SK Hynix's Board Signs Off $38.3 Billion for Two New Memory Fabs as AI Demand Reshapes the Math

SEOUL — SK Hynix's board met on Friday and approved roughly 54.3 trillion won, or $38.3 billion, in spending on two new memory fabrication plants, the company said in a regulatory filing. The money splits two ways: 35.2 trillion won heads to the Y2 fab in Yongin, south of Seoul, and 19.1 trillion won goes to the M17 plant at the company's Cheongju campus.

The decision turns the memory maker's June master plan into shovels in the ground. Y2 becomes the second of four fabs planned for the Yongin Semiconductor Cluster and will focus on DRAM plus high-bandwidth memory for AI servers. M17 adds NAND flash capacity to the Cheongju base, which already runs the M11, M12 and M15 plants. Construction begins next year at both sites, with SK Hynix saying the timing follows customer demand rather than a fixed market outlook.

The Numbers Behind the Bet

The concrete figures matter more than the headline. Yongin Y2 spans about 341,000 pyeong, or 1.13 million square meters — roughly 160 soccer fields. Ground breaks in July 2027, with the first cleanroom targeted for June 2029. That line will produce high-bandwidth memory and other next-generation DRAM. Its sibling, Y1, is already under construction in the same cluster and should open its first cleanroom in February 2027.

Cheongju M17 is the smaller of the pair at 206,000 pyeong, about 680,000 square meters. Work starts in February 2027 and the target for the first cleanroom is December 2028. The investment periods stretch to October 2031 for Y2 and April 2031 for M17, according to the filing.

The full program is much larger than this single board vote. The Yongin cluster, spread over 4.16 million square meters in Wonsam-myeon, carries a long-term budget of 600 trillion won, with another 100 trillion won assigned to the Cheongju footprint. SK Hynix has also pulled the cluster's completion target forward by 12 years, from 2045 to 2033, and Y2 is the second phase of that accelerated timeline.

Why Memory Got So Big That Fabs Got So Big

The AI buildout has changed the economics of chip making. A single AI accelerator often costs less than the high-bandwidth memory paired with it, so the two DRAM leaders — SK Hynix and Samsung — are both committing record sums to capacity. The company's own framing is that memory has moved from a component beside the processor to core infrastructure that decides how fast an AI system runs.

The approval covers more than cleanrooms. It includes a business support building for Y2 staff, an integrated research and development center that consolidates product testing and analysis, and auxiliary infrastructure. The company says it will build fabs on the master schedule while sequencing cleanroom expansions and equipment installation against customer demand, a model it argues protects capital.

NAND demand is coming from a different corner of the same AI wave. Enterprise solid-state storage is the main source, but SK Hynix points to key-value cache storage in AI inference as a fast-growing second one. The company expects those uses to broaden further as agentic and physical AI products reach the market, which is part of why it chose Cheongju for M17 — the campus already has power and water lined up.

The Timeline and the Infrastructure

The sequence carries its own logic. SK Hynix says Cheongju offers the fastest fab construction timeline among its sites because the M11, M12 and M15 plants already operate there, which lets M17 plug into existing power, water and logistics. Phase 1 power and water infrastructure for the Yongin cluster, meanwhile, is nearly complete at a 99% progress rate, which the company credits for keeping Y2's schedule on track.

The outlay is meant to be pre-emptive. "This investment is a decision made to seize opportunities in line with the market's growth speed," an SK Hynix official said in the company statement. "By proactively securing production capabilities, we aim to establish ourselves as a key partner in AI infrastructure, contributing to the stability of the global AI semiconductor supply chain."

The board's timing lines up with an industry that is printing record numbers. The Semiconductor Industry Association reported on August 6 that global chip sales reached $403.3 billion in the second quarter of 2026, up 35.1 percent from the first quarter, with June sales of $134.5 billion up 123.6 percent from a year earlier. SIA expects full-year 2026 sales to pass $1.5 trillion. SIA's release has the full regional breakdown, including year-over-year gains of 160.9 percent in the Americas and 112.8 percent in China.

The Money Already Behind the Money

SK Hynix can afford the build. The company posted a record 60.54 trillion won operating profit for the second quarter, roughly $41.2 billion, and in July raised $26.5 billion through a Nasdaq listing of American depositary receipts, billed as the biggest US IPO by a foreign firm. Both give it headroom to fund the fabs without breaking stride, though shareholders have asked hard questions about the pace — a minority investor publicly challenged the government's proposed 1,000 trillion won cluster plan before it was finalized.

The approval also lands as the memory market's price trends are turning in SK Hynix's favor. AI server demand has kept DRAM contract prices climbing through the year, and NAND pricing has firmed as hyperscalers buy enterprise storage in bulk. CNBC noted the timing of the board vote against that pricing picture, with memory suppliers telling customers that 2027 supply is already largely committed. When a buyer asks for more HBM than a fab can ship, the answer is a new cleanroom — and the board decision is effectively SK Hynix's answer to that question in advance.

The sales data and the capex approval reinforce each other. DRAM and NAND demand is forecast to grow at a compound annual rate of 19 percent through 2030, per Omdia, and the June sales jump suggests the forecast is not out of reach. That is the picture SK Hynix's board weighed when it cleared this week's spending.

What the Board's Signature Does

Friday's approval converts intent into obligation. The master plan announced in June described the outlines of the clusters; the board vote names two specific fabs, two groundbreak dates and two cleanroom targets, and it binds the company to a spending schedule that runs through 2031. Suppliers can now book capacity around it, and the Korean government's December cluster policies have a concrete anchor.

The risk is memory's history. The industry has swung through boom-and-bust cycles before, and a 19 percent CAGR forecast is only as good as the AI services feeding it. If the agents on the enterprise storage curve slow, SK Hynix could be left holding cleanrooms several years too early. The company is hedging by phasing cleanroom fit-out and equipment to follow customer orders, a design choice buried in the release. That is also why the two fabs are starting at different times — Y2's DRAM line targets mid-2029 while M17's NAND hall arrives about six months earlier, spreading the capital outlay across several years of expected demand.

“Memory transcends its role as a mere component to become core infrastructure determining AI performance itself,” SK Hynix said, citing Omdia's forecast to explain the timing. The shareholder return review, flagged for the third quarter, is a reminder that the same board is balancing spending against stock price.

For readers tracking the memory industry, this is the top single board-level commitment by a Korean chipmaker at this point in the cycle. Semiconductor coverage has followed the memory race as it moved from layer-count wars to fab-level math, and the SK Hynix vote ties the two together: the money moved only after the demand numbers did.