The numbers say Samsung Electronics posted an operating profit of 10.4 trillion Korean won in Q2 2024. That is a 1,460% surge year-over-year. The headline screams AI-driven memory super cycle. But the data beneath the surface tells a different story. One where the market leader in DRAM and NAND is losing the most profitable battle. The HBM war.
Context: The Memory Giant’s Core Business
Samsung is the world’s largest memory chipmaker, holding ~42% of the DRAM market and ~33% of the NAND market. Its semiconductor division accounts for roughly 55% of total company revenue, with the rest coming from mobile and consumer electronics. The Q2 profit spike was driven by two forces: a cyclical recovery in conventional DRAM/NAND prices, and the explosive growth of High Bandwidth Memory (HBM) used in AI accelerators. Yet the data reveals a structural weakness. Samsung’s HBM market share sits at only ~30%, versus SK Hynix’s ~50%. And in the fastest-growing segment—HBM3E—Samsung is still ramping production, lagging behind SK Hynix by at least six months.
Core: The On-Chain Evidence of Competitive Strain
Let me verify the past. I have tracked HBM-related capital expenditure and supply chain data across the three major memory players. Samsung’s 2024 semiconductor capex is estimated at 50 trillion won, with a significant portion allocated to HBM capacity expansion at its Pyeongtaek P3 facility. The target is to triple HBM production by late 2025. But execution risks are visible. The P3 line has reportedly faced construction delays of three months due to labor shortages and cost overruns. Meanwhile, SK Hynix is already shipping HBM3E in volume to NVIDIA, with an estimated monthly output of 250,000 stacked chips. Samsung’s HBM3E output is projected to reach only 300,000 units per month by end of 2024, still behind SK Hynix’s 250,000? Wait—the math does not weep, it merely liquidates. The gap is closing but SK Hynix is moving faster.
Diving deeper into the technology stack. Samsung’s HBM3E uses TC-NCF (Thermal Compression Non-Conductive Film) for stacking, while SK Hynix uses MR-MUF (Mass Reflow Molded Underfill). The latter offers better thermal dissipation and higher yield. Samsung’s yield on 1β nm DRAM, the base die for HBM3E, has been rumored to be 60-70%—below the 80%+ threshold needed for profitable HBM production. Industry checks suggest Samsung’s HBM yield has improved recently, but the company lost the initial NVIDIA certification because of inadequate yield and thermal performance. The cost of being late is not just lost revenue; it is a permanent loss of trust in the supply chain.
Now look at the NAND side. Samsung’s V-NAND is on the 8th generation at 238 layers, with the 9th generation (290+ layers) expected in H2 2024. SK Hynix has already shipped 321-layer NAND, and China’s YMTC is using hybrid bonding to reach 232 layers. Samsung is roughly 12 months behind in layer count. This matters because enterprise SSDs are a major beneficiary of AI inference workloads—datasets are growing, and capacity per drive is jumping from 2TB to 8TB+. Samsung’s SSD business could see 30% revenue growth in 2024-2025. But the delay in NAND layer advancement creates a window for competitors to lock in capacity contracts.
On the demand side, the market is real. HBM revenue is expected to grow from $20 billion in 2024 to $40 billion in 2025, and $60 billion by 2027. AI training consumes the bulk today, but inference is accelerating. The math says HBM penetration in the total memory market will rise from 3% to 30% over four years. That is a sea change. But Samsung’s share of that growth is capped by its technology lag. The company is betting everything on HBM4, due in 2026, which will use hybrid bonding—a shift that could leapfrog SK Hynix’s MR-MUF. Yet the development cycle is 18-24 months, and any hiccup in hybrid bonding yield could widen the gap further.
Contrarian: Correlation Is Not Causation
The narrative that Samsung is riding a perpetual AI wave is dangerous. Let me inject a dose of pre-mortem rigor. The data shows that conventional DRAM and NAND prices have already recovered from the 2023 trough. DRAM contract prices rose 15-20% quarter-on-quarter in Q2 2024, and NAND rose 30-40%. But this cyclical uplift is temporary. All three memory makers are aggressively expanding capacity. Samsung, SK Hynix, and Micron combined will add significant new output by H2 2025. Historical memory cycles last 8-12 quarters; we are already in quarter 4 of the upcycle. If AI demand softens—say, hyperscalers pause GPU procurement—the market could swing back to oversupply quickly. The market euphoria masks a lurking commodity risk.
Moreover, Samsung’s heavy reliance on NVIDIA for HBM sales (est. 80% of its HBM output) is a single-client risk of the highest order. NVIDIA is not loyal; they diversify across SK Hynix, Samsung, and now Micron. If Samsung’s HBM4 fails to meet NVIDIA’s specs, SK Hynix could capture an even larger share. The math does not weep, it merely liquidates. Samsung’s stock trades at 15x trailing PE versus SK Hynix’s 18x—the discount reflects an implicit market judgment that Samsung is a follower, not a leader, in the highest-growth segment.
Takeaway: The Signal for Next Quarter
Watch Samsung’s official Q2 earnings report on July 30. The key signal is not the headline profit—it is the HBM3E shipment volume and gross margin. If Samsung discloses a meaningful ramp in HBM3E volume and confirms NVIDIA certification for the next-gen B200 GPU, the market can reassess. If it stays vague, the technology gap remains. I do not predict the future, I verify the past. And the past six months of data says Samsung is still catching up. Liquidity is not a promise, it is a state of flow. The capital is flowing to SK Hynix in the HBM pool. Samsung needs to prove it can swim faster.