by depreciation and cash flow on financial statements.

From this point on, an internal review of Return on Invested Capital (ROIC) is inevitable.

If the growth rate of AI service revenue does not sufficiently justify the growth rate of facility investment, the option of adjusting the pace of expansion may emerge.

This change is likely to manifest first as a gradual slowdown in the pace of new investment, rather than an immediate cancellation of orders.

And this signal is more likely to be detected first in the slowing trend of HBM Average Selling Price (ASP) rather than a decrease in GPU orders.

Korean Semiconductors: A Matter of Profitability, Not Volume

The Korean memory industry remains a key supplier in the AI value chain.

However, the key for 2026 is not the increase in shipment volume itself, but the sustainability of profitability.

HBM has higher margins than general DRAM. However, as supply expands and supply-demand moves towards balance, the price increase elasticity may slow down.

If the growth rate slows, performance is likely to show a gradual normalization trend rather than a sharp decline.

Currently, considering the GPU backlog and HBM supply-demand situation, the possibility of a short-term shock is limited.

However, the market has already begun to ask, not 'AI is growing,' but 'How efficiently is AI making money?'

2026 could be the first year the AI industry moves beyond a thematic phase to a capital efficiency phase. The Korean semiconductor industry also stands at the center of this structural transition.

The debate over the HBM peak is not yet concluded. However, the question has clearly begun.