CNBC's Katie Tarasov got a rare thing this summer: a camera crew inside SK Hynix's construction sites in South Korea, watching the world's leading maker of high-bandwidth memory build what the company says will become the largest network of memory factories on Earth. The reporting, published August 13, 2026, put a number on the ambition behind that buildout: up to $720 billion, aimed at tripling SK Hynix's production capacity by 2034.

That figure isn't hypothetical anymore. On August 7, SK Hynix's board formally approved the next concrete tranche of the plan, committing 54.3 trillion won, roughly $38 billion, to two specific new facilities: the Yongin Y2 DRAM fab, at approximately $24.6 billion, and the Cheongju M17 NAND flash plant, at roughly $13.4 billion. Both fall under a broader master plan the company announced last year, covering 600 trillion won for its Yongin Semiconductor Cluster alone and another 100 trillion won expanding Cheongju.

The Scale Nobody Else Is Attempting

What sets SK Hynix's buildout apart isn't just the dollar figure. It's the physical design of the facilities themselves. CNBC's on-site tour of the Yongin cluster found fabs built vertically rather than horizontally, with the first structure expected to reach the height of a 50-story apartment building, housing six interconnected cleanrooms stacked across multiple floors. That's a meaningfully different approach than the sprawling, single-story layout TSMC has used for its newer fabs in Arizona, and it reflects Korea's more limited flat land relative to the sheer scale of capacity SK Hynix is trying to pack in.

One cluster alone, at Yongin, is projected to cost $390 billion by itself. SK Hynix has accelerated its own timeline by more than a decade to get there, with four fabs at the site now scheduled for completion by 2033, a compression that speaks to how urgently the company believes it needs capacity online, not just how much capital it's willing to spend to get it. SK Hynix's own trajectory since its record-breaking Nasdaq debut this summer has been defined by exactly this kind of aggressive, front-loaded capital commitment, betting that AI memory demand justifies building faster and bigger than the industry has ever attempted before.

The Wafer With Jensen Huang's Handwriting On It

The relationship driving much of this urgency has a name attached to it. SK Group chairman Chey Tae-won showed CNBC a memory wafer bearing a handwritten message from Nvidia CEO Jensen Huang: "Please make more." It's a small, almost informal detail, but it captures the actual dynamic underneath the spending numbers. Nvidia signed a $500 billion deal with SK Group specifically to secure stable HBM supply and co-develop next-generation memory, an arrangement that also includes building new data centers alongside SK Telecom by 2027.

That relationship isn't limited to Nvidia. Chey told CNBC he meets occasionally with Microsoft's Satya Nadella, Google's Sundar Pichai, and Meta's Mark Zuckerberg, and both Huang and AMD CEO Lisa Su have personally traveled to South Korea in recent months specifically for meetings about memory supply. When a company's factory floor is drawing personal visits from the CEOs of the world's largest AI chip buyers, that's a different kind of customer relationship than a typical supplier contract, closer to the kind of direct, high-level courtship AMD has been running through its own recent acquisitions aimed at securing inference capacity as every major AI hardware company scrambles to lock in supply ahead of competitors.

The Gap America Still Hasn't Closed

Buried inside this story is a genuinely stark strategic fact, laid out clearly in TechTimes' analysis of the same investment: the United States currently has zero domestic high-bandwidth memory wafer fabrication capacity, and no CHIPS Act funding specifically targets closing that gap. SK Hynix did receive $458 million in federal grants and loans to build an advanced HBM packaging plant in West Lafayette, Indiana, expected to reach full operational capacity by late 2028. But packaging isn't fabrication. That Indiana facility will assemble and stack the finished memory components into their final form. The actual silicon wafers those components are built from will still be manufactured entirely in Korea, shipped to the US only after the most technically demanding part of the process is already complete.

Micron's own roughly $200 billion domestic manufacturing buildout represents the most ambitious American memory investment underway, but even that commitment doesn't fully close the specific HBM wafer fabrication gap SK Hynix's Korean facilities are built to fill. For now, and for the foreseeable future given how long fabs take to build, the physical origin point of the memory chips powering American AI data centers, including the enormous compute buildouts companies like OpenAI are financing through arrangements running into the hundreds of billions of dollars, remains firmly outside US borders.

Why Prices Aren't Coming Down Anytime Soon

SK Hynix's own press release framed the investment logic plainly, stating that in the AI era, the ability to supply the required volume at the exact moment customers need it has become the ultimate competitive advantage, more decisive even than raw technological capability alone. Neil Shah, vice president of research at Counterpoint Research, offered a blunter read on what that means for anyone hoping memory prices might ease soon: memory prices are unlikely to soften before the end of 2028.

That's a genuinely long runway of continued price pressure, and it's not because SK Hynix is holding back supply deliberately. Shah noted that multi-vendor expansion from Samsung, SK Hynix, Micron, and China's CXMT will meaningfully expand global memory supply through 2028. The problem is that demand is growing even faster than that expanded capacity can come online, a dynamic that's already been pushing up prices on ordinary consumer phones and laptops throughout 2026, well beyond the AI data centers actually driving the underlying shortage.

What $720 Billion Actually Buys Strategically

SK Hynix's own financial trajectory shows why the company can credibly commit to spending at this scale. Revenue nearly tripled from roughly 2023 to 2025, reaching about $65 billion, and analysts polled by LSEG now expect 2026 revenue to more than triple again, toward roughly $235 billion. That kind of growth curve is what makes a $720 billion, decade-spanning capital commitment plausible rather than reckless.

What this buildout ultimately represents is South Korea consolidating a position at the physical center of the global AI supply chain for at least the next decade, regardless of how aggressively any other country tries to build competing capacity in the meantime. Every gigawatt of AI data center capacity announced anywhere in the world, every new frontier model trained, every AI infrastructure deal signed between now and 2033 will draw, in some meaningful part, on memory chips manufactured inside towering, multi-story cleanrooms outside Seoul. That's not a temporary supply chain quirk. Given how long it takes to build a wafer fab from scratch, it's closer to a structural fact about how AI gets built for the remainder of this decade.