Five days before SK Hynix reported Q2, we published our forecast on the quarter.
We wrote:
“July priced an industry demand break. The evidence points instead to a company-specific pricing-realisation dispute inside a physically strong quarter.”
The shares fell roughly 15% on the print.
Sell side research is rarely graded. Ours is, in public, misses included. This is the retrospective: what we got wrong, what held, and how we’re improving our forecast model.
The grade
Here was our forecast:
✗ 65% probability that operating profit would land at or above ₩65T. Printed ₩60.54T. Miss.
✗ 40% probability of ₩68T or above. Miss.
✗ 20% probability of ₩70T or above. Miss.
Revenue printed at ₩79.32T versus our estimate of roughly ₩85T.
KIS's bear estimate of ₩60.4T operating profit, which we argued against all week, landed within ₩0.15T of the reported number.
They were right. We were wrong.
The variable we said would decide the quarter did decide it. We registered blended DRAM price realization near +40%. The bear case required approximately +30%. It printed near +30%.
What we got wrong
Our model assumed that Q1's relationship between industry DRAM prices and SK Hynix's reported selling prices would hold in Q2.
In Q1, industry DRAM prices rose roughly 93% to 98%, while SK Hynix's reported ASP rose in the mid-60% range. We assumed a similar pass through rate in Q2.
That was the error.
Reported pricing is increasingly determined by long term contracts, not just the market. SK Hynix has concluded around 10 long-term agreements (LTAs) of up to 5 years, with deposits and pricing mechanisms designed to reduce volatility. Market prices still matter, but reported ASPs now depend more on contract resets, product mix and customer specific terms.
If you model Korean memory names, this is the transferable lesson from the quarter: the spot tape is no longer a clean input to reported ASP.
What did not break
The quarter missed on price realization - not on memory demand.
In fact the memory demand language was as strong as it gets. Nothing suggests the memory cycle is breaking.
Revenue reached a record, rising 51% quarter on quarter. DRAM and NAND bit shipments landed in line with guidance, and Q3 DRAM bit growth was guided to approximately 10%. For the full year, management expects DRAM demand to grow in the mid-20% range and NAND demand in the high teens.
HBM4 is shipping, HBM4e samples are with a major customer, and 2027 HBM volume and pricing negotiations are underway.
Two of our other registered calls did come true:
✓ HBM4 schedule reaffirmed. Mass production shipments began in Q2, with the full ramp planned for the second half.
✓ Supply remains tight into 2027. Management described an "extreme supply shortage", said customers wanted more memory than SK Hynix could ship, and confirmed that PC and mobile customers were being rationed.
How our model changes
Our new base case is that roughly half of the increase in industry DRAM prices flows through to SK Hynix's reported ASP.
We will place less weight on spot pricing alone and more on contract resets, LTA coverage, product mix, customer deposits and shipment timing.
The implication is a different earnings profile: less upside torque from short term price spikes, but greater visibility and potentially lower cyclicality.
What the market is testing now
The selloff priced lower upside from spot prices. We think the market may be underweighting the offset: more durable earnings in a supply constrained market.
The H2 test is whether contract resets, improving mix and higher volumes keep ASPs and earnings growing. Management expects deferred high value shipments to return, blended ASP to rise, and second half bit growth to exceed the first half.
Our view remains constructive. If those claims land, Q2 will look like a contract timing miss, not a structural ceiling on earnings. We will grade that view against the Q3 print, exactly as we graded this one.
The next test is Wednesday
SanDisk reports Wednesday after the close: the first memory print since this quarter broke the pricing realization assumption we, and the Street's bulls, were carrying. Our call goes on the record before the print and gets graded after it, either way.
Every call we make is dated before its event and stays on the public ledger.
If you want to interrogate any of this yourself, use our research workspace. Search any company we cover and the read opens with the thesis, the evidence behind it, and what would break it. Your first question is on us.
If you want our full SanDisk read before earnings, and Thursday’s public grade after the print, you can become a Pro member here.
Teng