SK HYNIX · PEAK-OUT EARLY WARNING

SK Hynix Peak-Out
Early-Warning Radar

"Catch the conditions for a top faster than anyone, and first." A field monitoring dashboard cross-checked against historical cycles, leading indicators, and credible research.

Data as of: June 26, 2026 · Sources: TrendForce, Morgan Stanley, SemiAnalysis, Counterpoint, DigiTimes, SK Hynix filings, etc.

SK Hynix is living through its strongest quarter ever. HBM is effectively sold out, and conventional DRAM contract prices jump double digits every quarter. Yet the most dangerous moment for a memory investor has always arrived when the numbers look their best, because the stock turns before the fundamentals do.

This radar is a tool for catching that turn before the crowd. It puts the common script of four cycles across forty-nine years, the eight leading indicators worth watching daily and weekly, and a map that splits "when, why, and how far" the stock falls, all on a single screen. The starting premise is simple: if you wait for the earnings release, you are already late.

Current cycle position (as of 2026.06)
Not at the top yet
Late-cycle entry, but no peak signal lit
EarlyRisingCautionPeak

In all four historical cycles (1995, 2010, 2014, 2018), the five leading indicators turned negative just before the peak. Today, five of the eight are still green, a combination not seen ahead of any prior peak. The key, though, is this: "strongest fundamentals now" does not equal "the stock is safe."

🎯 If you watch only three things today

1
HBM4 / HBM4E contract-price direction (③) — the heart of SK Hynix's profit. The first crack is when next-gen prices go flat or fall versus the prior generation.
2
Hyperscaler capex guidance (④) — the source of demand. Two or more cutting guidance is a strong signal.
3
Big Tech ROI and depreciation news (⑤) — the trigger of the causal chain. Watch for write-downs and "pacing our investment" remarks.

⚡ The 30-second takeaway

  • The stock turns about two quarters (six months) before the fundamentals. In 2018, Micron's stock topped two quarters before its earnings peak. Waiting for the print is too late.
  • This is the late cycle of "strongest now, doubted ahead." Even Apple has caved to price hikes (a peak of good news), while OpenAI's IPO delay marks the start of ROI doubt.
  • The real triggers are ④ HBM contract prices and ⑤ hyperscaler capex guidance. If those two roll over, they outweigh any Apple-driven good news.
  • The fastest alarm = spot-price deceleration + double-ordering caught + a Big Tech depreciation shock.

1The peak-out causal chain

The mechanism running from "ROI doubt" to "peak-out." Here is the current status of each stage.

① ROI doubt raised lit

Depreciation > profit. The five Big Tech names plan ~$2T in AI assets by 2030, implying ~$400B in annual depreciation (above their combined 2025 profit). Amazon FCF is projected negative for 2026.

② Sentiment cooling early signs

OpenAI's IPO pushed to 2027 (ROI and market-volatility reasons). Despite a Q4'25 earnings beat, the absence of estimate upgrades signals a perceived profitability ceiling. AI and chip stocks fell together.

③ Order cuts not yet

None so far. HBM is sold out for 2026, with new 16-Hi HBM4 orders running into Q4'26. Nvidia demand is, if anything, accelerating.

④ Memory-price decline not yet

The opposite. HBM4 is priced 50%+ above HBM3E (mid-$500 per stack). Q1'26 DRAM contract prices rose +90–95% QoQ (a record). Even Apple gave up resisting the pass-through.

⑤ Earnings/price peak not arrived

SK Hynix's conventional DRAM operating margin is in the high-70s% (an all-time high). Most institutions forecast the earnings peak for Q4 2027.

The point: ①–② (causes) are already lit; ③–⑤ (effects) are not. The question to ask is "when do we catch the moment ② spreads into ③." That moment is the sell trigger.

2What history tells us

Memory has repeated the same script for 49 years. These are the common features of a peak-out.

CycleUp-legPrice collapseStock drawdown
1995 (1st supercycle)~2 yrs-51% → -65%-60 to -80%
2018 (pre-HBM supercycle)~30 monthsmargin 59% → 27%Micron -56%
2021–22 (COVID)~14 monthsSK '23 net margin -28%Micron -50%
2024– (AI/now)~30 months+none yetnone yet

Three recurring lessons

Capex is announced 12–18 months into the up-leg, and that capacity creates a glut 2–3 years later that ends the cycle. (In 1995, 50 fabs announced → prices collapsed 60%+.)

The stock turns about two quarters before the fundamental peak. The market prices the turn in before it shows up in the numbers.

"This time is different" is dangerous, yet AI is price-insensitive demand, so the argument that the structure differs from past (consumer-led) cycles also holds. That is why we split it into scenarios.

Where we are: ~30 months from the mid-2023 trough equals the longest cycle on record (2018). By duration alone, we are already near a peak. But the five leading indicators are still accelerating, an unusual phase where "duration vs. indicators" conflict.

3Eight leading-indicator lights

Watch just these, daily and weekly. The moment three or more turn red is the real peak alarm.

① Memory spot price
Tier 1 · the fastest lead
green
Now: DDR5 spot ~3x its September level, +171% YoY. Contract prices also accelerating at +90–95% QoQ in Q1'26.
⚠ Alarm: spot gains slow → flatten → turn down. Spot moves 1–2 months ahead of contract. The first light to flicker.
Where: TrendForce / DRAMeXchange spot quotes, the DXI (DRAM eXchange Index) trend.
② Supplier inventory (in weeks)
Tier 1 · core of supply-demand balance
green
Now: SK Hynix and Micron ~2 weeks, Samsung ~6 weeks. Extremely tight versus normal (8 weeks) and prior pre-peak levels (15+ weeks).
⚠ Alarm: supplier inventory reverses up, 4 → 8 → 10+ weeks. If buyer (Big Tech) inventory also starts piling up, it signals a double-order peak.
Where: DIO (days of inventory) commentary in SK, Samsung, and Micron quarterly results and IR.
③ HBM contract price / pricing power
Tier 1 · SK Hynix's core profit engine
green
Now: HBM4 priced 50%+ above HBM3E (mid-$500 per stack). Price-hike talks with Nvidia succeeded; Samsung also refused to discount.
⚠ Alarm: HBM4E / next-gen contract prices come in flat or down versus the prior generation. Or Nvidia gains the leverage to push prices down (successful supply diversification).
Where: DigiTimes / TrendForce HBM price reports, coverage of Nvidia–SK negotiations.
④ Hyperscaler capex guidance
Tier 1 · the source of demand
caution
Now: $600–725B for 2026 (+36–74% YoY). Still rising, but with signs of slowing increments and no next-quarter upgrade.
⚠ Alarm: capex guidance is cut, or "slower growth" is stated in quarterly results. Two or more of MSFT, Google, Meta, and Amazon makes it a strong signal.
Where: capex-guidance remarks on the MSFT, Alphabet, Meta, and Amazon quarterly calls.
⑤ Big Tech ROI / depreciation shock
Tier 1 · the chain's trigger
caution
Now: OpenAI's IPO delayed to 2027 (ROI worries). Warnings of under-stated depreciation and "Fugazi" accounting. FCF deteriorating. Not yet reflected in orders.
⚠ Alarm: Big Tech announces accelerated depreciation or an AI-asset write-down, or officially says it is "pacing AI investment." Confirm a slowdown in AI revenue growth.
Where: the depreciation line in Big Tech 10-Q/10-K, AI-revenue disclosures, major press and short-seller reports.
⑥ New fabs / expansion (supply)
Tier 2 · a 12–18-month lagging threat
green
Now: Micron's first new fab (Idaho) starts mid-2027, the NY fab broke ground in Jan 2026. Samsung targets 150k wafers/month of 1c (end-2026). No meaningful new supply before 2027–28.
⚠ Alarm: all three makers announce aggressive capex competition at once (especially conventional DRAM). The true cause that ended every past cycle. 2027 is the watershed.
Where: the three makers' capex guidance, national fab-investment announcements, TrendForce capacity outlooks.
⑦ Double / triple ordering (phantom demand)
Tier 2 · a late-bubble signal
caution
Now: buyer inventory at ~10 weeks, higher than suppliers'. Signs of panic ordering and forced bundling (1:1 motherboard tie-ins in Taiwan, RAM purchase limits in Japan).
⚠ Alarm: with real and phantom demand impossible to tell apart, buyers cancel or defer orders all at once. This feeds straight into ③ order cuts.
Where: channel-check reports (Taiwan/China distribution), signs of module/finished-goods price inversion.
⑧ AI memory-efficiency breakthroughs
Tier 3 · structural demand destruction
green
Now: quantization and KV-cache compression can cut inference DRAM by 30–50%. But model scaling is offsetting it (absolute demand keeps rising).
⚠ Alarm: efficiency gains overtake scaling and absolute memory demand starts to fall. A long-term, structural threat (low probability near-term).
Where: inference-infrastructure papers and tech blogs, Big Tech trends in memory used per token.

4Four peak-out scenarios

When, why, and how far it falls. Each scenario's trigger and expected drawdown. Probabilities are a synthesized judgment as of now.

A · Soft landingprob. 35%
Timing: H1–mid 2027 · the most likely base case

New fabs (Micron Idaho and others) come online mid-2027 and supply gradually catches up. Hyperscalers digest the inventory they bought, and demand normalizes. Prices ease gently instead of crashing.

Trigger: ① spot prices flatten + ⑥ 2027 new supply starts + ④ slowing capex increments. The three overlap gradually.
Expected: margins retreat from the 60s% to the 40s% (still healthy). The stock corrects -25 to -40% off the high, then stabilizes. Earnings peak around Q4 2027.
B · ROI disappointment → freezeprob. 30%
Timing: H2 2026–H1 2027 · the chain fires fastest

Big Tech ROI doubt crosses a threshold. After the OpenAI delay come further blows (a depreciation shock, AI-revenue disappointment, asset write-downs). Hyperscalers abruptly cut capex guidance → orders shrink → HBM pricing power reverses.

Trigger: ⑤ ROI shock lit → ④ capex cut (two or more) → ③ orders canceled or deferred. The ②→③ transition in the chain is the crux.
Expected: the stock falls sharply about two quarters before the fundamental peak. -40 to -55% off the high. The "trap zone" where the stock drops first while earnings still look good.
C · Oversupply hard landingprob. 20%
Timing: H2 2027–2028 · the history-repeats worst case

The three makers race to expand aggressively in 2026–27 (especially conventional DRAM) → new fabs come online together in 2027–28 just as AI demand slows. The 1995 and 2018 script plays out: glut → price crash.

Trigger: ⑥ simultaneous expansion announcements + ⑦ double orders canceled en masse + ⑧ efficiency gains pile on. The phantom-demand bubble bursts.
Expected: historical pattern = prices -50 to -65%, stock -56 to -80%. SK Hynix could swing to a loss (precedent: -28% net margin in 2023).
D · Supercycle extensionprob. 15%
Timing: peak 2028+ · if "this time is different" is right

AI inference demand offsets the slowdown in training. Servers become "memory machines" and the HBM TAM hits $100B by 2028. Price-insensitive AI demand structurally breaks the cycle, and new supply still cannot keep up.

Holds if: ④ HBM prices stay strong + ② inference demand stays firm + ⑥ supply stays constrained (through 2028). Every light stays green.
Expected: the earnings peak is deferred past 2028. But even this scenario ends eventually, and the drawdown then could be larger (accumulated leverage).

5SK Hynix-specific variables

Strengths and weaknesses unique to SK Hynix, apart from the whole market.

50–62%
No.1 HBM market share
~70%
projected HBM4 share for Nvidia Rubin (UBS)
high-70s%
conventional DRAM operating margin (record)
~90%
Micron HBM's reliance on Nvidia (concentration risk)

Strengths

Dual-generation HBM3E and HBM4 supply at once, packaging partnership with TSMC, primary-supplier status to Nvidia. Cheongju M15X fab online May 2026. Record margins.

Weaknesses / risks

Samsung's HBM4 catch-up — Nvidia brought Samsung to the table a week after the SK deal. If Samsung supplies early in Q2 2026, SK's H1 HBM4 near-monopoly breaks → weaker pricing power.
China's CXMT — racing to build HBM3E capability; export-control wildcard.
Nvidia concentration — single-customer reliance is a double-edged sword.

SK-specific sell signals: "Samsung breaks 30%+ of Nvidia's HBM4" or "Nvidia forces price cuts via HBM supply diversification" are SK Hynix-specific negatives, independent of any broad market peak-out.

6The hands-on monitoring checklist

Check at these cadences and you will know before almost anyone.

DRAM/HBM spot-price trendDXI index, TrendForce spot. Watch for slowing gains or a turn downweekly
Supplier inventory weeks (DIO)SK, Samsung, Micron quarterly results. Any 4 → 8 week reboundquarterly
Hyperscaler capex guidancecapex direction and commentary in MSFT, Google, Meta, Amazon quarterly resultsquarterly
Big Tech depreciation / ROI newsasset write-downs, AI-revenue slowdown, "pacing investment" remarks, further OpenAI movesongoing
HBM4/HBM4E contract-price talksprice direction vs. prior gen, Nvidia's leverageongoing
Samsung's HBM4 share at Nvidiabreak of SK's near-monopoly = idiosyncratic negativeongoing
Three makers' new capex/fab newsespecially conventional DRAM expansion = a historical cycle-end signalongoing
Major IB peak-call changeswhether Morgan Stanley, TrendForce, SemiAnalysis revise their peak timingmonthly
Sell-decision rule (example): when three or more of the Tier 1 indicators (①–⑤) turn red at once → high odds of a cycle turn. Especially if ④ capex cut and ③ order shrinkage confirm together, it is the most reliable signal in history. Do not wait for the earnings release (the stock moves two quarters ahead).

Key tracking keywords

DRAM spot DXI supplier inventory weeks HBM4 contract price hyperscaler capex guidance AI capex ROI depreciation write-down double ordering Samsung HBM4 Nvidia share book-to-bill

iKey sources

  • History cycles and the five leading indicators — useLuminix DRAM Cycle Analysis (2026.03)
  • Four-cycle drawdown data — UncoverAlpha "Every Memory Cycle Ends the Same" (2026.03)
  • Inventory weeks and contract prices — TrendForce, The Economy (2025.10), IntuitionLabs (2025.12)
  • Margins and peak-timing outlook — The Diligence Stack "Memory's $200B Inflection" (2026.02), Morgan Stanley
  • HBM4 price negotiations — DigiTimes (2025.11), SeDaily / TrendForce
  • Hyperscaler capex — CreditSights, A.L. Capital Advisory (2026.05)
  • OpenAI IPO delay — NYT / Bloomberg (2026.06.25)
  • Apple price hikes — CNBC, MacRumors (2026.06.25)
  • Supercycle-extension thesis — SemiAnalysis "Memory Mania," Blocks & Files (2028 outlook)