Same "Sold Out" Headline, Opposite Trade - Micron vs SK Hynix
Saturday, July 25, 2026
Deep Dive
Same “Sold Out” Headline, Opposite Trade — Micron vs SK Hynix
1. The One-Line Thesis
Both memory giants just locked up huge chunks of future output under long-term contracts, but they signed opposite deals. SK Hynix is moving to contracts that keep its pricing upside, after old fixed deals stoked fears its boom quarters are capped; Micron traded its upside away for a guaranteed floor. Both trade at low multiples on record earnings, and our read is the market may be mispricing both, in different directions.
2. Lets Start With Some Context
DRAM (a computer’s fast working memory) Function: Serves to be the high-speed hardware meant for “short-term memory” of laptops, the cloud, and pretty much any type of personal smart device. The data is short-term, so it is gone when the device is turned off. Server grade DDR5 DRAM will run you anywhere from $4.50 to $5.50 per Gigabyte.
HBM (High-Bandwidth Memory) - Stacked DRAM for AI. Function: It’s premium 3 - Dimensional DRAM stacked vertically through microscopic wires known as Through-Silocon Vias (TSVs). Instead of sitting in the typical RAM slots on the motherboard that you’d see next to the CPU, you can find HBM next to AI chips like Nvidia’s GPUs. Because of the complexities HBM is meant to handle, it will run you an upmark of over a 10x what you’d pay for standard DRAM. So expect to pay roughly $40 - $50 per gigabyte.
RTX 4070 from NVIDIA NAND or “Flash” Function: These are far more stable, calm and larger memory storages that you can find within laptop and desktop SSDs, smartphones, and yes, even inside your motherboard or video card. Note that the last two are usually for something like ASUS bloatware, or of course, the BIOS. Since it is slower and more stable, it can be found far cheaper and for far larger quantities. This is what is often used to store memory permanently, while DRAM rids the data whenever the device is powered off. Cheapest of the bunch, expect to pay around $0.06 - $0.12 per Gigabyte.
Plain English: What Just Happened to the Memory Order Book
Two prices matter. The spot price is what memory sells for on the open market today. The contract price is negotiated between producer and customer for months or years ahead.
An LTA (Long-Term Agreement) locks in volume, and sometimes price, for years. The seller gets revenue visibility; the buyer (think Microsoft, Google, Nvidia’s supply chain) gets guaranteed chips in a shortage.
Take-or-pay means the buyer must purchase the committed volume or pay anyway. That’s a hard revenue floor, cancellation doesn’t zero the check.
A price cap is the fine print that decides everything. A seller who caps contract prices gives up the upside if spot prices keep ripping. A seller who removes the cap lets its contract book float up with spot. Same “sold out” press release; radically different economics.
Why now: DRAM contract prices rose 81% quarter-over-quarter in Q1 2026 and are still climbing (+13-18% QoQ expected in Q3), NAND +10-15%, and meaningful new supply is unlikely before late 2027. Whoever keeps pricing flexibility owns the shortage.
3. SK Hynix, The Upside Trade (That Just Got Punished)
Entire 2026 output, DRAM, NAND, HBM, is sold out, with orders booked into 2027, on ~57-60% of the global HBM market and roughly 60-70% of Nvidia’s Vera Rubin HBM4 volume. In HBM4, Hynix isn’t a supplier; it’s the supplier.
The paradox: its old contracts stoked fears it’s missing the boom. Legacy HBM deals were fixed-price for 12-36 months, so while DRAM average selling prices jumped ~30% QoQ and NAND ~50%, the fixed book doesn’t participate. The stock fell ~15% on exactly that concern (”contracts limit earnings upside”), and one broker cut its Q2 2026 operating-profit estimate to ₩60.4tr, 8% below the ~₩65tr consensus, before the report. That’s a pre-report estimate cut, not a reported miss; the report itself is the test.
The pivot is the story, but it’s a transition in progress, not a done deal. As of July 2026, Hynix is moving to uncapped / floating contracts that ride spot prices up. How fast the old fixed book rolls off, how much of the new book floats, and what it does to blended ASP (the average selling price across everything shipped) are all still unproven.
The HBM4 kicker is enormous, if it lands. HBM4 mass production starts Q3 2026 at ~$500/stack, and industry projections have HBM4 pricing roughly doubling toward $4-5/Gb by 2027. Hynix has 3-5 year LTAs with hyperscalers plus prepayments running to the end of the decade.
Our read: the valuation gives little credit for any of this. ~$856B market cap (briefly over $1T earlier this year), ~4.4x forward earnings on consensus FY2026 operating profit of ~₩148 trillion (+213% YoY), after a Q1 that printed a 72% operating margin. To be fair, a low multiple on elevated earnings can already embed both the repricing hope and an expected downturn, we can’t show a clean bridge for what’s “in” the stock. What we can say: if the floating-contract migration and HBM4 pricing land, the current multiple leaves plenty of room to be wrong to the upside.
4. Micron, The Floor Trade (Upside Sold, Durability Bought)
16 strategic customer agreements, 14 of them “foundational,” totaling ~$100B of cumulative committed revenue over CY2026-2030, take-or-pay, with automatic 3-year extensions. Spread it out and that’s ~$20B/year, roughly a quarter of MU’s ~$76-79B annual revenue carrying a hard floor today.
A separate, bigger number: contracted exposure grows toward ~50%+ of revenue as the agreements ramp, the SCAs cover ~20% of DRAM output and ~a third of NAND. But that broader ~50% is on various price terms; it is not all floor-priced take-or-pay. Don’t conflate the two: the hard floor is ~a quarter of revenue, the contracted book is the larger, softer wrapper around it. HBM is 100% sold out for CY2026, price and volume locked including HBM4.
The catch, and it’s the whole trade: MU capped its contract prices at roughly Q2-2026 market levels. Micron sold its cycle for certainty. If spot keeps running, the contracted book does not participate. The upside was traded away, on purpose.
And MU’s HBM4 hand is weak. Nvidia certified all three memory makers for Vera Rubin (June 5, 2026), but Hynix takes ~60-70% of the volume, Samsung ~25-30%, and Micron gets the remainder. MU’s samples show materially lower pin speeds, volume production slips to H1 2027 (~6 months behind Hynix), and one source has it out of initial Rubin builds entirely. A strong #3 in HBM3E is at risk of shrinking in HBM4.
So what’s actually mispriced? Our read: the crash that may not come. MU trades at ~6-11x forward earnings (sources vary) on ~$76-79B FY2026 revenue and ~$34 EPS, a multiple consistent with the market bracing for memory to do what memory always does: collapse. A take-or-pay floor under a quarter of revenue makes the next downturn structurally shallower than prior cycles, that’s the durability argument. Whether the current multiple already pays for it, we can’t prove; the September print (fiscal Q4, ~Sept 23-29) is the first evidence.
5. The Un-Priced Upside, Real, But Unproven
This is contingent optionality, not a sure thing. For the hidden upside to materialize, three things all have to go right, and none has printed yet:
The shortage has to persist into 2027 (no meaningful new capacity is expected online before late 2027/2028, but that’s a forecast, not a fact).
Hynix’s book has to actually reprice, the fixed-to-floating migration has to show up in blended ASP and margin, not just in a press release.
HBM4 pricing has to land near projections, the “~2x by 2027” path is an expectation, and “surge multiples higher” is analyst language, not a signed invoice.
It cuts both ways. Floating contracts ride spot down just as faithfully as up, moving to uncapped pricing raises Hynix’s downside sensitivity in a bust, exactly when Micron’s capped floor would start looking like the smartest deal in memory. The same fine print that makes Hynix the upside trade makes it the higher-beta one.
Honest framing: both stocks carry low multiples on record earnings, which bears read as the classic top-of-cycle setup, and bulls read as un-credited structural change. The contracts say this cycle is different; the multiples say the market isn’t convinced. The signals below tell you which side is winning, in close to real time.
6. Signals to Watch, the Point of This Piece
Bull-confirmation (the hidden upside is materializing)
TrendForce / DRAMeXchange contract-price prints stay positive QoQ into Q4 2026 and 1H 2027. This is the master signal, and the one genuinely monthly data feed here. DRAM and NAND contract gains are already decelerating (+81% → +13-18% QoQ); decelerating-but-positive keeps the thesis alive. Watch every print.
SK Hynix contract-mix disclosure flips (quarterly). On earnings calls, listen for the share of the book moved from fixed to uncapped/floating, and whether blended ASP and operating margin re-accelerate as old fixed contracts roll off. This is the direct proof the July pivot is real.
HBM4 ASP prints versus the ~2x projection. Mass production starts Q3 2026 at ~$500/stack; track reported HBM4 pricing and Hynix’s yield/ramp commentary against the $4-5/Gb 2027 path. Every quarter it holds, the optionality gets more real.
Prepayment commentary strengthens. Customers wiring cash years ahead is the market’s own scarcity signal; there’s no standardized disclosed “ratio,” so listen for it in earnings commentary and balance-sheet movements.
Hyperscaler capex guides keep climbing (quarterly). MSFT, GOOGL, META, AMZN capex guidance is the demand side of the entire structure. Rising guides = the buyers of all this contracted memory still want more.
MU-specific: guidance HOLDS the first time spot softens. This is the floor’s proof-of-life, the first soft price print that doesn’t dent Micron’s outlook is the moment the durability thesis earns credit. Also watch for new SCAs beyond the current 16, and any sign of HBM4 share recovery (pin speeds, ramp timing, a Rubin design win).
Bear-warning (the optionality is dying)
QoQ contract-price increases turn negative. The cleanest tell in the sector. The moment TrendForce prints a down quarter, floating contracts start moving the wrong way and the whole repricing thesis inverts.
Supply arrives earlier than expected. Announcements of late-2027/2028 capacity are consistent with the thesis, the bear signal is capacity coming online ahead of that timeline, or a material acceleration in supply, including credible China ramps (CXMT in DRAM, YMTC in NAND) pulling the shortage’s end date in.
Hyperscaler capex cuts or an AI air-pocket. One flat-to-down capex guide from a major buyer is a yellow flag; two is the cycle top.
Take-or-pay stress. Reports of customers renegotiating or walking back committed volumes mean the “hard floor” is softer than the contracts read, this hits the Micron durability thesis directly.
MU’s HBM4 share slips further. If Micron confirms exclusion from initial Rubin builds or slips past H1 2027, its premium-product story shrinks to the capped commodity book.
Generic HBM
7. Counter-Case, What Breaks It
Micron’s floor is a REVENUE floor, not an EARNINGS floor, the strongest single objection to the durability thesis. ~$20B/year of take-or-pay covers ~a quarter of revenue, and it guarantees the top line only: margins stay fully exposed to costs, product mix, execution, and HBM4 share loss. In a real downturn MU could hit every contracted dollar and still see earnings compress hard. The floor is softer than “durability” reads.
The “peak earnings” bears might simply be right. Every memory cycle has ended with an explanation of why this one was different. Low multiples on record earnings is the market’s standard top-of-cycle posture, and it has usually been vindicated.
Take-or-pay has never met an AI-scale shock. Contracts of this kind can get renegotiated under severe demand stress, committed volume is a legal claim, not a law of physics, and this structure is untested against a genuine AI demand air-pocket.
Hynix’s repricing could disappoint on timing. Old fixed contracts run 12-36 months; the drag that spooked the market doesn’t vanish in a quarter. The stock already fell 15% on the gap between the “sold out” headline and the fixed book’s economics, it can happen again before the floating book dominates.
Micron’s floor has a ceiling problem. Even if the floor holds, a business that capped its contracted book at Q2-2026 prices in a rising market has structurally muted earnings torque, durability may earn a better multiple, but it won’t print upside surprises.
8. Bottom Line
One headline, two opposite postures: Micron sold its cycle for a $100B cumulative take-or-pay floor ($20B/yr, ~a quarter of revenue) with capped prices; SK Hynix, sold out for 2026, booked into 2027, is mid-transition to uncapped contracts that keep the upside Micron gave away.
The mispricing is different in kind, our read, not a proven bridge. MU at ~6-11x looks priced for a crash its floor makes shallower, the un-priced asset is durability (revenue durability, note, not earnings durability). Hynix at ~4.4x looks lightly credited for the repricing and HBM4 (~60-70% of Vera Rubin, prices projected to ~double by 2027), the un-priced asset is optionality. Either multiple could also simply be the market correctly pricing a downturn.
The optionality is real but unproven, and it’s symmetric: floating contracts raise downside sensitivity exactly as they raise the upside. Nothing here is a sure thing until it prints.
The scoreboard is public, monthly for prices, quarterly for the rest: TrendForce contract prints (monthly), Hynix’s fixed-to-floating mix and margins (quarterly calls), HBM4 ASPs and ramp commentary, hyperscaler capex guides, prepayment commentary. Both cases will show up in that data before they resolve in either stock.
Nearest checkpoints: Hynix’s HBM4 mass-production ramp and Q2 report against the ~₩65tr consensus in Q3 2026, hyperscaler capex guides through the summer, and Micron’s fiscal Q4 report in late September, the first real test of whether the floor holds when the market squints at it.