# Why Did Investors Sell Record Profits From Samsung and SK Hynix?

According to Bloomberg's July 30, 2026 report, Samsung Electronics posted an operating profit of 28.9 trillion won (~$21 billion) for Q2 2026, while SK Hynix delivered 19.4 trillion won (~$14 billion) — both all-time highs. The combined figure exceeds the entire annual profit of the global memory industry in 2023. Yet Samsung shares fell 4.2% the day after earnings, and SK Hynix dropped 6.1%. It took four trading days for both stocks to recover their pre-earnings levels.

The selloff wasn't a rejection of the numbers — it was a rejection of the trajectory. Investors have watched hyperscaler capex guidance climb every quarter for two years, and the fear is that 2026 is the peak. The memory makers' record profits are, in this reading, the last great gasp before the AI capex cycle turns. That interpretation is plausible but ignores one structural fact: HBM supply is constrained by wafer starts decided 18 months ago, so even a demand dip in 2027 won't immediately crash prices.

# Who Controls the AI Memory Bottleneck Now?

TrendForce's July 2026 market share analysis reported that Samsung and SK Hynix together control 91.3% of the HBM market, with Micron trailing at 8.7%. This is not a commodity business anymore — it's a duopoly with pricing power that rivals OPEC's in the 1970s. SK Hynix alone holds 56% of the HBM3E segment, the current generation used in Nvidia's flagship accelerators.

The strategic consequence: hyperscalers no longer negotiate memory prices; they accept them. According to the Bloomberg report, Samsung signed a multi-year HBM supply agreement with an unnamed major AI cloud provider during the same week — a deal structure that was virtually unheard of in memory, where spot pricing dominated for decades. This shifts the risk from the memory makers to the cloud providers, who are now locked into volumes they must justify to their own CFOs.

# How Does Samsung Compare to SK Hynix After This Quarter?

# What Does This Mean for the AI Capex Cycle?

The key tension is whether memory profits are a leading or lagging indicator of AI demand. Bloomberg's report notes that SK Hynix's HBM inventory is effectively zero — every wafer produced through Q1 2027 is already allocated. This is not a supply glut story; it's the opposite. The only way this becomes a bubble is if AI accelerators get built but not deployed — and there is no evidence of that yet.

My read: the market is confusing the memory cycle with the AI cycle. Memory has historically been brutally cyclical, and investors are correctly wary. But HBM is not DRAM — it's a custom, co-designed product with qualification cycles that lock in suppliers for years. The record profits are the result of decisions made in 2024; the 2027 decisions are being made now, and they point to continued, if slower, growth.

# What Happens Next in the Memory Wars?

According to TrendForce's July 28, 2026 report, HBM4 production qualification will begin in Q1 2027, with SK Hynix holding a 12-month lead over Samsung. This is the next battleground: HBM4 doubles stack height and requires advanced packaging that both firms are racing to secure. The winner of HBM4 will control the AI memory market through 2029.

The deeper question is whether the memory duopoly can resist the temptation to overbuild. History says no — every memory upcycle ends in overcapacity. But the contract structure has changed: multi-year agreements with penalty clauses make the classic boom-bust cycle harder to execute. The AI era may have accidentally created the first stable memory oligopoly.

1. Prediction 1: By March 2027, Samsung will announce a $40B+ investment in HBM4 dedicated fabs, matching SK Hynix's capacity and triggering a price war in HBM3E that erodes 2027 margins by 15% (estimated).
2. Prediction 2: By Q3 2027, at least one major hyperscaler (Microsoft, Meta, or Alphabet) will report an AI-specific memory cost line item in earnings, making the duopoly's pricing power a public political issue.
3. Prediction 3: The Korean government will designate HBM as a 'national strategic technology' by December 2026, unlocking tax breaks and effectively subsidizing the duopoly's next capex cycle.

• The market's panic after record profits is the strongest evidence yet that AI sentiment has detached from AI fundamentals.
• SK Hynix's zero-inventory position means the AI memory shortage extends at least through Q1 2027 — no near-term supply relief exists.
• Multi-year HBM supply deals represent a structural break from memory's commodity past; pricing power has permanently shifted to suppliers.
• Micron's 8.7% HBM share makes it a spectator in the AI memory trade — its future depends on NAND, not HBM.
• The HBM4 race, not current profits, will determine the next three years of the memory duopoly's fortunes.