SK Hynix Just Solved the Hardest Problem in HBM — Now It Needs to Build It at Scale

SK Hynix completes 12-layer HBM4 validation with hybrid bonding, orders $15M equipment, and invests $3.87B in new fabs to lead Samsung in next-gen AI memory.

The race to dominate next-generation AI memory just entered a new phase. SK Hynix has cleared a critical technical hurdle that will determine whether it can hold its lead over Samsung when the industry shifts to HBM4.

The company confirmed it has completed internal validation of 12-layer HBM stacks using hybrid bonding technology — a manufacturing technique that eliminates the physical bumps traditionally used to connect memory layers. Instead, chips make direct contact, enabling faster speeds and lower heat generation. For 16-layer HBM4 stacks destined for Nvidia's upcoming Rubin platform, the lower overall stack height is a decisive advantage.

Kim Jong-hoon, a technical leader at SK Hynix, told an industry gathering that all engineering challenges around high-density stacking have been resolved. He declined to disclose specific yield figures. The next challenge is scaling production to commercial volumes.

The $15 Million Machine That Could Decide the Race

That scaling effort now has a price tag. SK Hynix has placed an order for a hybrid bonding inline system jointly developed by Applied Materials and Besi. The system costs roughly $15 million and combines polishing and plasma processing tools with a hybrid bonder. It marks the first time SK Hynix has purchased hybrid bonding equipment explicitly for mass production.

The company is also preparing to acquire tools from Korean manufacturer Hanwha Semitech for quality testing.

The capacity expansion is unfolding on multiple fronts. In Indiana, SK Hynix is investing approximately $3.87 billion in a new packaging fab dedicated to AI memory. In Cheongju, a separate facility is under construction with a budget of around 13 billion euros, targeting completion by the end of 2027.

Samsung is not standing still. The rival Korean chipmaker is testing Besi equipment and additional bonders from domestic manufacturer SEMES. But SK Hynix enters this phase from a commanding position. According to Counterpoint Research, it holds 62 percent of HBM shipments and 57 percent of HBM revenue.

Record Margins Fund a Shift Into Physical AI

The financial numbers explain why SK Hynix can afford to move aggressively. Revenue hit 52.6 trillion won in the latest quarter, up 60 percent from the prior period. Operating margin reached 72 percent. The stock has surged roughly 90 percent since the start of the year and trades just below its 52-week high.

That cash flow is now being redirected toward a broader strategic pivot. SK Hynix is deepening its partnership with Nvidia beyond data center memory. Senior managers from both companies met in Seoul this week, with Nvidia sending Madison Huang, who leads technical marketing for physical AI platforms. The two firms plan to integrate memory solutions more deeply into Nvidia's Omniverse platform, which simulates industrial processes.

The logic is straightforward. Industry is building massive GPU clusters for autonomous robots and digitized factories. These systems require memory with very different characteristics than standard server workloads. The shift from pure training models to what the industry calls "agentic AI" — where AI agents process data in real time and react continuously to their physical environment — is expected to drive a massive increase in memory consumption per computing cycle.

Analysts See Room to Run

KB Securities has raised its price target on SK Hynix to 2 million won, arguing that the current market capitalization represents a structural floor. Daol Investment & Securities goes further, setting a target of 2.1 million won, citing the company's dominance in AI memory and its successful entry into physical AI infrastructure.

The stock currently trades at around 1.29 million won. The gap between today's price and the highest analyst target implies roughly 63 percent upside.

Industry expectations place the gradual introduction of hybrid bonding into HBM4 products starting in the second half of 2026. Whether SK Hynix can master the technology at sufficient volumes by then will determine how long its lead over Samsung lasts — and whether it can supply the memory that Nvidia's next-generation platforms will demand.