SK海力士AI内存需求强劲但面临周期风险
SK海力士AI内存需求强劲但面临周期风险
Conclusion: SK hynix’s latest results and customer agreements support a strong near-term outlook for AI memory, with demand extending beyond high-bandwidth memory (HBM) into server DRAM and data-center storage. But they do not establish a smooth, durable earnings boom: margins are exceptionally high, memory prices are a major part of the current surge, and future orders still depend on AI infrastructure spending producing adequate returns. Capacity additions, competition and the industry’s history of oversupply remain material risks.
Data as of October 31, 2026.
What supports durable demand
SK hynix’s preliminary third-quarter report showed record revenue of and operating profit of . Revenue was up and operating profit ; sequentially, they rose and . The company reported an . CNBC reported that both revenue and operating profit beat analyst estimates. The release cautions that the results are preliminary and subject to audit.
The mix and customer commitments matter as much as the headline growth. SK hynix said HBM4 shipments began in earnest and expects HBM4 to overtake HBM3E in both shipments and revenue in the fourth quarter. It plans to ramp its M15X facility to expand HBM4 production. Separately, CNBC reported server DDR5 shipments rose more than 40% quarter over quarter and DRAM average selling prices rose by a high-20s percentage range. That points to broad data-center memory demand, but also suggests higher prices and product mix are contributing substantially to the earnings surge.
The company said long-term agreements with cover of its production capacity for 2027–28. That gives better demand visibility than relying only on spot-market sales. The agreements use pricing mechanisms that reflect market conditions within agreed ranges, however; they are not evidence that all future output or margins are guaranteed. SK hynix’s expectation that tight supply will persist through 2030 is management’s outlook, not a settled industry outcome.
What could interrupt growth
-
AI infrastructure spending must earn its keep. On October 29, CNBC reported Microsoft lowered its annual capex forecast to from , while Alphabet held guidance at US$195–205 billion. Both companies’ reported capex came in below analyst estimates. Microsoft also said delays at third-party AI-chip providers would push some capacity availability into the . These reports do not show that AI investment has stopped, but they illustrate how timing, financing and returns on infrastructure can affect the pace at which memory demand turns into orders.
-
Supply could catch up—or competition could intensify. In October, CNBC reported investor concerns about Chinese competition and a possible memory glut around 2029, while noting that analysts questioned whether CXMT could scale its new technology quickly. The same report cited a Counterpoint Research forecast for DRAM prices to fall around 23% in 2028 versus 2027, while remaining well above 2025 levels. That is a forecast, not a certainty; long-term contracts may soften price swings, but they cannot eliminate the risk that new capacity eventually pressures pricing. Bloomberg also reported Samsung planned to raise spending and target share from SK hynix and Micron, adding to the competitive supply risk.
-
The boom is exposed to customer and financing strain. CNBC reported that concerns about AI-investment returns and funding pressure at neocloud providers were weighing on AI-infrastructure sentiment. SK hynix also discontinued its review of a possible Solidigm IPO, citing market conditions; the proposed listing had been intended to help fund Solidigm’s expansion. That is not direct evidence of weakening HBM orders, but it shows that financing conditions can affect expansion plans elsewhere in the AI supply chain.
-
Earnings are not uniformly accelerating. Third-quarter net income was , down even as operating profit rose. The release does not provide a reconciliation for that sequential decline, so it should not be attributed to a specific cause from these materials alone. It is a reminder that the operating-profit surge and the bottom line are not moving in lockstep.
Assessment
The strongest evidence for durability is the combination of demand across HBM, server DRAM and data-center storage, plus contracted production coverage into the medium term. The strongest counterpoint is that today’s extraordinary profitability depends on a tight market, elevated pricing and continued investment by customers. If AI infrastructure spending slows, customer returns disappoint, or industry capacity arrives faster than demand, revenue may remain substantial while margins and earnings retreat. The next tests are whether HBM4 ramps as planned, whether customer commitments translate into sustained shipments, and whether new supply is absorbed without a sharp reversal in memory pricing.