SK Hynix Says Memory Stays Short Through 2030 as Long-Term Contracts Cover Over Half of 2027 Capacity
SK Hynix reported its third quarter on October 28 and the numbers are hard to take in on first read. Revenue was 104.6 trillion won, up 32% on the quarter and 328% on the year, the first time the company has crossed 100 trillion in three months. Operating profit came to 83.3 trillion won. That’s an 80% operating margin, up from 76% in the second quarter and 47% a year ago. A memory maker, in a business famous for boom and bust, now earns margins that look like software.
The headline everyone picked up was the forecast. SK Hynix says demand will run ahead of supply through 2030, and gives two reasons: leading-edge processes are hard to get right and new fabs take years to build. CFO Kim Woo-hyun repeated the 2030 line on the call. In July, CEO Kwak Noh-jung had already called 2027 the worst year in the industry’s history from the supply side.
That’s the line that travelled. Headlines led with 2030, ahead of the profit itself. A four-year shortage call from the company that sells into the shortage deserves a slower read, so here is what in the report backs it and what doesn’t.
What the contracts actually prove
The strongest evidence sits in the long-term agreements. SK Hynix has signed them with 14 large customers, leading cloud providers among them. The typical term is five years. Volume under those deals covers more than half of total capacity for 2027 and 2028, and several customers want to extend. Kim said the contracts include safeguards such as deposits, so that a customer’s purchase plan turns into real orders.
That’s real information. Customers who believe prices will fall don’t sign five-year volume commitments with deposits attached. Management went further on the call and said DRAM and HBM are effectively sold out through 2029 once volumes under negotiation are counted in (which is a softer number, since negotiation isn’t a signature).
Now the part to read twice. Pricing in those agreements “reflects market conditions within a range agreed with customers.” So the buyers have locked in supply, and the price still floats inside a band. If the shortage eases, prices can slide toward the bottom of that range while the volume stays booked. The contracts prove that buyers fear running short. They don’t prove that today’s prices, or an 80% margin, last until 2030.
Contracted coverage also stops at 2028. Everything said about 2029 and 2030 is forecast.
Who pays for it
The cost shows up somewhere else first. Kim said PC and mobile shipments are expected to fall because of what rising memory prices do to device makers’ costs. DRAM and NAND prices both rose steeply for a second straight quarter, led by HBM, server DRAM and enterprise SSDs. SK Hynix expects AI infrastructure demand to more than make up for the lost phone and laptop volume.
For a laptop brand that’s a bad trade. Memory is one of the few parts of a PC bill of materials that it can’t design around, and the supplier has just told investors that the squeeze continues. PC makers have already looked at Chinese memory to fill the gap; this site covered what Apple’s version of that move means for Micron and SanDisk in an August post.
Which raises the obvious test of the 2030 call. Can new supply from China close the gap sooner?
On current evidence, not soon. CXMT, China’s largest DRAM maker, is set to start limited HBM3E production within about two months. Its yields are put at 20% to 30%, against roughly 70% at SK Hynix, and its parts have tested closer to the older HBM3 in performance. SK Hynix has been mass-producing HBM3E since 2024, and says HBM4 will overtake HBM3E in both volume and revenue at the company this quarter. A supplier a generation behind, throwing away most of what it makes, adds some HBM for Chinese AI chip designers who can’t easily buy Korean parts. It doesn’t change the global balance in 2027.
What would break the forecast
The supply side is the honest risk. SK Hynix says 2027 capex will rise significantly, the first phase of its Yongin cluster goes into full operation next year, and the M15X line ramps to full capacity by year end. Kim said Yongin’s cleanrooms will open in phases so the company can adjust timing. That flexibility is a hedge, and companies only build hedges against outcomes they think are possible.
Demand is the other variable. The 2030 view assumes AI infrastructure spending keeps climbing, from hyperscalers, from governments building national AI capacity, and later from cars and robots. Hyperscaler capex guidance this earnings season is the near-term check on that, and it will arrive within days.
Then there’s the capital return. SK Hynix is targeting about 100 trillion won for shareholders this fiscal year, with at least 30 trillion in dividends and at least 60 trillion in buybacks, and sits on 98.7 trillion won of net cash. A company that truly saw no end to tight supply might put more of that into fabs. Returning half of free cash flow says management wants to keep some discipline in reserve.
So the call holds up through 2028 on signed paper. After that it’s an opinion from the seller, a well-informed one, and worth checking against capex plans every quarter.
The deposits are the number to watch.