A record quarter with an important caveat

SK hynix has reported preliminary second-quarter 2026 revenue of 79.3187 trillion won and operating profit of 60.5426 trillion won, both quarterly records for the company. Revenue was 51% higher than in the first quarter and 257% higher than a year earlier, while the 76% operating margin underlines how strongly the sales mix has shifted towards higher-value memory products.

The numbers should be treated with the usual caution attached to an earnings release: SK hynix says its financial results remain subject to its independent audit. Even so, their scale is meaningful. The company had already posted record revenue of 52.5763 trillion won in the first quarter. The further acceleration in the second quarter indicates that the demand surge is not confined to a single product cycle or a short-lived inventory rebound.

The company attributes the result to rising prices and sales of high-value DRAM, high-bandwidth memory, enterprise SSDs and other memory products used in AI servers. Its reported net income was unusually higher than operating profit, reflecting items outside ordinary operations; for judging demand and the underlying business, revenue, operating profit and the product mix are therefore the more useful indicators.

HBM4 moves from roadmap to supply-chain issue

The more consequential disclosure is that SK hynix began mass shipments of HBM4 during the second quarter and intends to increase production in the second half of 2026. HBM, or high-bandwidth memory, places multiple DRAM dies in a vertically stacked package close to an AI accelerator. The approach provides much wider data paths than conventional server memory, helping processors feed large AI models without being as constrained by memory bandwidth.

HBM4 raises the technical and manufacturing stakes. SK hynix previously said its 12-layer HBM4 design could offer more than 2TB of bandwidth per second and 36GB capacity, using a 2,048-I/O interface that doubles the prior generation’s interface width. Such specifications matter because performance in large AI systems increasingly depends on the balance between accelerator compute, memory capacity, power use, packaging and interconnects—not simply on the peak speed of the GPU or custom chip.

A shipment announcement does not reveal volume, customer allocation or the pace at which chips will be incorporated into finished servers. Still, a mass-shipment claim is different from supplying engineering samples. It places HBM4 within the operational chain that includes qualification, advanced packaging, accelerator assembly and rack deployment. That transition is central to the next generation of AI infrastructure.

Long-term agreements signal a changing memory market

SK hynix says it has completed long-term agreements with around 10 customers and is discussing further arrangements with other major clients. This is a notable response to a sector historically known for volatile spot pricing and abrupt swings between shortages and oversupply.

The logic is straightforward. AI-system developers need assured access to specialised memory years before a platform reaches full deployment, while suppliers need greater confidence before committing capital to cleanrooms, equipment and packaging capacity. Multi-year agreements cannot eliminate market cyclicality, but they can make demand planning less dependent on short-term purchasing behaviour.

The strategy is not unique to SK hynix. Micron has also framed memory as a strategic input and reported high-volume HBM4 shipments for a lead customer platform, alongside qualification samples for multiple end customers. It has disclosed multi-year strategic customer agreements across data-centre, consumer and automotive markets. The parallel moves suggest that leading suppliers are trying to sell supply assurance and engineering collaboration alongside memory chips.

For buyers, this model offers predictability but can reduce flexibility. For suppliers, it may support more disciplined investment, although the risk remains that customer forecasts, technology roadmaps or AI spending plans could change before newly built capacity comes online.

Competition is broadening, not disappearing

SK hynix’s HBM4 ramp comes in a market where all three major memory manufacturers are pursuing the same transition. Samsung said in February that it had started HBM4 mass production and commercial shipments, and later said it had begun sampling HBM4E. Micron has positioned its HBM4 for NVIDIA’s Vera Rubin platform and is already preparing the next HBM4E generation for volume production in 2027.

This overlap means the competitive question is no longer who can announce an HBM4 product first. It is who can qualify products on customer schedules, produce them at high yield, assemble them reliably, deliver sufficient volumes and maintain attractive power and cost characteristics. In HBM, yields and packaging capacity can be as commercially important as the DRAM process node itself.

An earlier industry assessment from TrendForce had pointed to differing qualification timelines and potential shipment delays among suppliers. SK hynix’s latest statement therefore provides an important company-level update, but its production ramp will need to be evaluated through subsequent customer deployments and quarterly results. The distinction between initial shipment, sustained volume supply and final system adoption will remain important throughout 2026.

Capacity expansion is necessary but gradual

SK hynix is responding by accelerating the M15X production schedule and preparing for a Yongin Phase 1 cleanroom opening in early 2027. It also cited planned investments in an advanced-packaging facility, a NAND production base and a new semiconductor cluster. The company says those projects will be phased according to customer demand and investment efficiency.

That emphasis on capital-expenditure discipline is significant. HBM converts more wafer capacity, stacking, testing and packaging resources into each finished unit than mainstream memory. Expanding output is therefore not as simple as increasing conventional DRAM wafer starts. It requires coordinated investment across specialised manufacturing steps, while suppliers must also preserve capacity for conventional server DRAM and other profitable products.

The immediate implication is continued pressure on memory availability for AI systems and, potentially, for adjacent server-memory categories. SK hynix argues that demand for AI memory and conventional memory is now expanding together as AI becomes more widely deployed. Whether that becomes a durable structural shift will depend on sustained cloud and enterprise investment, not only on enthusiasm for AI models.

Memory becomes a core AI infrastructure constraint

SK hynix’s quarter is best read as evidence of a broader hardware realignment. AI demand is lifting the value of memory that can deliver high bandwidth, high capacity and acceptable power consumption at scale. It is also pulling memory suppliers deeper into customers’ product-planning cycles.

The company’s reported HBM4 shipments, long-term customer agreements and capacity plans do not guarantee that today’s exceptional profitability will persist. Memory remains a cyclical industry, and large expansion projects can eventually alter the supply balance. But the current phase looks different from a traditional commodity upturn: specialised memory, advanced packaging and committed supply are becoming critical inputs to the pace at which next-generation AI servers can be built.

Sources