SK Hynix's costs, leverage, and underlying fundamentals of front-end and back-end operations
I'm LongbridgeAI, I can summarize articles.Why you shouldn't look at book profit for $SK Hynix(SKHY.US)
To determine if SK Hynix's profitability has peaked, many people's first instinct is to check the net profit attributable to shareholders on the last line of the income statement. This approach doesn't work.
In its most recent quarter, the Q3 2026 report released on October 28, SK Hynix saw single-quarter revenue surge to 104.62 trillion KRW, with operating profit hitting 83.33 trillion KRW and an operating margin of 80%. This profitability has left TSMC in the dust.
However, looking at the bottom line, Q3 net profit was only 39.78 trillion KRW, a 58% QoQ plunge from 93.92 trillion KRW in Q2. The net margin dropped from 118% in Q2 to 38%. Many investors, seeing profits halved, combined with rampant market rumors about price cuts, order grabbing, and TSMC price hikes, immediately concluded that the cycle had peaked and costs were spiraling out of control.
This understanding is too superficial. The reason Q2 net margin hit an absurd 118% was due to 62.17 trillion KRW in non-operating net income recognized on the books, mostly from fair value revaluation and partial divestment gains on SK Hynix's Keenex holdings. Just this one item accounted for 63.27 trillion KRW. In Q3, Keenex's stock price pulled back significantly, wiping out a large chunk of paper gains, compounded by FX gains/losses and standard income tax provisions, creating the illusion of a near-60% collapse in net profit.
Remove the distraction of Keenex's stock volatility and look only at the true manufacturing ledger of the wafer fab, and the numbers tell a completely different story. In Q2, SK Hynix revenue was 79.32 trillion KRW, with total operating costs (COGS + R&D + SG&A) totaling 18.78 trillion KRW. By Q3, revenue surged by 25.30 trillion KRW, while total operating costs increased by only 2.50 trillion KRW to 21.28 trillion KRW.
In other words, for every extra won of revenue SK Hynix generated in Q3, 90 cents flowed directly into operating profit, resulting in a marginal operating margin of 90.1%.
This leads to the first question we need to explore further down the article. Why does a heavy-memory storage manufacturer achieve such exaggerated profit elasticity compared to top-tier software companies? If the cycle turns downward, will this heavy asset base and debt load drag the company down as it did in 2023? To understand this, we must first break down SK Hynix's cost structure and two layers of leverage.
Where does the 90% marginal profit come from?
Before calculating industry-wide figures, we need to thoroughly explain the cost logic of wafer fabs. If you continue to view memory manufacturers through the lens of assembly plants that buy parts and earn processing fees, all subsequent calculations will be wrong.
Let's break down SK Hynix's 21.28 trillion KRW total operating cost in Q3 into four parts.
First, depreciation and amortization. According to data from the Q3 earnings call, D&A was 4.3 trillion KRW, only slightly up from 4.02 trillion KRW in Q2, accounting for about 20% of total operating costs. As long as cleanrooms are powered on and maintained at constant temperature, and lithography and etching machines run 24/7, this cost remains largely fixed each quarter.
Next are the cash variable costs of wafer fabrication, including silicon wafers, photoresist, specialty gases, facility electricity, and packaging materials. This category sits around 9 to 10 trillion KRW. As long as the wafer fab is running at full capacity, whether the resulting DRAM modules sell for $2 or $20 on the market, the water, electricity, and chemicals consumed per wafer remain virtually identical.
Then there are daily R&D and SG&A expenses, which also follow a steady quarterly rhythm.
Finally, the truly variable items tied to profit, primarily employee performance bonus pools and some outsourced foundry fees. SK Hynix currently accrues a bonus pool of about 9% of operating profit, bringing R&D and SG&A together to roughly 6 to 7 trillion KRW.
Laying these four components side-by-side reveals the formidable operating leverage of memory manufacturers. When production lines are fully loaded, most costs are already locked in. Every downstream price increase, after paying employees slightly more bonuses and outsourcing partners slightly more processing fees, flows almost entirely into pure profit. Conversely, calculating the safety cushion similarly, reversing from Q3's cost base, even if future prices crash, SK Hynix's single-quarter revenue would need to plummet 78% from current levels before operating profit truly falls below the breakeven point.
At this point, someone might ask: operating leverage looks good when going up, but can be deadly when going down. SK Hynix's capex is nearing 50 trillion KRW this year, and it is building new factories in Cheongju and Yongin. If depreciation pressures ease later, combined with heavy asset borrowing, could profits be instantly eaten away? To answer this, we need to check its balance sheet and clearly calculate the second layer of financial leverage and depreciation accounts.
Will 50 trillion KRW in capex eat up profits?
Veteran stock investors fear the memory cycle essentially because they fear a repeat of the 2023 scenario. Back then, SK Hynix lost several trillion KRW in cash each quarter, burdened by net debt equivalent to $15.9 billion. High operating leverage叠加 high financial leverage meant it could only borrow new debt to repay old debt.
But checking the balance sheet at the end of Q3 this year, SK Hynix's financial leverage has turned completely negative. As of late September, cash and cash equivalents on the books rose to 114.5 trillion KRW, while interest-bearing debt decreased by 2.8 trillion KRW QoQ, leaving only 15.8 trillion KRW, representing just 4.0% of total assets.
After deducting all borrowings, SK Hynix holds 98.7 trillion KRW in net cash. Its Q3 operating cash flow was 6.1 times its capex. The most致命 debt burden of the past has been cleared.
Next, let's calculate the 'depreciation cliff' favored by bears. SK Hynix raised its 2026 capex guidance to 45-49 trillion KRW. Let's assume the worst case and calculate based on a full 50 trillion KRW spend for the year.
The investment structure of wafer fabs is transparent: about 30% is for buildings and cleanroom civil engineering, depreciated over 25 years; about 70% is for core production equipment, depreciated over 5 years even under the most conservative estimates among Korean peers. Breaking down 50 trillion KRW: 15 trillion for civil engineering results in 0.6 trillion annual depreciation; 35 trillion for equipment results in 7 trillion annual depreciation. Total new annual depreciation amounts to 7.6-8 trillion KRW, translating to just 2 trillion KRW per quarter.
Placing this 2 trillion KRW quarterly new depreciation into SK Hynix's Q3 revenue of 104.62 trillion KRW accounts for less than 2%. More bluntly, the 83.33 trillion KRW operating profit SK Hynix earns in a single quarter is enough to cover all future capex for the next two years, and it can still set aside approximately 100 trillion KRW for dividends and buybacks in 2026 alone,注销ing over 3.3% of total shares in the second half of the year.
Once internal company accounts are settled, a more core industrial question arises. Since SK Hynix is earning an 80% operating margin, why do savvy North American cloud giants willingly endure such high premiums instead of squeezing prices or switching to cheaper memory solutions? The answers lie in the next two sections: first, how cloud giants calculate their total cost, and second, how much front-end wafer capacity has actually been siphoned off.
Why must cloud giants buy expensive HBM?
Many people view memory through the lens of consumer electronics, assuming memory is a standard commodity. They argue that HBM4 selling for over $1,000 is too expensive, and cloud providers will surely find ways to use less or downgrade. This mindset fails to grasp the basic logic of AI model cost accounting.
Let's examine this from three perspectives: bandwidth, power consumption, and comprehensive cost.
First, bandwidth. In recent years, GPU compute core speeds have increased so rapidly that memory transfer speeds cannot keep up. The bottleneck for LLM training and long-context inference is actually VRAM bandwidth—the amount of weights and cache data fed to compute units per second. Mass-produced HBM4 doubles the interface bit-width from the previous generation's 1024 bits to 2048 bits. If you refuse to buy high-bandwidth HBM, top-tier GPUs costing $30,000-$40,000 will sit idle waiting for data.
Second, power consumption. The biggest constraint for large data centers globally is not money, but power quotas. HBM is attached directly microns away from the GPU via advanced packaging, consuming a fraction of the energy required to move the same data volume compared to externalized standard memory.
Finally, cost—this is what cloud giants truly calculate. Cloud providers never look at the price of a single HBM stack; they look at the total cost of ownership (TCO) per token generated, amortizing the entire server rack's depreciation, five years of electricity, data center rent, and operations & maintenance.
In a latest-generation NVIDIA Rubin architecture server, a single GPU paired with 8 HBM4 stacks incurs VRAM costs of approximately $8,000-$10,000. However, the comprehensive investment per card, including GPU, high-speed networking, power racks, and five years of electricity, exceeds $50,000-$60,000. If you save $3,000 on HBM by using inferior products, causing the whole card's token throughput to drop by 30%, the amortized comprehensive cost per token ends up being significantly higher. No purely commercially driven cloud giant would fail to see this math.
This explains why HBM commands ultra-high gross margins, but it doesn't fully explain SK Hynix's 80% overall operating margin. After all, HBM shipment volume accounts for less than half of SK Hynix's total DRAM. Why did average selling prices for traditional server DDR5 and even nearly obsolete DDR4 surge 198% YoY in Q2? We need to go upstream to the front-end wafer process and calculate the black hole account of global wafer capacity.
How much front-end wafer capacity has been siphoned off?
In this section, we break down global front-end DRAM capacity. First, a disclaimer: specific production schedules for each original manufacturer's line are trade secrets, so numbers cannot be precise to the unit. However, the physical monthly capacity of global wafer fabs and chip area are undeniable, allowing us to converge on the magnitude.
Let's look at the global baseline. By the end of 2026, the total monthly 300mm DRAM wafer capacity of the three major originals is approximately 720k for Samsung, 590k for SK Hynix, and 375k for Micron. Combined, the global monthly plate is only 1.685 million wafers.
HBM drains this plate because front-end manufacturing imposes a harsh wafer consumption penalty coefficient. Producing one HBM core die requires drilling thousands of vertical silicon vias through the silicon for interconnects, plus leaving physical isolation zones for crack prevention and massive I/O circuits. A 1GB HBM die occupies significantly more silicon area than standard DDR5, and multi-layer stacking introduces yield losses.
During the HBM3E phase, producing 1GB of capacity consumed approximately 3x the wafer area of standard DDR5. With this year's 2048-bit HBM4, accounting for yield losses, the wafer area consumed per GB has expanded further to 3.5x-4x that of DDR5.
With this coefficient, we can convert totals. Currently, 20% of the three majors' DRAM wafer capacity is allocated to HBM, moving towards 30%. Taking the midpoint at 25%, roughly 420k wafers globally per month are used for HBM. Converting this using a 3.5x+ consumption coefficient, if these wafers had produced standard server memory instead, it would equate to consuming nearly half of the global effective output of traditional DRAM.
Thus, the near-doubling of ordinary memory prices this year, and even the奇观 of DDR4 prices surpassing DDR5 at one point, is not due to collusion and hoarding by manufacturers, but because all wafers were sucked into the HBM black hole. SK Hynix's 80% Q3 operating margin is the combined result of HBM's high premium叠加 surging traditional DDR5 and enterprise SSD prices.
Can front-end wafer capacity be replenished immediately by throwing money at it? The answer is no.
Building a large memory cleanroom from land acquisition to line opening and ramp-up takes at least two to three years. SK Hynix signed a 11.95 trillion KRW EUV lithography machine contract with ASML in March this year; delivery for these machines is scheduled until the end of 2027. Lam Research, Tokyo Electron, and Applied Materials' high-aspect-ratio silicon via etch and deposition equipment also face delivery delays exceeding one year.
Looking at SK Hynix's own schedule makes this clearer. The Cheongju M15X factory is pushing to max capacity by year-end, adding only about 50k monthly wafers. The Yongin Y1 super base, planned with 6 cleanrooms each at 60k monthly capacity, won't start equipment installation until early 2027 at the earliest. For the next year-plus, front-end physical supply is essentially locked in the open.
With no flaws in the front-end, short-selling logic naturally focuses on back-end packaging and external foundries. The most prevalent rumor recently is that SK Hynix outsourced the bottom-most logic base die for HBM4 to TSMC. TSMC charges 3-4x more and continues to raise prices, allegedly turning SK Hynix into a mere assembler for TSMC. Is this foundry bill really like that?
Will TSMC foundry fees for base dies drain gross margins?
The structural aspect of this claim exists, but the numbers are purely designed to scare people with multipliers.
The correct part is the business structure. In HBM3E and earlier, the Base Die at the bottom, responsible for signal control and data scheduling, was made by SK Hynix using its own DRAM lines. In the HBM4 era, with interface bit-width doubling to 2048 bits, NVIDIA and cloud giants require custom logic circuits embedded in the base die. SK Hynix, lacking advanced logic foundry capabilities, must indeed outsource base die production to TSMC using 12nm and 5nm processes, before sending them to TSMC's CoWoS line for hybrid bonding. Industry surveys confirm that TSMC-foundry base dies cost 3-4x more than SK Hynix's self-produced ones.
The incorrect part is focusing only on the multiplier while ignoring the absolute dollar base. Many hear costs quadrupled and assume gross margins will collapse, but are too lazy to calculate exactly what this base die is worth in dollars.
Let's calculate using the most extreme assumptions. Previously, SK Hynix made a base die using its own DRAM process for $6-$8. After switching to TSMC foundry, let's ignore the cheaper $5,000/wafer 12nm wafers and assume all use $15,000/wafer TSMC 5nm wafers.
A HBM4 logic base die is approximately 110 sq mm. A 300mm wafer, accounting for edge loss and yield, yields about 450-500合格 chips. This translates to a single-wafer foundry cost of $30-$33. Pushing to the extreme, including front-end IP licensing amortization and future TSMC price hikes, let's hard-cap it at $45 per die. Note that this is already a massively overstated figure.
In other words, SK Hynix pays TSMC an additional ~$35 per HBM4 stack. How much more does SK Hynix charge customers? A 12-layer 36GB HBM3E previously sold for ~$500-$600, while customized 12-layer HBM4 single-die prices jumped to $1,000-$1,300.
Paying an extra $35 toll fee brings back $500-$600 in price increases. Even if TSMC raises 5nm foundry fees by another 20% next year, each base die costs only $7-$8 more. Against a $1,200 HBM4 selling price, the impact is just 0.6%. The claim that TSMC drains SK Hynix's gross margins collapses with simple addition and subtraction.
If TSMC isn't taking the bulk of the profit, what about the back-end packaging technology routes and equipment? There are two more rumors: one claims that at the 16-layer HBM4 stage, existing thermocompression bonding can't hold, requiring a switch to bumpless copper-to-copper hybrid bonding, rendering SK Hynix's old工艺 advantages zero; another focuses on patent lawsuits involving Korean equipment makers, claiming SK Hynix's thermocompression bonders will face supply cutoffs. Will these two back-end gates suddenly change weather?
Will back-end packaging routes and equipment be choked?
To answer this, we must know why SK Hynix has been beating competitors on HBM yield rates these past two years. SK Hynix maintains a comprehensive yield rate of around 80% for HBM3E and HBM4 thanks to its exclusive MR-MUF (Mass Reflow Mold Underfill) batch process. It stacks over a dozen thin chips, performs a single heating reflow soldering, then injects liquid epoxy mold compound supplied by Japanese firms like Namics into the gaps. Compared to competitors' early layer-by-layer application of non-conductive film thermocompression, batch reflow offers superior heat dissipation and single-tool output efficiency.
First, will hybrid bonding replace it immediately? This confuses lab viability with commercial mass-production economics.
Hybrid bonding requires direct atomic-level alignment of copper contacts between upper and lower layers, demanding sub-nanometer wafer surface flatness. Any micron-sized dust particle falling in scraps the entire multi-layer stack. Cleanroom requirements and polishing costs are nearly as expensive as front-end lithography. With JEDEC standards relaxing HBM4 height limits to 775 microns and discussing 900 microns for the next gen, SK Hynix can easily fit 12-layer and 16-layer HBM4 using improved MR-MUF, with no need to risk a massive yield drop by forcing hybrid bonding.
In April this year, SK Hynix did spend 20 billion KRW on a hybrid bonding tool jointly developed by Applied Materials and BESI. However, a regular thermocompression bonder costs 3 billion KRW. 20 billion KRW isn't even enough for 7 regular machines, clearly indicating this purchase is merely for R&D lines to probe HBM5 20+ layers in the future.
Regarding rumors of back-end bonder supply cutoffs: Korean equipment maker Hanmi Semiconductor historically held 71.2% of the global HBM thermocompression bonder share, tightly bound to SK Hynix. But this year, SK Hynix proactively dismantled single-source dependency. In June, it placed a 44.2 billion KRW (~15 units of HBM4 equipment) order with Hanmi while simultaneously elevating ASMPT's fluxless thermocompression bonder, allocating nearly half the incremental orders to it, and introducing Hanwha Semitech as a third备选.
Previously a sole supplier, now three equipment vendors compete to supply SK Hynix, making SK Hynix's bargaining power and security in back-end equipment more stable than in previous years.
So here's the question: since front-end capacity is locked, and back-end processes and equipment are secure, why does market statistics show SK Hynix's global HBM revenue share dropping from 64% a year ago to 58% in Q1 this year, and further to 50% in Q2, while Samsung surged from 21% to 33%? Did Samsung steal existing orders?
Did the share drop to 50% because Samsung stole orders?
Focusing solely on market share percentages without looking at wafer fab utilization rates easily leads to the erroneous conclusion that SK Hynix's competitiveness is declining.
Comparing two sets of objective facts clarifies this. While SK Hynix's HBM market share dropped from 64% to 50%, its Q3 total revenue surged 32% QoQ. Management explicitly guided that Q4 HBM4 shipments and revenue will both exceed HBM3E. Furthermore, all HBM capacity for the three majors in 2027 is already fully booked by customers.
Why did share drop, yet revenue and orders exploded? Because the downstream大盘 expanded too fast. The current order fulfillment rate for HBM4 across the industry is only 60%-70%.
SK Hynix's own M16 and M15X lines have reached physical limits per wafer; it simply cannot produce more. The share drop from 64% to 50% isn't because old customers switched to Samsung; it's because the 30%-40% overflow demand that SK Hynix's capacity couldn't feed finally found a buyer in Samsung once its HBM4 yields ramped up.
There's also a frequently misread detail. In Q2, among the three majors, SK Hynix's overall DRAM ASP QoQ growth lagged slightly behind Samsung and Micron. Some cited this data to claim SK Hynix was cutting prices to protect share.
The real reason is SK Hynix has the highest HBM proportion. Its HBM is covered by long-term supply agreements lasting five years with about ten key clients, including prepayments and pre-locked price bands. Samsung and Micron, having lower prior HBM proportions, had a larger share of ordinary server memory exposed to the spot market, which surged nearly two-fold. The long-term agreement mechanism caused SK Hynix to miss out on some quick cash during the spot market frenzy, but in exchange, it secured extremely high order certainty for the coming years and a price cushion during cyclical downturns.
Having deconstructed Samsung's追赶 logic, the remaining external ghost stories focus on geopolitical chokepoints and domestic memory expansion. US restrictions on upgrading advanced equipment at the Wuxi fab, coupled with domestic players like ChangXin Memory pushing server DDR5 and higher-density 4F² architectures by year-end—what is the actual杀伤力 of these two risks within SK Hynix's total revenue denominator?
How significant are Wuxi restrictions and domestic expansion?
First, calculate the Wuxi exposure. Many hear that nearly 40% of SK Hynix's DRAM wafer capacity was once located in Wuxi, China, and assume US tightening of equipment exemptions will cripple SK Hynix.
But capacity and profit contribution are not the same thing. Since the US banned EUV lithography machines from entering China years ago, SK Hynix's Wuxi C2 fab has zero EUV equipment. Production lines remain on mature nodes like 1z and 1a nm for traditional memory.
The core assets driving over 75% of SK Hynix's DRAM revenue and contributing 80% of its operating margin—including all HBM lines, 1b/1a nm high-end server DDR5, and 321-layer enterprise SSDs—are 100% located in SK Hynix's local facilities in Icheon and Cheongju, Korea. Even if future equipment updates at Wuxi are further restricted, the impact will only be on mature consumer-grade products with shrinking profit margins.
Regarding domestic memory追赶: Domestic leaders are indeed发力 on server DDR5 modules and 4F² new architectures by year-end. The author never underestimates the攻坚 speed of Chinese engineers. Coupled with low baselines in domestic government clouds and local server markets, and strong domestic substitution demands, domestic memory shipment growth rates will certainly be very high in the coming years.
But qualitative progress and quantitative share must be viewed separately. The HBM4 used in overseas cloud giant core AI clusters faces bottlenecks not in single-chip manufacturing, but in front-end EUV lithography, yield of thousands of ultra-deep silicon vias, back-end MR-MUF packaging, and deep integration with TSMC CoWoS lines and NVIDIA GPU underlying physical design. Before 2028, domestic memory is unlikely to cut into overseas AI core computing chains. The real impact surface is concentrated on domestic consumer PCs, mobile phone memory, and mid-to-low-end general-purpose server DRAM.
Of course, as buyers, we cannot only report good news. Looking ahead to 2028-2029, when SK Hynix's Yongin new plant, Samsung's Pyeongtaek P5, and new domestic capacity all come online, ordinary memory prices halve from highs, and HBM excess margins compress as supply from three majors becomes ample, can SK Hynix's performance and current stock price withstand it?
How much profit remains if margins are cut in half?
Let's state the ugly truth upfront: the 80% operating margin in Q3 is absolutely not a long-term steady state for the memory industry. It contains暴利 from traditional memory shortages triggered by HBM siphoning. If anyone simply multiplies Q3's 83.33 trillion KRW operating profit by four to claim SK Hynix earns over 330 trillion annually with a 3x P/E, they lack basic reverence for manufacturing cycles.
The prudent approach is to calculate a post-bubble bottom-line figure for SK Hynix assuming all ghost stories materialize in a bearish scenario.
Assume by 2028-2029, new wafer capacity is fully released, slashing booming traditional DRAM and NAND prices by half, while HBM excess margins also compress significantly. Under this extreme bearish assumption, we slash SK Hynix's single-quarter revenue from Q3's 104.62 trillion KRW by 40%, reverting to an annualized 250 trillion KRW level. Simultaneously, we cut the operating margin from 80% by more than half, pressing it to a 35% median. If you feel 35% isn't conservative enough, adjust downwards to 30% yourself.
Calculate how much money remains under this revenue-cut-by-40%, margin-cut-in-half scenario. 250 trillion KRW annual revenue multiplied by 35% operating margin yields 87.5 trillion KRW annual operating profit. After taxes, steady-state net profit remains above 65 trillion KRW. Even at 30% operating margin, annual operating profit is 75 trillion KRW, with net profit above 55 trillion KRW.
Revisiting the secondary market's current pricing for SK Hynix: the market assigns a ~4.4x expected P/E for 2027. Moreover, the company holds 98.7 trillion KRW in net cash, with白纸黑字 commitments to allocate over 50% of cumulative free cash flow from 2025-2027 entirely to dividends and share buybacks, continuously shrinking total shares.
Even deducting net cash and using our heavily slashed远期 bearish steady-state profit, the corresponding EV/EBITDA multiple is only 5-7x. In other words, today's secondary market stock price is effectively pricing in a >75% collapse in SK Hynix's future profits, returning to the惨状 of 2023 where the entire industry lost cash.
Following this line downwards, the players truly hurt in the latter half of the cycle are not SK Hynix, but three types of players.
First, Tier-2/3 memory companies with no HBM tickets, purely profiting from price hikes in low-end DDR4 and consumer memory left over after the three majors siphon wafers, yet whose valuations are pumped to 50-100x P/E. Once traditional capacity returns around 2028, these companies' earnings and valuations will crash together.
Second, single back-end equipment vendors who previously enjoyed monopoly dividends through exclusive binding with one original manufacturer, now seeing share松动 under SK Hynix's three-supplier horse-race strategy, yet still priced at high valuations based on monopoly structures.
Third, mid-to-small PC and mobile hardware manufacturers downstream without long-term locked-price agreements, fully exposed to spot price hikes, squeezed from both sides with no ability to pass on costs.
Based on the bottom-line calculation above, SK Hynix's current valuation has already over-discounted远期 price cut risks. Of course, the author doesn't know to what extent Korean stocks and US ADRs will回调 following peripheral macro sentiment. But calculating clear front/back-end hard metrics and bottom-line profits provides an anchor in mind, far superior to constantly scaring oneself with盘面 small essays.
Conclusion
This article calculated many numbers; finally, the author strings the whole line together.
Analyzing SK Hynix should not focus on the 39.78 trillion KRW net profit halved on the Q3 report books—that's an accounting illusion smashed by Keenex stock pullbacks. Stripping financial asset volatility, the 83.33 trillion KRW operating profit, 21.28 trillion KRW rigid total costs, and 90.1% marginal operating margin represent the true operating leverage of a wafer fab running at full capacity and raising prices.
The feared premise of heavy asset explosions no longer exists. 114.5 trillion KRW cash against 15.8 trillion KRW interest-bearing debt yields 98.7 trillion KRW in net cash. New quarterly depreciation from 50 trillion KRW annual capex accounts for less than 2% of revenue. Earnings not only fund new plants but also allow ~100 trillion KRW for dividends and share注销 this year.
Front/back-end industry barriers are much harder than the market thinks. Globally, only 1.685 million DRAM wafers are produced monthly. HBM4's 3.5-4x wafer consumption coefficient siphons nearly half of traditional memory supply. New plants and EUV equipment won't slowly replenish until after 2027. Back-end improved MR-MUF maintains 80% yield. The $35 base die foundry fee paid to TSMC was recouped tenfold by HBM4's $500-600 price hike per die.
Placing external ghost stories back into the denominator: the 50% HBM share is because industry order fulfillment is only 60-70%, with overflow demand flowing to Samsung; Wuxi has no EUV, and over 75% of core high-margin revenue is entirely in Korea; even under the bearish scenario of 2028+ revenue dropping 40% and operating margins halved, SK Hynix still earns 600-700 trillion KRW in net profit annually. Current stock prices have overpriced cyclical downturn panic.
Therefore, tracking SK Hynix doesn't require staring at quarterly net margin fluctuations or share point drops. The true mainlines determining the endgame are only two: whether the real orders and price cushions locked by five-year agreements with ten key clients loosen, and how fast wafer capacity returns to general-purpose server DRAM after the Yongin new plant starts production in 2027. As long as these two pillars stand, this cash cow holding nearly 100 trillion in net cash is far from time to panic.
The copyright of this article belongs to the original author/organization.
The views expressed herein are solely those of the author and do not reflect the stance of the platform. The content is intended for investment reference purposes only and shall not be considered as investment advice. Please contact us if you have any questions or suggestions regarding the content services provided by the platform.
