SK Hynix sets record as quarterly profit jumps five-fold, says AI chip demand exceeds capacity

SK Hynix reported a record quarterly profit, with earnings up five-fold, and said AI chip demand for HBM already exceeds its production capacity over the next three years. The company sees pricing remaining favorable for now as AI demand offsets weaker PC and smartphone demand, supporting continued robust earnings growth. It also signaled higher 2025 capex versus last year's 30.2 trillion won and is reviewing shareholder returns including dividends, buybacks, and cancellations.

Analysis

The key read-through is that AI demand is no longer just absorbing incremental capacity — it is pulling the entire memory supply stack into a multi-quarter rationing regime. That is structurally bullish for NVDA because constrained HBM availability forces customers to optimize around premium GPU deployment rather than substitute away, which preserves Nvidia’s pricing power and mix even if unit growth normalizes. It is also a strong signal for ASML: when memory vendors are still pulling forward capex and committing to EUV tooling, the bottleneck shifts from demand uncertainty to tool delivery and fab readiness, supporting a longer-than-consensus order runway.

Second-order, this is less about one quarter of outsized earnings and more about the industry’s inability to build through the cycle fast enough. Memory is effectively transitioning from a cyclical commodity to a strategic AI input, which should keep spot/contract spreads elevated and delay the usual mean reversion that shorts often rely on. The implication is that upstream equipment and materials suppliers likely capture the best risk/reward over the next 6-18 months because they benefit from multi-year capacity expansion without taking end-demand pricing risk.

The contrarian risk is that the market may be overconfident on duration: once new HBM lines come online, the pricing curve can inflect sharply because memory supply responds in chunks, not smoothly. Another underappreciated risk is capex intensity — if every vendor chases the same AI prize, return on incremental fab investment could compress by 2026 even while near-term earnings look excellent. The setup is bullish now, but the trade becomes more differentiated after the next 2-3 quarters, when investors start discounting eventual oversupply rather than current shortages.

Geopolitics and supply-chain diversification reduce headline risk, but they also increase resilience and therefore the probability that this supercycle lasts longer than expected. That lowers the odds of a near-term supply shock reversal and argues for staying with the secular beneficiaries rather than fading the move on traditional memory-cycle logic. The cleanest expression is to own the picks-and-shovels names on pullbacks and avoid chasing the memory producers after outsized prints, where upside is increasingly priced and downside will come from capex execution or future oversupply, not current demand.

Key Decisions for Investors

• Stay long NVDA on any post-earnings consolidation; use a 3-6 month horizon and focus on the setup where HBM scarcity protects GPU mix and pricing. Risk/reward remains favorable so long as supply constraints persist, but trim into any sharp multiple expansion.
• Add ASML on weakness over the next 1-3 months as a higher-quality way to express multi-year AI capex. The risk/reward is attractive because it monetizes capacity expansion without direct exposure to memory pricing normalization.
• Initiate a basket long on memory equipment and materials versus memory producers: long ASML / short a generic memory proxy if one rallies into strength. Thesis is that capex beneficiaries have better durability than cyclical earnings peaks over a 6-12 month window.
• If the market starts pricing in 2026 oversupply, consider taking profits in pure-play memory exposure and rotating into NVDA/ASML. The catalyst to watch is commentary on new fab ramp schedules and any signs of order elongation beyond the next two quarters.
• Use call spreads rather than outright longs for NVDA if implied volatility spikes after the next AI demand read-through. This keeps upside participation while limiting premium paid into a name where the demand story is strong but consensus is already crowded.