Samsung and SK Hynix Double Down on AI Chips: The Paradox of the Semiconductor Cycle

Context

Samsung Electronics and SK Hynix are confirming a massive acceleration in their investments in HBM (High Bandwidth Memory) chips dedicated to AI, despite growing signs of oversupply in the standard DRAM market. This paradox—massive investment while the semiconductor cycle shows signs of slowing—is explained by the unique structure of AI demand: contracted over 2-3 years with hyperscalers and Nvidia, it bypasses the usual mechanisms of the spot market.

Data

• Samsung: Estimated semiconductor CAPEX for 2026 at ~$40 billion (annual reports/Economic Times).
• SK Hynix: World's leading supplier of HBM3E—~50% market share, majority of revenues from AI.
• Each Nvidia B200 GPU requires 8 stacks of HBM3E (192 GB), compared to 4 for the H100—memory density doubled.
• Standard DRAM: Prices down ~15% in Q2 2026 (Trendforce)—cyclical oversupply in PCs/smartphones.
• GPU compute utilization rate: 85% (RUM Group, July 2026)—no oversupply on the AI side.
• Argentum AI: $4.1 billion contract for 27,000 GB300 GPUs (Nvidia Blackwell) signed in June 2026—confirmation that AI GPU demand remains under pressure.

Analysis (Mechanism)

The semiconductor market is structurally bifurcated. Standard memory (DRAM, NAND) follows a classic supply-demand cycle with current cyclical oversupply. HBM memory for AI operates in a quasi-captive market: Samsung and SK Hynix do not rely on the spot market but consolidate order books with Nvidia, AMD, and hyperscalers committed for several years. The risk of a traditional cycle is partially neutralized by this contractualization. The risk lies in technological transition: if HBM4 (expected in 2027) requires a major overhaul of production lines, current HBM3E CAPEX could depreciate faster than anticipated.

Probabilistic Scenarios

• Sustained HBM supercycle (50%): AI demand grows faster than capacity → HBM remains tight until 2027. Samsung and SK Hynix outperform in their AI segments despite weak standard DRAM.
• Worsening bifurcation (35%): HBM remains tight but standard DRAM collapses (-30% in 2026). Overall results disappointing despite AI growth.
• Technological shock (15%): Adoption of HBM4 or alternative architectures (CXL, in-package memory) faster than expected → accelerated depreciation of HBM3E CAPEX.

Portfolio Implications

SK Hynix remains the purest bet on HBM (majority of revenues from AI). Samsung offers foundry diversification (partial alternative to TSMC) but at a higher P/E ratio. ASML (lithography equipment) remains the most defensive exposure to the AI semiconductor cycle. To monitor: HBM unit margins, which could compress if TSMC enters HBM production (H2 2027).

Risks & Blind Spots

SK Hynix's dependence on a limited number of key customers (Nvidia in particular) poses a significant concentration risk. A Blackwell→Rubin absorption dip in H1 2027 could create a temporary weakness window.

To Watch

Samsung Q2 results (end of July) · SK Hynix CAPEX guidance for H2 2026 · Nvidia Rubin announcements (GTC September 2026) · HBM3E spot prices (Trendforce monthly).