SK Hynix 1Q26: The Mix Cycle, Not the Volume Cycle

Every prior memory up-cycle in history had bits AND prices rising together. This one doesn't.

SK Hynix just reported its first-ever KRW 50T+ quarter — revenue of KRW 52.58T (~$38B), +60% QoQ / +198% YoY, operating margin of 72%, and a balance sheet that flipped from net debt to KRW 35T (~$25B) net cash in two quarters. The surface-level takeaway is "record revenue, AI demand strong." The analytically important content is underneath the surface.

Note on units: SK Hynix reports in Korean won (KRW). We show USD approximations in parentheses using ~KRW 1,400 = $1, the mid-April 2026 rate. Rounded to the nearest billion for readability.

The one-sentence read: This is a mix cycle, not a volume cycle — and if your 2026 memory model assumes a 2017-style print, it is wrong by approximately 40%.

This note walks through six things mainstream coverage will miss: why flat bits with rising prices is a structurally different signal than any prior memory cycle, how the NAND inflection finally arrived, what 72% operating margin means for the industry, why the balance sheet transformation matters more than the income statement, what the SOCAMM2 disclosure actually signals, and how this plays for procurement teams, designers, and investors over the next two quarters.

1. The flat-bit paradox

SK Hynix grew DRAM revenue roughly in line with the consolidated +60% QoQ print. It did this on flat DRAM bit growth. DRAM ASP rose mid-60% QoQ. Essentially all of the revenue expansion came from mix and price, none from volume.

1Q26 is the first quarter where bit growth stalled while ASP accelerated — the signature of a mix cycle.

Source: Silicon Analysts estimate using SK Hynix quarterly earnings disclosures, 2Q25–1Q26. Earlier quarters are approximate reconstructions for trend illustration.

Why flat bits with rising ASP is structurally different

A normal memory up-cycle has bits and ASP climbing together. What 1Q26 shows is different. Two non-exclusive readings:

Reading 1 — The mix is crowding the fab floor. HBM (sold per-stack, not per-bit), high-density server DIMMs (128GB and 256GB modules), and LPDDR5X/LPDDR6 for AI-capable devices are absorbing disproportionate wafer capacity per bit shipped. HBM alone consumes enormous die area per usable bit because of the TSV and base die overhead. Holding total wafer starts constant, a shift toward richer-mix products produces flat bit volume and higher blended ASP. That is exactly what we see.

Reading 2 — SK Hynix is holding supply to defend pricing. Capex and wafer-starts allocation are discretionary. In a quarter when contract prices were being negotiated aggressively, operating below maximum bit output keeps the pricing anchor high. The balance sheet makes the choice cheap to maintain.

Why both readings matter: Both are bullish for 2026 DRAM pricing. Mix shift is a structural drift — it does not reverse on a quarter. Supply discipline is a choice, but the capital position removes the compulsion to chase volume.

The practical implication: if you do not have LTAs locked in, this is your reality for at least two more quarters. Management guided 2Q26 DRAM bit growth at high-single-digit % QoQ — a step up from flat, but still constrained relative to demand signals. Model the BOM-level impact in our AI Chip Cost Bridge or the Chip Price Calculator.

2. NAND's forgotten inflection

For three years, NAND was the memory industry's stepchild. Excess capacity from 2022-2024 capex, weak enterprise and client demand, and the QLC transition kept prices pinned near cash cost. SK Hynix — through Solidigm — was in a worse spot than pure-play NAND competitors because enterprise SSD was the softest corner of the softest market.

That ended in 1Q26. NAND bits fell ~10% QoQ. NAND ASP rose mid-70% QoQ.

NAND's ASP inflection is sharper than DRAM's. This is the AI-inference eSSD cycle turning.

Source: Silicon Analysts estimate using SK Hynix quarterly earnings disclosures, 2Q25–1Q26. Earlier quarters are approximate reconstructions for trend illustration.

The supply-side response has already started shipping

The demand driver is enterprise SSD — specifically QLC eSSD for AI inference clusters and training checkpoint storage. Hyperscalers are absorbing QLC product faster than the industry can produce it. Hard-drive alternatives don't meet the latency requirement.

SK Hynix's supply-side response is already in market: 321-layer QLC NAND ("PQC21"), CTF-based architecture, now shipping to AI PCs (as cSSD) and data-center eSSDs. Kioxia, Micron, and Samsung are on similar nodes. The capex cycle to add meaningful QLC eSSD capacity runs four to six quarters — until mid-2027, the supply side cannot meaningfully catch up.

Don't confuse bit growth sources next quarter. 2Q26 NAND bit growth guidance is mid-teen % QoQ — but that includes the first full Solidigm consolidation quarter. Pure-play operational bit growth is lower. When comparing to Micron or Samsung, back the Solidigm effect out.

See NAND Shockwave: AI Demand Triggers SSD Price Explosion for the full demand-side breakdown.

3. 72% operating margin in context

Consolidated 1Q26 operating margin was 72%. That is a record. It is also roughly 20 percentage points above TSMC's 1Q26 operating margin.

OPM expanded 30 points in five quarters. 1Q26 marks memory out-earning the most advanced logic foundry for the first time.

Source: SK Hynix quarterly earnings disclosures, 1Q25–1Q26.

Historically anomalous. And temporary.

In every prior memory up-cycle — 2017-2018, 2021 — memory OPM peaks sat in the 45-55% range. Leading-edge logic foundry margins were always higher because of structurally better pricing power on bespoke wafers. 1Q26 inverts that relationship.

Three things are working in memory's favor simultaneously:

Base case for 2H27-2028: memory OPM normalizes toward the 45-55% range. Still excellent. Not 72%. Model a glide path, not a cliff — the three factors above fade at different speeds, which smooths the descent.

4. The SOCAMM2 signal buried on slide 11

The earnings deck noted: "Mass production of 1cnm 192GB SOCAMM2 optimized for NVIDIA's Vera Rubin platform." This line will not get airtime in financial coverage. It should.

What is SOCAMM2, and why now

SOCAMM (Small Outline Compression Attached Memory Module) is a server memory module form factor NVIDIA has been pushing to replace RDIMM/MRDIMM in specific AI server configurations. It delivers higher density per channel, better signal integrity at DDR5-8000+ speeds, and a physical footprint that fits the thermal envelope of tightly-spaced GPU server trays. SOCAMM2 is the second generation.

Two signals in the disclosure:

Platform-specific qualification. SK Hynix's production is specifically for NVIDIA's Vera Rubin platform — the post-Blackwell generation. "Mass production" started in April 2026 — that's a hard, dated signal that Rubin is transitioning from sample to pre-ramp.

Memory BOM lock-in. Custom module form factors are hard to dual-source. Once a hyperscaler has qualified SK Hynix SOCAMM2 on a Vera Rubin reference platform, switching to Samsung or Micron requires re-qualification of the entire memory subsystem — not just DRAM component swap. This is effectively a multi-quarter preferred-supplier position.

For cluster buyers: 2H26 AI server capex plans need a recalibration line item for SOCAMM2-compliant memory. The cost step-up over conventional RDIMM is material. Factor it before placing Vera Rubin orders.

Model the server-memory BOM impact in the AI Chip Cost Bridge. For HBM-specific pricing and share dynamics, see HBM Market Analysis or the HBM Pricing dataset.

5. The balance sheet transformation

The income statement is where the headlines are. The balance sheet is where the industry structure is changing.

Cash more than 3.8x'd while debt fell 17% — a net swing of ~KRW 44T (~$31B) in six months.

Source: SK Hynix quarterly balance sheet disclosures, 3Q25–1Q26. Cash includes cash and short-term financial investments. USD equivalents: 1Q26 Cash = KRW 54.3T (~$39B); Debt = KRW 19.3T (~$14B).

From KRW 9T (~$6B) net debt to KRW 35T (~$25B) net cash — in two quarters

The net cash to equity ratio moved from roughly +11% to −21% in six months. Management has explicitly stated a target of >KRW 100T (~>$70B) cash balance, implying the accumulation is not done. Critically, they said on the call they will "review dividends, share buybacks & cancellations — prepare implementation plans within the year." This is not if. It is when.

Why this matters for the industry: the memory industry's long-running problem has been that all three suppliers carry capex intensity that swamps free cash flow across the full cycle. A "good" memory balance sheet historically meant modest net debt at the peak and manageable leverage at the trough.

SK Hynix is no longer capital-constrained. With KRW 54T (~$39B) cash on the balance sheet and ~KRW 40T (~$29B) annualized operating cash flow at current run-rate margins, the company can:

• Self-fund Yongin + M15X simultaneously. Yongin alone is a ~KRW 120T (~$86B) multi-decade buildout. M15X is another ~KRW 25T (~$18B). Both are now fully underwritten without incremental debt issuance.
• Return meaningful capital to shareholders via dividends and buybacks, without starving capex.
• Out-invest competitors on HBM capacity in 2027-2028 without the balance sheet deterioration that previously constrained cycle-peak capex.

Samsung Memory is buried inside Samsung Electronics' capital allocation. Micron is ~$10B smaller in annual FCF and does not have this capacity. For the first time in memory industry history, one supplier has a materially stronger balance sheet than the other two — and has stated it will act on it.

6. Working capital as a supply-tightness indicator

Buried in the cash flow statement: accounts receivable rose from KRW 18.2T to KRW 33.8T (~$13B → $24B) in one quarter — a KRW 15.6T (~$11B) jump. Inventory rose only KRW 1.7T (~$1B) over the same period. The difference tells the story.

Receivables tripled in three quarters; inventory barely moved. Product is flying out the door faster than customers pay — the opposite of a glut signal.

Source: SK Hynix balance sheet disclosures, 1Q25–1Q26. Earlier quarters approximate from statement of financial position disclosures. USD equivalents at 1Q26: Receivables KRW 33.8T (~$24B); Inventory KRW 16.0T (~$11B).

This is not a collections problem

SK Hynix's largest receivables counterparties are the major hyperscalers and NVIDIA. Customer quality has not changed. What changed is the invoice-to-cash cycle under supply pressure: product is being pulled out of the fab and onto aircraft/ships faster than customers cut checks.

Why this is the opposite of a glut signal: In a glut, product sits in inventory because demand softens. Inventory builds while AR stays flat. In 1Q26, inventory barely moved (+KRW 1.7T / ~$1B) while AR exploded (+KRW 15.6T / ~$11B). Demand is absorbing supply as fast as it leaves the fab. Headlines will miss this signal entirely — it's a cash-flow-statement artifact.

The −KRW 17.6T (~−$13B) line in the cash flow statement ("changes in working capital, etc.") is the accounting footprint of this dynamic. Operating cash flow was still +KRW 26.4T (~+$19B) after absorbing the drag — a testament to the operating profit underneath.

7. Why net income is a red herring

Do not use KRW 40.35T (~$29B) for modeling. It contains:

• KRW 9.94T (~$7B) in valuation gains on investment assets (Kioxia-related positions and other holdings). Mark-to-market. Will reverse in either direction next quarter.
• KRW 1.57T (~$1B) in FX gains from KRW/USD movement. Also non-operational.

Strip both out. Clean operating profit baseline is KRW 37.61T (~$27B). That is the figure for DCF, multiples-based valuation, and capital-allocation models.

Headline coverage will round KRW 40.35T (~$29B) into forward estimates. Forward estimates built on that figure will overshoot once the valuation-gain tailwind reverses. Anchor on operating profit.

8. Memory efficiency as a demand driver — the Jevons pitch

Slide 9 of the earnings deck framed AI memory demand through a Jevons paradox lens: more efficient AI memory utilization (better compression, larger context windows in HBM, sparsity exploitation) expands rather than contracts demand, because the efficiency unlocks use cases that consume more aggregate memory than the original workload saved.

This is the explicit counter-argument to the "AI capex digestion" bear case. The bear case holds that as inference costs drop and model efficiency improves, aggregate memory demand plateaus. The Jevons reading holds that improved per-token memory efficiency expands the accessible use-case surface — agentic systems, multi-modal inference, persistent per-user memory — each scaling memory consumption super-linearly with usage.

We are not endorsing or rejecting the pitch. It is self-serving for a memory supplier to make it. What is worth noting: it is the explicit framing from the single best-positioned company in memory, and it directly contradicts the "memory is about to deflate" thesis. If your model has a 2H26 memory digestion scenario, SK Hynix's view is now clearly on record that it expects the opposite.

What this means for you

Different readers have different exposure. Here is where the 1Q26 print should update your model:

What to watch next quarter

2Q26 guide: DRAM bit growth high-single-% QoQ, NAND bit growth mid-teen % QoQ (including the Solidigm consolidation effect). Management expects the mix cycle to continue, with HBM share of DRAM rising further.

Three earnings in ~10 weeks will resolve whether this is SK-Hynix-specific or a full memory supercycle:

• Micron FQ3'26 (late June). Can Micron replicate margin expansion? Micron starts with lower HBM mix share, so its DRAM ASP delta is the cleanest test of whether this is industry-wide. Anything materially below +mid-60% QoQ signals Micron is under-indexed to HBM.
• TSMC Q2'26 (mid-July). CoWoS-L is the gating factor for HBM4 ramp. Any update on the three N3 sites announced in 1Q26 and HBM base-die allocation feeds directly into SK Hynix's 2H26 HBM4 cost structure.
• Samsung Memory (late July). HBM4 qualification status with NVIDIA is the load-bearing disclosure. Every quarter Samsung is out of qualification is a quarter SK Hynix holds share.

Internal readthroughs:

• HBM4 customer schedule. Management referenced "agreed schedules" — language that implies signed LTAs. Watch for named volume commitments on the Q2 call.
• Yongin groundbreaking cadence. The pace of Yongin buildout is the single biggest supply-side variable for 2028-2030 memory balance.
• Solidigm disclosure granularity. As Solidigm's relative contribution rises, disclosure should separate operational NAND performance from Solidigm consolidation effects more cleanly.

The one-paragraph takeaway

SK Hynix crossed KRW 50T (~$36B) quarterly revenue for the first time — KRW 52.58T (~$38B) exactly — posted 72% operating margin that exceeds TSMC's, and flipped from net debt to KRW 35T (~$25B) net cash in two quarters. Underneath: a mix cycle — flat DRAM bits, mid-60% DRAM ASP expansion, and a sharp NAND ASP inflection driven by AI-inference eSSD — not a volume cycle. KRW 40.35T (~$29B) net income is not a clean operating baseline once KRW 11.5T (~$8B) of valuation + FX gains are stripped out. The KRW 15.6T (~$11B) one-quarter receivables build is a sell-out signal headlines will miss. SOCAMM2 mass production for NVIDIA Vera Rubin started this month — the earliest concrete signal post-Blackwell memory BOM is locking. And the balance sheet transformation — self-fundable Yongin + M15X plus buybacks telegraphed for within the year, with KRW >100T (~>$70B) cash targeted — is the structural development that will shape memory-industry competitive dynamics through 2028. This is not a "record revenue, AI demand strong" story. It is a mix-and-balance-sheet story, and those are the two variables to model from here.

References & Sources