Memory Split in Two: SK hynix's 80% Margin Is an HBM Mix Story, Samsung's Record Profit Is a Volume Story

On October 28, SK hynix closed up 11% in Seoul after reporting operating profit of 83.3 trillion won on revenue of 104.6 trillion won — an 80% operating margin and an operating profit roughly eight times the LSEG estimate of 76.7 trillion won. Three weeks earlier, Samsung Electronics had guided to a record 107.4 trillion won of operating profit, beat the LSEG estimate of 106.1 trillion won, missed the higher Bloomberg consensus of 108.67 trillion won, and fell 2.4% while the KOSPI dropped 2.6%. In Seoul trading on October 29, with the full results out, Samsung was down roughly 4.8% and the index 2.4% lower.

The two companies sold into the same shortage, for the same buyers, at the same point in the AI capital-expenditure cycle. That is what makes the split a mix story rather than a demand story.

Note what is not in either release. SK hynix does not disclose HBM revenue separately, and Samsung does not disclose HBM4's share of its own HBM revenue, so the mix gap has to be measured outside both companies — which is exactly what the market did with its share prices.

The second-order read is more important than the headline beat. SK hynix's operating profit rose 38% quarter on quarter while its net profit fell 58%, from 93.9T won in the June quarter to 39.8T won, because non-operating items swung from a 62.2T won gain to a 32.6T won charge. Operating cash flow over the trailing twelve months is 127.2T won; net income over the same window is 189.2T won. Neither company itemizes that swing in the headline release, and it is the reason the tape treats SK hynix's earnings as a cash question rather than an earnings-per-share question.

The supply chain did not hand SK hynix more demand than Samsung. It handed SK hynix more of the part of demand that carries a premium. In Q2 2026, SK hynix held 25% of the conventional DRAM market but 50% of HBM; Samsung held 38% of DRAM and only 33% of HBM; Micron held 24% of DRAM and 18% of HBM. Samsung sold the most bits. SK hynix sold the most value per bit of the scarce stack.

That is the whole divergence: Samsung's record was a volume record at commodity DRAM prices, and commodity DRAM pricing is decelerating into Q4.

There is a second reason the mix gap widens from here. HBM4 stacks require more bonding steps, more known-good-die testing and more advanced packaging than HBM3E, so each incremental stack consumes more of the constrained parts of the supply chain. SK hynix has said HBM demand is running at twice what it can supply, and that its DRAM and HBM volumes through 2029 are effectively sold out under a combination of long-term agreements and customer discussions, with deposits used to turn customer intentions into firm orders.

A memory manufacturer now converts 80 won in revenue into 80 won of operating profit, while the company designing the accelerator that consumes the memory earns a 66.2% operating margin and the company assembling servers earns 11.7%. In a cycle where the binding constraint is packaging and qualified HBM stacks rather than logic wafers, the scarce rent has migrated to the layer that owns the stack, and it is being paid to the supplier with the most HBM-weighted revenue, not the one with the most bits.

Samsung's own disclosure makes the point sharper. Device Solutions produced 106.5T won of operating profit on 147.1T won of revenue in Q3, while everything outside it — displays, mobile, appliances — turned 48.2T won of revenue into roughly 0.9T won of operating profit. Samsung is no longer a conglomerate that also owns memory; for this cycle it is a memory company with a very large, very low-margin tail attached.

SK hynix's 83.3T won of operating profit was the highest in its history and beat the estimate by nearly9%. The same release carried a net profit of 39.8T won, down 58% from 93.9T won in the June quarter, because non-operating items moved from a 62.2T won gain to a 32.6T won charge. Neither the release nor the earnings call items the driver of that roughly 95T won swing, and the gap between the two lines is the single largest unexplained item in the print.

The cash statements are the cleaner read, and they are the reason the company is returning roughly 9% of its market value in a single year.

Here is the tension that explains why a record quarter did not re-rate the stock. Trailing EV/EBITDA of 4.9 times and a 16.4% trailing earnings yield are the kind of multiples you apply to peak-cycle cash, and the market has been cutting SK hynix's market value by about 41% from its June 30, 2026 level while profits tripled. The market is not disputing the cash. It is refusing to capitalize it.

The payout is part of that refusal. Returning 100T won in one year, roughly 9% of market value and well above the 91.2T won of trailing free cash flow, converts a cyclical windfall into a one-off distribution rather than a compounding franchise — while 2027 capex is set to rise significantly for the Yongin ramp and inventory days have stretched to 189. The company is funding a capacity build and a dividend out of the same peak-cycle cash flow at the same time.

If HBM4 stacks are the scarce product, the money moves to whoever supplies the stacking. SK hynix placed a 44.2 billion won order with Hanmi Semiconductor on June 8, 2026 for TC Bonder 4.5 Griffin thermal-compression bonding equipment, with delivery, installation and test operation completed by September 2 at SK hynix's post-package and test facility in Cheongju. Those bonders are the specific step that turns a stack of dies into a sellable HBM4 module, and SK hynix is the only supplier with volume orders in that equipment class.

The pull-through is visible across the chain, and it is visible in reported numbers rather than in forecasts.

What is not disclosed: neither SK hynix nor Samsung publishes an HBM revenue split, so the mix gap is an estimate built from market-share data and management commentary rather than a reported line item. Samsung does not disclose foundry profitability inside Device Solutions, and Hanmi Semiconductor does not disclose how much of its order book is HBM4 bonding rather than other packaging work. Anyone treating the HBM share gap as audited revenue is over-reading it.

The same mix that pays the suppliers is paid for by the buyers of AI capacity, and TrendForce puts the scale of it precisely: DRAM and NAND together take 47% of cloud service provider capex in 2026 and 68% in 2027, with total cloud capex rising 98% in 2026 and a further 50% in 2027. HBM and RDIMM together already account for 51% of DRAM bit supply in 2026, and server DRAM plus HBM bit supply is forecast to rise 27% in 2027.

For AI server economics, this is the number that matters: the accelerator is no longer the largest line item in the build. That does not show up yet in the gross margin of the chip designer, because HBM costs are being absorbed inside a 75% gross margin and passed into system pricing. It shows up in the operating margin of the companies that assemble the box.

Micron is the cleanest available read on the same quarter, because it is the third supplier with the least HBM weighting of the three. It reported fiscal Q4 2026 revenue of 54.23 billion dollars on Sep 3, 2026, net income of 37.70 billion dollars and an 86.8% gross margin, and guided fiscal Q1 2027 revenue to 61.5 billion dollars with gross margin near 86%.

Micron also said the thing that matters more than the numbers: fiscal Q1 2027 is the floor for gross margins for the year, and HBM revenue grew faster than total company revenue in the September quarter. It is also guiding fiscal 2027 capex higher than prior plans, with the majority of the increase in construction capex for cleanroom space landing in late calendar 2028 and beyond.

Micron is therefore the read-through with the most duration and the least mix. It has the longest contracted visibility of the three, because it sold75% of 2027 output earlier than SK hynix sold more than half of 2027-28, and because its new clean rooms do not add meaningful bits until 2028. The reason its HBM share is lower is timing: its HBM4 ramp is on track but its advanced packaging output only starts in early calendar 2027, and it does not begin volume production of the next generation until the second half of calendar 2027.

The near-term calendar is unusually well defined, and it is a test of mix rather than of demand. If HBM4 really does overtake HBM3E in revenue at SK hynix in the December quarter, the ASP tailwind inside a 39% DRAM price increase becomes visible in the segment split, and the market can finally mark the difference between HBM revenue and DRAM bits. If it does not, then a decelerating commodity curve is doing more of the work than the release implies, and the 4.9 times trailing EV/EBITDA stops looking like a discount.

On the 2027-28 horizon the question is no longer pricing but capacity, and the answer currently favours the suppliers. SK hynix has DRights and HBM volumes sold out through 2029, Micron has 75% of 2027 committed and no meaningful new clean rooms before 2028, and Samsung's own expansion plans are landing into a market that will be tighter in 2027 than in 2026. Samsung and Micron expect the shortage to persist into 2028.

Three things are simply not disclosed, and they are the gaps a reader should keep open. Samsung has not published HBM4's share of its own HBM revenue, so the 33% HBM share figure is a market estimate. SK hynix has not published its HBM revenue split either, so the mix argument rests on management commentary and third-party share data rather than a reported line. And neither company itemizes the roughly 95 trillion won swing in non-operating items that turned a record operating quarter into a 58% drop in net income at SK hynix.

The first clean divergence in the memory complex is a statement about which company owns the scarce product, not about which one owns the AI cycle. Both sold into the same shortage. SK hynix earned 80% operating margins on a business that is now 50% of HBM by market share and has more than half of its 2027-28 capacity under long-term agreements with 14 customers; Samsung earned a larger absolute profit on a business that is 33% of HBM, and its non-chip businesses turned 48.2 trillion won of revenue into less than 1 trillion won of operating profit while commodity DRAM contract prices decelerate from 60% growth to 10-15%.

For investors the practical read is that memory has stopped being one trade. The premium is now collected by HBM-weighted suppliers and by the bonding, test and packaging chain that gates HBM4 supply — Hanmi Semiconductor's bonder orders, Lam Research and Applied Materials' DRAM-driven tool revenue, TSMC's packaging, Advantest's test content. The cost is collected from the buyers: cloud capex with memory rising from 47% to 68% of the budget, server assembly at 11.7% operating margin, PCs at 5.7%. And Micron, with the longest contracted visibility and the least HBM weighting, is the cleanest way to own the cycle without owning the mix.

What the divergence does not do is re-rate anything. SK hynix's market value has fallen about 41% since June 30 even as operating profit tripled, because100 trillion won of dividends and buybacks and a rising 2027 capex program convert peak cash into a distribution and a capacity bill rather than a compounding franchise. The trade here is contract structure and mix quality over the next eight quarters — and the 2027 HBM book is already written.