SK Hynix Sees Memory Shortage Through 2030 as AI Demand Squeezes PCs and China Lags
For most of its history the memory business has been a commodity trade with a bad temper. Prices go up, everyone builds fabs, prices crash, someone nearly goes bust. Investors learned to sell memory stocks when the numbers looked best.
This week’s news says that pattern is on hold. SK Hynix told investors it expects memory shortages to last until at least 2030, and it backed that view with a quarter so strong it barely looks like a memory company anymore. Around it, the rest of the week’s headlines fill in the picture: a Chinese challenger shipping high-bandwidth memory at yields most fabs would scrap, a PC market shrinking because parts cost too much, PC makers moving assembly back to China to save money, and even Fujifilm betting on magnetic tape.
Taken together it’s one story. AI data centers are eating the world’s memory supply, the leaders are being paid handsomely for it, and everyone downstream is paying the bill.
The Quarter That Broke the Memory Playbook
SK Hynix reported Q3 revenue of about $78.8 billion, up 328% from a year earlier and ahead of the roughly $74.1 billion analysts expected. Operating profit came in near $62.8 billion, up 632%, against estimates of about $57.8 billion.
Run the arithmetic and the margin is the real headline. $62.8 billion of operating profit on $78.8 billion of revenue is an operating margin of roughly 80%. Back out the growth rates and the same quarter a year ago was about $18.4 billion of revenue and $8.6 billion of operating profit, a margin near 47%. So revenue roughly quadrupled while profit grew more than sevenfold. Almost all of the extra revenue fell straight to the bottom line.
That’s what pricing power looks like. Costs in a fab don’t quadruple in a year. Prices do, when buyers have nowhere else to go.
The beat was also bigger on profit than on sales: about 6% above consensus on revenue, closer to 9% on operating profit. Analysts got the volume roughly right and still underestimated how much SK Hynix could charge for it.
Why the Shortage Has Legs
The company’s 2030 call rests on something more concrete than optimism. SK Hynix says it has long-term agreements with 14 major buyers covering more than half of its output in 2027 and 2028.
That changes the character of the business. A memory maker selling on spot and quarterly contract prices lives and dies by the cycle. A memory maker with half its capacity committed two years out looks more like a utility or a foundry. Its customers have decided that guaranteed supply is worth more than a cheaper price later, which tells you how worried they are about running short.
There’s a physical reason the squeeze is so stubborn. High-bandwidth memory, the stacked DRAM that sits next to AI accelerators, uses far more wafer area per bit than ordinary DRAM. Every wafer turned into HBM is a wafer that doesn’t become memory for phones, laptops or servers. As AI chips need more HBM per package with each generation, capacity gets pulled away from everything else, and building new fabs takes years.
The demand side isn’t easing either. In the same news cycle Meta raised money for AI spending with Nvidia, AMD and Broadcom as cornerstone investors, Andy Jassy said AWS capacity for 2027 is largely sold out, and an industry report pointed to PCBs and chip substrates as new bottlenecks. When the buyers of compute are pre-selling capacity two years out, the suppliers of memory can do the same.
SK Hynix isn’t alone in saying so. Micron guided its November quarter to $61.5 billion of revenue with gross margin around 86 percent, numbers covered in Micron Guides Fiscal Q1 2027 Revenue to $61.5 Billion.
Handing the Windfall Back
SK Hynix is now doing something memory companies traditionally avoided at the top of a cycle. It’s returning cash in size.
The company plans more than $22 billion in dividends and about $15 billion in buybacks. That comes after a roughly $30 billion buyback it completed in August. An earlier report from Korea’s DongA Ilbo pointed to a special dividend of around $12 billion, helped by a gain of about $15 billion on SK Hynix’s stake in Kioxia.
Two readings are possible. The bullish one says management is confident enough in contracted demand that it doesn’t need to hoard cash for the next downturn. The cautious one says a company that can’t find enough fab projects to absorb this much money is admitting that capacity can’t be added fast enough, which is itself a sign of how tight things are. Either way, the old memory habit of pouring every windfall into new capacity until prices collapse isn’t what’s happening here.
China’s Catch-Up Comes at a Steep Price
The one company trying hardest to break the incumbents’ grip is CXMT, China’s leading DRAM maker. Reuters reports it plans to start HBM3E production even though yields are only 20% to 30%. Its HBM3E reportedly performs closer to SK Hynix’s previous-generation HBM3.
A yield of 20% to 30% means most of what comes off the line gets thrown away. HBM is especially unforgiving because a single bad die can ruin a whole stack, so poor yields compound. The cost per working part is a multiple of what a mature line achieves. No commercial operator would ship on those numbers unless it had a reason beyond profit.
CXMT does. Export controls have cut Chinese AI chip makers off from the leading HBM suppliers, so domestic memory, even late and expensive, is a strategic necessity. The goal is to have something rather than nothing, and to climb the learning curve on real volume.
Markets are less patient. CXMT is down about 25% in the past month, and China’s Star 50 index of AI and chip leaders has fallen roughly 40% in under four months, wiping out about $480 billion in value. Investors seem to be pricing in a long, costly road. In the near term that’s good news for SK Hynix. A serious second source would have been the main threat to its pricing, and on current evidence it’s at least a generation behind.
Everyone Else Pays
Memory doesn’t disappear when AI data centers take more of it. It gets more expensive for everybody else.
Counterpoint Research says global PC shipments fell 21% year on year in Q3 to about 57 million units, with memory and other component costs pushing prices up. Lenovo stayed in first place despite a 23% drop of its own. That’s not a normal soft quarter. It’s demand being priced out.
PC makers are responding by cutting costs elsewhere. Nikkei Asia reports that HP, Asus and Acer are moving some production from Thailand and Vietnam back to China as component costs climb, and that Dell is weighing a similar move. Over the past few years the trend ran the other way, as brands diversified out of China for political reasons. Now cost pressure from expensive parts is strong enough to reverse some of that. It’s a small detail with a big implication: memory prices are now shaping where electronics get made.
The squeeze reaches into storage too. Fujifilm plans to expand magnetic tape production capacity fivefold, its biggest production expansion in about 20 years, to meet AI-driven demand. Tape is the cheapest place to keep huge volumes of data that rarely gets read, such as training sets, archives and backups. When flash and disk are tight and expensive, cold data moves to tape. Disk is tight too: Seagate now earns half its revenue as operating profit and has most of its nearline output allocated through 2028, as set out in Seagate Guides Fiscal Q2 2027 Revenue to $4.8 Billion. Fujifilm’s bet shows that the AI build-out is pulling on every layer of the storage stack, down to technology most people assumed was obsolete.
What Could Go Wrong
None of this is guaranteed to last until 2030. Memory has fooled investors before, and the risks are real.
The biggest one sits with the customers. AI spending depends on hyperscalers, model labs and neoclouds continuing to raise and spend enormous sums. There are signs of strain. Bloomberg reports investors growing impatient for returns on AI capex ahead of earnings from Alphabet, Amazon and Microsoft. Elon Musk said debt markets have shut off credit for AI data center companies and that hundreds of neoclouds face consolidation. Apollo’s Torsten Slok argued that outside tech, companies haven’t yet seen a financial payoff from AI. If the money tightens, memory orders would be among the first things renegotiated.
Long-term agreements reduce that risk but don’t remove it. Contracts can be restructured when a buyer is in trouble, and a buyer that goes under doesn’t take delivery.
Supply is the second risk. Every memory maker is adding HBM capacity, and new fabs eventually come online. If they arrive just as AI demand flattens, the old cycle could return with force. And CXMT, for all its yield problems, will get better with volume. Chinese policy will keep funding it whether or not it’s profitable.
The third risk is the one already visible in the PC numbers. Demand destruction in consumer electronics hurts the volume business that memory makers still rely on. A 21% fall in PC shipments is a market telling suppliers that prices have gone too far for ordinary buyers.
The Bottom Line
The memory industry has always been a story of boom and bust. This week SK Hynix argued that AI has changed the shape of the cycle, and the numbers back it up. An 80% operating margin, half its output booked two years ahead and tens of billions going back to shareholders aren’t the marks of a company bracing for a downturn.
The cost of that strength is being spread around. China is spending heavily to stay a generation behind. PC buyers are paying more and buying less. Manufacturers are rearranging supply chains to absorb the hit. Even magnetic tape is having a moment.
For now the leaders hold the cards. Whether that lasts to 2030 depends less on memory than on whether the people buying AI chips keep finding the money to pay for them.