Mirae Asset Securities logo

Demand strength reaffirmed

Exceptionally strong end-market conditions

We maintain our Buy rating and target price of W2,000,000 on SK Hynix, applying a target P/B of 3.4x―the average multiple of global pure memory peers―to our 12-month forward BPS of W596,154. While we slightly lowered our ASP forecasts for 2027, the impact on 12-month forward BPS is limited. As such, we leave our target price unchanged.

Following earnings releases from North American big tech companies last week, the aggregate 2026 capex estimate for global hyperscalers has been revised up to US$806bn (+73% YoY), with further expansion likely in 2027. Major tech firms are now disclosing concrete metrics demonstrating rapid growth in AI-related businesses, and given their sizable order backlogs, a sharp downward reversal in capex appears unlikely.

As of end-1Q26, the order backlog for Microsoft Azure reached US$627bn (+99% YoY), while that for Google Cloud jumped to US$467bn (+93% QoQ). In addition, Amazon stated that it has already secured customer commitments for a significant portion of its planned 2026 capex (US$200bn). Meta not only raised its 2026 capex guidance (to a maximum of US$145bn) but also emphasized that it had previously underestimated the scale of computing demand.

The demand outlook for HBM used in big tech ASICs remains strong. Microsoft disclosed that its MAIA 200 incorporates 216GB of HBM3E and delivers a 30% efficiency improvement over the previous generation. Furthermore, Google’s recently unveiled 8t TPU (for training) features 216GB of HBM3E―up 13% vs. the previous Ironwood generation―while its inference-optimized TPU 8i features 288GB of HBM3E, a 50% increase. We expect these developments to help diversify the HBM3E customer base.

We are also seeing clear signs of robust inference-oriented CPU demand. Meta disclosed that it is deploying a significant volume of AMD chips to complement its new Nvidia systems, which are likely intended for inference workloads. In addition, Meta has signed an agreement for tens of millions of AWS Graviton cores, signaling a full-scale push to develop inference-oriented CPU infrastructure. Increased CPU deployment should inevitably translate into higher demand for conventional memory.

The industry-wide trend toward long-term supply agreements is also encouraging. Competitor SanDisk disclosed that it has signed long-term supply contracts of up to five years, covering more than one-third of its total estimated FY2027 bit shipments (with a minimum disclosed order backlog of US$42bn across three contracts). The contracts also include US$11bn in financial guarantees (e.g., advance payments). We believe such contract structures are becoming increasingly common across the industry.

We forecast SK Hynix’s operating profit to reach W65tr (+72% QoQ) in 2Q26, W279tr (+491% YoY) in 2026, and W398tr (+43% YoY) in 2027.