> ## Content Index > Fetch the complete content index at: https://nexi.fund/llms.txt > Use this file to discover other available public pages before exploring further. # SK hynix sold out its memory through 2029. The glut thesis has a problem. - URL: https://nexi.fund/sk-hynix-hbm4-memory-shortage-2030-2026/ - Published: 2026-10-30T12:10:49.000Z - Updated: 2026-10-30T12:10:49.000Z - Description: An 80% operating margin, 14 long-term customers and output sold out through 2029. SK hynix is betting the memory cycle broke, not paused — and NVIDIA's quieter HBM4 spec cut supports it. - Author: Nexi.fund Labs - Tags: AI & Infrastructure, #mode-6, #hook-statistic, #track-A, #layout-radar-scan KRW 83.3 trillion in operating profit on KRW 104.6 trillion of revenue. SK hynix had never printed a quarter like it. That is an 80% operating margin for a company that builds fabs, etches silicon and stacks memory dies for a living. It is also the least important number in the October 28 disclosure. 2x demand vs. available HBM supply #### The booking gap Customer requests for HBM exceed what SK hynix can ship by more than two times, per its Q3 call · *SK hynix, 3Q26* 2029 DRAM and HBM output sold out through #### The booking horizon Secured LTAs plus demand under negotiation cover three years of output · *SK hynix, 3Q26* The disclosures that matter sit three paragraphs into the earnings release. Long-term agreements now cover 14 customers, up from 10 in the second quarter. Those contracts already account for more than half of planned production capacity in 2027 and 2028. Add the demand still under negotiation, and DRAM and HBM output is effectively sold out through 2029\. Management expects the shortage to run through 2030. ## The contract book replaced the forecast Memory is a commodity business with a memory of its own. Every upcycle since the mid-1990s ended the same way: price signal, capacity response, collapse. So when a supplier says the shortage lasts five more years, the correct instinct is disbelief. SK hynix is asking buyers to disbelieve differently this time, and it brought paperwork. Typical agreements run five years. The company is discussing extensions with several of the 14 customers. More than half of 2027 and 2028 output is committed at terms that reflect HBM4's higher value. That changes who carries the cycle risk. Pre-sold capacity is not a demand forecast. It is a signed contract with a delivery schedule attached. 📊 **Four numbers from the call** Revenue KRW 104.62 trillion, up 328% year on year. Operating profit KRW 83.33 trillion, up 632%. Operating margin 80%. Net cash KRW 98.7 trillion, against KRW 15.8 trillion of borrowings. ## Why NVIDIA lowered the bar instead of the order NVIDIA did something in September that looked bearish and probably is not. The Vera Rubin specification sheet now lists 19.2 TB/s of HBM4 bandwidth per GPU, down from 22 TB/s. NVLink 6 dropped from 3.6 TB/s to 3.0 TB/s. Both changes were made quietly, without a press release. Neither came with an explanation. Read as a demand signal, that looks like a cut. Read as a supply signal, it is the opposite. Memory makers have already disclosed HBM4 pins at 11.7 Gbps from Samsung and SK hynix, and more than 11 Gbps from Micron. Rubin's original target needed roughly 10.7 Gbps per pin. The revised target needs about 9.4 Gbps. The memory was fast enough. Lowering the requirement widens the pool of modules that can qualify for Rubin, and raises the number that can be produced per wafer. > With HBM supply constrained, customers have to build more AI accelerators and systems using limited volumes.— Kim Ki-tae, head of HBM sales and marketing, SK hynix He is describing stack-lowering: fewer memory dies per cube, so more cubes are needed for the same bit volume. Reports suggest NVIDIA may use 8-Hi HBM4 in Rubin Ultra racks from the second quarter of 2027, replacing a planned 12-Hi HBM4E configuration. Fewer dies per stack is cheaper per stack. It also means more base dies, more packaging steps and more back-end work. | Vera Rubin memory spec | September 1 | Current | | ---------------------------- | ----------- | ---------- | | HBM4 bandwidth per GPU | 22 TB/s | 19.2 TB/s | | NVLink 6 throughput | 3.6 TB/s | 3.0 TB/s | | Pin speed needed to hit spec | \~10.7 Gbps | \~9.4 Gbps | ## The bear case was not cancelled The glut thesis has real evidence behind it. Memory makers are building fabs. Governments are subsidising domestic capacity. China's CXMT keeps appearing in analyst notes as the eventual price breaker. SK hynix's answer is timing, not denial. Leading-edge process steps are getting harder. New fabs take years. The company expects the imbalance to persist because supply cannot move as fast as demand. Its own numbers keep some scepticism alive. DRAM average selling prices rose by a high-20s percentage in the quarter. Pricing is doing as much work as volume. Then there is realizable demand, the company's own term for orders customers can actually pay for once financing and power constraints are counted. Management guides 2027 DRAM growth in the mid-20s and NAND in the high teens. That is roughly where 2026 lands. Growth, not acceleration. ## The power constraint behind the demand curve Cloud providers are forecast to raise capital spending more than 60% in 2027\. Secured power capacity is expected to grow more than 100% year over year. Those figures, cited from market research on the call, explain why memory demand is not levelling off: the buildings are already being energised. SK hynix projects the global AI data center infrastructure market growing at about 40% a year through 2030\. If that holds, the shortage is arithmetic. ## What HBM4 actually changed HBM4 entered mass production and shipment in the third quarter, on schedule for 12-Hi parts. Management expects it to overtake HBM3E in shipment volume and revenue share in the fourth quarter, little more than a year after HBM3E's own peak. Yields beat plan, and the company raised its target. HBM4E samples shipped in the second quarter. DRAM bit shipments grew about 10% sequentially. Server DDR5 shipments grew more than 40%. HBM demand, on the company's own account, runs at more than twice what it can supply. An 80% operating margin is a strange thing to see next to a 38% net margin. Most of the gap is below the operating line: KRW 11.6 trillion of foreign-exchange losses and KRW 22.9 trillion of valuation losses on investment assets, for KRW 32.6 trillion of net non-operating expense. ## What an 80% margin buys Cash and equivalents reached KRW 114.5 trillion. Borrowings fell to KRW 15.8 trillion. Net cash sits at KRW 98.7 trillion and the debt-to-equity ratio is 5%. The company will spend it on capacity first. M15X in Cheongju runs to full capacity by year-end to lift HBM4 output. The 321-layer NAND process is targeted at more than half of domestic capacity in the fourth quarter. Shareholders get roughly KRW 100 trillion this year: KRW 30 trillion in dividends and KRW 60 trillion in buybacks and cancellations, including the KRW 40 trillion announced in August. The Solidigm IPO is shelved, for now. ## China, and the price of being wrong Chosun Biz reported that SK hynix used the call to push back on CXMT as a near-term threat. Its position on 4F² DRAM is that commercialisation is too early, and it is staying on 6F² for now. That is a confident answer from a company with 5% debt-to-equity and KRW 98.7 trillion of net cash. It is also the kind of answer that ages badly if the cycle turns. Six consecutive record quarters. An 80% margin. A customer book that runs to 2029. And a market that still prices memory as a cycle. ## The overproduction question An analyst on the call asked the obvious question. SK hynix keeps building fabs, Yongin included. What stops this from becoming the same overbuild that broke every previous memory cycle? The answer was a policy, not a number. The company will prepare supply infrastructure as planned, then ramp actual capacity in line with demand and profitability. If conditions diverge from plan, it will flex investment disbursement, production schedules and resource allocation. Read that carefully. Capital committed to buildings and cleanrooms is hard to reverse. Tool installs and wafer starts are not. SK hynix is separating the two, so the part it cannot undo stays cheap and the part it can undo stays late. That is a reasonable defence, and it is unverifiable until the cycle turns. Every memory maker in history has promised to be disciplined about capacity. Several meant it. The ones that did not still shipped. The difference this time is visible on the balance sheet. KRW 98.7 trillion of net cash and a 5% debt-to-equity ratio mean SK hynix can absorb a downturn that would previously have forced it to keep fabs running flat out to service debt. Discipline is easier when you do not owe anyone. There is also a version of 2027 where the 8-Hi shift cuts the value of each stack faster than it expands the number of stacks. Eight dies instead of twelve is roughly a third less memory per cube. If customers buy cubes rather than bits, revenue per accelerator falls even as unit counts rise. SK hynix says it plans around bit volume, and that base dies and back-end processing absorb the difference in work. That is the hinge. Either this market pays for bits or it pays for packages. The guidance assumes bits. None of this resolves the argument. It moves the burden of proof onto the bears, who now have to explain why 14 signed customers, more than half of 2027 and 2028 capacity committed, and a 13% lower bandwidth bar at NVIDIA all fail to hold prices. They might. Memory has broken better stories. ## The numbers that will be tested Every claim here has a shelf life. The 2029 sold-out figure depends on customers taking delivery, not just signing. The 2030 shortage depends on fabs that do not exist yet finishing on time. Watch four things: - 12-Hi HBM4 qualification volume at NVIDIA through the fourth quarter - Whether the 8-Hi HBM4 decision for Rubin Ultra lands in the second quarter of 2027 - Whether LTA coverage of 2029 capacity moves above half - Whether DRAM contract prices keep rising once the current quarter's high-20s increase annualises 🎯 **The bottleneck moved** The constraint in AI infrastructure is no longer wafer capacity. It is contract capacity — who has already signed for the bits that will exist in 2028\. ## Sources [ SK hynix announces 3Q26 financial results The primary release: revenue, operating profit, margin, net cash, LTA coverage and the 2030 shortage guidance. SK hynix Newsroom ](https://news.skhynix.com/en/q3-2026-business-results/?ref=nexi.fund) Company filing, October 28, 2026 — the anchor for every figure above. [ SK hynix Q3 2026 earnings call transcript Where the stack-lowering answer and the capex and power forecasts were said out loud. Quartr via Yahoo Finance ](https://finance.yahoo.com/quote/000660.KS/earnings/000660.KS-Q3-2026-earnings%5Fcall-717350.html?ref=nexi.fund) Full call transcript, October 28, 2026. [ Record 80% operating margin, output sold out to 2029, 14 long-term customers Independent reconciliation of the LTA count, ASP moves and the HBM4 versus HBM3E crossover. TrendForce ](https://www.trendforce.com/news/2026/10/28/news-sk-hynix-3q26-opm-hits-record-80-says-dram-hbm-effectively-sold-out-to-2029-14-lta-customers-signed?ref=nexi.fund) Useful cross-check on the pricing and contract details.