By Hamza Faisal

SK hynix is moving to expand production of HBM4, the high-bandwidth memory that sits beside AI accelerators in data-centre servers, even as uncertainty hangs over its biggest customer. The South Korean chipmaker is in talks with equipment suppliers to buy the tools it needs to add a new HBM4 production line, according to a report in Korean outlet ETNews on November 15.

The orders cover back-end equipment — the machines that bond stacked memory dies together, cut wafers into chips and test the finished stacks. Suppliers have been asked to deliver and install the tools in the first half of 2027, with some lines requested as early as the first quarter.

The expansion would use SK hynix’s P&T6 packaging and testing plant in Cheongju, a site formerly known as the M8 fab that the company picked as an HBM production base in 2025. At full operation, the line would add more than 20,000 wafers a month of HBM4 output, according to people in the equipment industry cited by ETNews.

A bet on demand beyond Nvidia

The timing is notable. Nvidia, SK hynix’s largest HBM customer, is in the middle of a leadership transition after chief executive Jensen Huang resigned last week, and questions remain about how the chipmaker’s purchasing plans will evolve under new management. SK hynix’s answer, at least for now, is to keep building.

The company’s bet is that HBM4 demand is spreading beyond a single customer. As rival chip designers such as AMD and Broadcom ramp up their own AI accelerators, memory suppliers are finding new buyers for the same high-bandwidth stacks — a shift that gives SK hynix more pricing power and less dependence on any one account.

Nvidia has also lowered the HBM4 bandwidth targets for its next-generation Vera Rubin platform, a move reported earlier this week that effectively lowers the bar memory makers must clear. With a more forgiving specification, more of the modules coming off the line should pass testing, improving yields.

Sold out through 2029 — but no price tag yet

SK hynix told investors on its third-quarter earnings call that its DRAM and HBM production capacity is essentially sold out through 2029, backed by long-term supply agreements. That backlog is the commercial case for the Cheongju expansion: the company is adding capacity it believes is already spoken for.

What it has not said is how much the expansion will cost. No investment figure has been disclosed, and the company has not confirmed the exact number of tools being bought or even the final site selection. In a brief comment, SK hynix said only that “equipment orders are proceeding according to our plan, and we will respond to market growth without disruption.”

The report is sourced from several equipment makers rather than a company announcement, so the details — including the 20,000-wafer figure — should be treated as industry estimates until SK hynix confirms them. But the direction of travel is clear: the world’s biggest HBM supplier is not waiting for the dust to settle at Nvidia before adding capacity.

For the AI industry, the message is that the memory bottleneck is being attacked from the supply side. If the Cheongju line ramps on schedule in the first half of 2027, it would put more HBM4 into a market where every major accelerator maker is competing for the same scarce stacks.

By Hamza Faisal