SK Group Chairman Chey Tae-won announced at Computex 2026 that SK Hynix will double its total wafer capacity over the next five years, citing an AI-driven memory shortage he expects to persist through the end of the decade.
The announcement landed in Taipei on Tuesday, straight from one of the most consequential stages in the semiconductor calendar. According to Reuters, Chey said SK Hynix aims to double wafer capacity over the next five years, after previously warning that the memory shortage could last until 2030. For a company already holding roughly 58 percent of the global high-bandwidth memory market, that is not a defensive posture. It is a statement of intent from a supplier that understands its own indispensability to the AI infrastructure stack.
The stakes are enormous. HBM has become the chokepoint of the AI hardware supply chain. Every Nvidia GPU that ships with HBM attached is drawing from a supply pool that, until recently, had been fully spoken for well into 2026. SK Hynix sells out its HBM allocation early because the demand picture is not softening. Hyperscalers continue to authorize nine-figure data center buildouts, and each new AI accelerator generation requires more HBM stacks, not fewer.
Chey was candid about the structural difficulty of the problem. Building a new fab takes years, and starting from a greenfield site takes even longer, which frames the challenge not as a short-term capacity blip but as a decade-scale infrastructure bet. That framing matters for anyone tracking the AI supply chain: memory cannot be conjured quickly. The lead times are measured in years, which means the commitments being made now determine what AI hardware builders can access in 2028 and 2029.
SK Hynix's spending trajectory backs up the rhetoric. The company's capital expenditure reached the mid-KRW 20 trillion range in 2025 and is projected to surpass KRW 30 trillion in 2026, the equivalent of roughly $21 billion. Its new Cheongju fab, M15X, is already coming online in the first half of this year, starting at around 10,000 wafers per month before ramping to 50,000 by the fourth quarter. That single facility alone moves the needle meaningfully on DRAM output, with the company targeting total DRAM capacity of approximately 620,000 wafers per month by the second half of 2026, nearly double the mid-300,000 level it was operating at in 2023.
The pivot to HBM4 is happening in parallel. SK Hynix completed its HBM4 mass-production setup last September and has been supplying Nvidia with paid samples that reportedly cleared final validation without issue. By most current estimates, SK Hynix is expected to supply roughly two-thirds of Nvidia's HBM4 requirements, a concentration that gives the South Korean company extraordinary pricing leverage but also concentrates execution risk in a way that Nvidia's procurement team must be watching carefully.
Samsung and Micron are not standing still
The competitive response is already underway. Samsung, currently one of SK Hynix's two major rivals in HBM, has reportedly outlined plans for a sharp increase in HBM capacity in 2026 with a focused push into HBM4. Its 2026 capex is expected to reach around $20 billion, up 11 percent year over year, with attention concentrated on advanced process nodes and its P4L facility. The catch for Samsung is that qualification timelines with Nvidia have lagged, and recovering the ground lost in HBM3E is a prerequisite to competing seriously in the HBM4 generation.
Micron is the wildcard. It has gained meaningful share in HBM after entering the race later than its Korean rivals, and it is investing aggressively, with 2026 capex projected at $13.5 billion, up 23 percent year over year. Micron has started shipping HBM4 samples rated at up to 11 gigabits per second and is working with foundry partners on future HBM4E products. For a company that was once treated as the third player in this market, that pace of investment signals genuine ambition rather than a token effort.
What this competitive picture means in practice is that the HBM market is transitioning from a near-monopoly dynamic, where SK Hynix could essentially name terms, toward a more contested landscape at the HBM4 generation. Whether Samsung and Micron can convert their investment into validated supply before demand peaks is the central uncertainty. Qualification cycles with Nvidia run long, and SK Hynix's head start at the HBM4 node is real.
For founders and operators building on AI infrastructure, Chey's Computex announcement is a useful signal to internalize. The memory bottleneck is not a temporary pricing anomaly. It is a structural feature of the current AI buildout cycle, and the companies best positioned to influence it are committing capital today that will only pay off in fabs that open three to five years from now. Pricing relief, when it arrives, will be gradual. The more immediate implication is that AI compute costs will remain elevated throughout the forecast horizon, which shapes the economics of every product and service built on top of that infrastructure. Watch the M15X ramp rate and Samsung's HBM4 qualification news closely. Those two data points will tell you more about AI chip availability in 2027 than almost anything else in the supply chain.
Also read: STMicroelectronics nearly doubles its data center revenue target as AI infrastructure demand reaches chipmakers far beyond the GPU giants • Jensen Huang called Marvell the next trillion-dollar company and the market believed him immediately • SoftBank is in early talks to anchor an $800 million round for Agile Robots as the humanoid race enters its capital-intensive phase
