Korea's Chip Exports Rose 179%. Volume Rose 20%.
Korea's semiconductor exports rose 178.8% year over year in July. Export volume rose 20.0%. Both figures describe the same month.
Most coverage stops at the headline number. Korea also publishes a volume index that strips price out, and it says this is not a shipment boom. It is a price event, and HBM is the mechanism.
KEY TAKEAWAYS
1. July semiconductor exports reached $41.02 billion, up 178.8% year over year. The Bank of Korea's export volume index rose 20.0% over the same period. DRAM export prices rose 270.3% in won terms.
2. HBM is why volume cannot follow price. Micron puts the wafer trade ratio above 3x and rising toward 4x for HBM4E, so every HBM bit displaces roughly three conventional DRAM bits of output.
3. The HBM premium is compressing, not expanding. TrendForce calculates that per-wafer revenue on HBM was overtaken by DDR5 64GB RDIMM in 1Q26, and Micron management confirmed non-HBM margins currently exceed HBM margins.
Two numbers, one month
What the July release actually said
Start with the headline. According to the July 2026 ICT trade release from Korea's Ministry of Trade, Industry and Energy and the Ministry of Science and ICT, semiconductor exports came to $41.02 billion, up 178.8% from a year earlier. That is the seventeenth consecutive monthly increase and the seventh straight month of triple-digit growth. Semiconductors alone accounted for roughly 77% of all ICT exports.
Now overlay the Bank of Korea's July trade index, published the same week. The export volume index came in at 153.33, up 20.0% year over year. The export value index rose 64.2%. In other words, the value of what Korea shipped grew more than three times faster than the quantity.
The item-level detail is where it gets stark. DRAM export prices rose 270.3% year over year and flash memory 248.1%. Against that, the volume index for computer, electronic and optical equipment, the category that carries semiconductors, rose 25.2%. Korea's net barter terms of trade improved 24.7%, the largest gain in the history of the series.
One correction worth making, because it is easy to miss from outside Korea. Those DRAM and flash price figures are quoted in won. The won averaged 1,497.43 per dollar in July, roughly 8.9% weaker than a year earlier. Adjusting for that, the increase in contract-currency terms is closer to 240%. Still extraordinary, but 30 percentage points lower than the number that circulates. That is my own arithmetic off the published indices, not a Bank of Korea figure.
Regionally, the growth was broad. Exports to China and Hong Kong reached $19.61 billion, up 246.0%, with the United States up 257.5% and the EU up 238.9%. Hong Kong is a long-standing entrepot, so reading the combined China figure as Chinese AI demand overstates the case. Customs data records where goods clear, not where they are finally consumed.
The wafer tax
Why volume cannot follow price
HBM stacks DRAM dies vertically and connects them with through-silicon vias. Producing a given number of bits in that form consumes considerably more wafer area than producing them as conventional DRAM.
Micron frames this as the trade ratio. On its FY26 third-quarter call, management put the current ratio above 3x and rising toward 4x for HBM4E, and described the consequence plainly: expanding HBM supply leaves less capacity for everything else. Stacking and bonding add process steps, which adds yield-loss surface on top of the raw area cost.
TrendForce estimates that HBM will absorb roughly 18% of the top three suppliers' total DRAM wafer input at the end of 2025, 22% in 2026 and 30% in 2027. Capacity is being converted rather than added. Industry capital spending is rising, but it is going into process upgrades, higher stacking and hybrid bonding rather than raw wafer capacity.
The conclusion follows arithmetically. The more HBM the industry builds, the tighter total bit supply becomes. July's export data, all price and almost no volume, is what that looks like when it reaches trade statistics.
The premium is compressing
The part of this cycle that surprised people
Here is the counterintuitive turn. HBM squeezed conventional DRAM supply so effectively that conventional DRAM prices ran faster than HBM prices.
TrendForce analyzed per-wafer revenue for both, estimated from die size, yield and per-gigabit pricing, and found that HBM was overtaken by DDR5 64GB RDIMM in 1Q26. Profitability followed. The cause is structural rather than demand-driven: the three major suppliers price HBM on annual agreements, which cannot reflect quarterly moves in conventional contract prices when those prices are rising 50% or more per quarter.
Micron's management said the same thing from the inside, confirming on the FY26 third-quarter call that non-HBM margins currently exceed HBM margins. The company also said it is deliberately holding HBM share close to its overall DRAM share rather than maximizing it.
SK hynix's second quarter is the clearest illustration. Revenue of 79.3187 trillion won and operating profit of 60.5426 trillion won, a 76% operating margin, were both quarterly records, and first-half revenue passed 100 trillion won for the first time. Both figures still landed about 5% below the consensus of roughly 84 trillion and 64 trillion won.
The takeaway is not that HBM is a bad business. It is that the reflex of the last two years, that a higher HBM mix is automatically better, stopped holding cleanly this quarter.
Where HBM4 actually is
Announcements versus recognized revenue
HBM3E remains the volume product. All three suppliers are in mass production, and it is where most 2026 HBM shipments still sit.
HBM4 is shipping everywhere but is early. Samsung announced its first mass-production shipment on 12 February, with the base die built on its own 4nm logic process. Korean media reported roughly $1 billion in cumulative HBM4 revenue about 130 days later, citing industry sources rather than a company disclosure, so it is worth holding at that confidence level. Micron disclosed passing $1 billion in HBM4 revenue in its FY26 third quarter and said its 12-high ramp is running twice as fast as HBM3E 12-high, with HBM3E and HBM4 both booked through calendar 2027. SK hynix began mass-production shipments to major customers in the second quarter and guided to expansion in the second half.
HBM4E is at the sampling stage. Samsung announced on 29 May that it had shipped 12-high samples, specifying 14 to 16 Gbps per pin and 3.6 TB/s per stack, built on 1c DRAM with a 4nm logic die. SK hynix completed sample supply in the first half. Micron is developing on its 1-gamma node with volume production targeted for calendar 2027.
"World first" means different things at different companies. Development complete, sample shipment, customer qualification and revenue shipment are four separate milestones, and press coverage routinely collapses them into one.
For anything investment-related there are only two questions worth asking: when was it recognized as revenue, and has it passed customer qualification.
The share table has not moved yet
Counterpoint Research puts Q1 2026 HBM revenue share at 58% for SK hynix and 21% each for Samsung and Micron. SK hynix's number is down sharply from 69% a year earlier, which is the more interesting movement.
Counterpoint is explicit that most of that Q1 revenue came from HBM3E and that HBM4 shipments should materialize in the second half. So the HBM4 news flow of the past six months is not yet in any share table anyone has published. The real scorecard arrives with third-quarter results.
One number in the same dataset deserves more attention than it gets. In overall DRAM revenue, CXMT reached 8% in Q1 2026 against 3% a year earlier. Given that most of this cycle's profit is coming from conventional DRAM pricing rather than HBM, that is not a distant concern.
How to find HBM inside Korean customs data
The part you cannot easily do from outside Korea
A common assumption is that HBM is invisible in trade statistics. That is not quite right. There is no HBM-specific code, but HBM is captured.
Memory falls under HS 8542.32. Within Korea's national tariff schedule that splits further, and two subheadings matter. 8542.32.1010 covers single-chip DRAM. 8542.32.3000 covers multi-chip packages, which is where HBM lands, because an HBM stack is exactly that: multiple DRAM dies assembled into one package.
The second half of the method is geography. Korea Customs publishes exports by customs district, and Korean memory production sits in a small number of places. Combine the two and the picture narrows considerably.
Korean outlets have used this approach for a while. In April 2025, Newsis and Seoul Economic Daily reported, drawing on Korea International Trade Association and customs data, that MCP exports from Icheon and Cheongju rose from $2.051 billion in the first quarter of 2024 to $4.285 billion in the first quarter of 2025. Cheongju alone was up 561.4%, consistent with HBM back-end capacity coming online there. Taiwan took 58% of Icheon's MCP exports and 70% of Cheongju's, which is the packaging route into GPU assembly. Over the same quarter, MCP exports from Cheonan fell 24%.
The limits are real and worth stating. The MCP line carries non-HBM multi-chip packages as well. Figures are booked by customs district rather than by company. Inventory movements and intra-group logistics are included.
This is a tool for reading direction and inflection, not for estimating any company's HBM revenue. Used that way it is a useful monthly signal that arrives well before quarterly results.
The database is public and free at tradedata.go.kr. Enter a code, pick a district, pick a destination country and pull the monthly series yourself.
What I actually watch
| Indicator | Why it matters | Next read |
|---|---|---|
| BOK export volume index | Separates shipments from price. The handoff from a price cycle to a volume cycle is the regime change worth catching. | August data, mid-September |
| Quarterly HBM4 revenue | Every supplier is announcing billion-dollar milestones. Sequential growth, not the announcement, is the ramp rate. | Q3 results, late October |
| Conventional DRAM versus HBM spread | If HBM per-wafer revenue stays behind DDR5 RDIMM, suppliers reallocate capacity and the HBM premium narrative weakens. | TrendForce quarterly |
| Long-term agreement coverage | Micron has disclosed roughly $100 billion in minimum contracted revenue across 16 agreements. SK hynix closed agreements with about ten customers. Longer contracts cut volatility and cap the upside. | Annual and quarterly filings |
| MCP exports by district | Monthly, public and roughly a quarter ahead of earnings. Directional only. | Korea Customs, monthly |
Value chain read-through
| Area | Why it connects |
|---|---|
| Bonding and stacking equipment | More layers means more process steps and more tools per unit of output. Hybrid bonding is absorbing a growing share of memory capital spending. |
| Test | Every generation transition raises test time and screening burden, and the burden peaks precisely during the ramp. |
| Foundry | HBM4 moves the base die onto a logic process. Samsung states its HBM4 base die uses its own 4nm node, which is a structural change in who captures the value. |
| Power and cooling | The data center bottleneck is migrating from memory availability toward power delivery and thermal capacity. |
Risks to this view
- Price-led cycles unwind quickly. An upturn not supported by volume has historically reversed faster than one that is. The current setup is almost entirely price.
- Premium compression. With three suppliers qualified and shipping HBM4, the negotiating balance shifts. HBM's annual pricing structure already lags conventional DRAM in a rising market.
- Customer roadmap slippage. Memory revenue is now tied to a small number of accelerator programs. If those schedules move right, memory results move with them.
- Currency. The won averaged 1,497.43 per dollar in July against 1,527.30 in June, a 2.0% move. Identical dollar exports translate into fewer won, which shows up in Korean reported earnings but not in the export headline.
- Chinese conventional DRAM. CXMT reached 8% of DRAM revenue in Q1 2026 from 3% a year earlier. Since conventional DRAM pricing is where most of this cycle's profit is being generated, that is a direct exposure rather than a peripheral one.
To put it plainly: the export data does not show a boom driven by HBM selling well. It shows HBM constricting supply until everything else repriced. HBM4 has only just started converting into revenue, and it will not appear in share statistics until the second half.
Which makes the next two numbers easy to name. The export volume index, and third-quarter HBM4 revenue.
The next post pulls the Korea Customs database directly and breaks memory exports down by district, product code and destination, month by month.
Everything here is from public sources, IR decks, conference papers and customs data. Sources: MOTIE and MSIT, July 2026 ICT trade release; Bank of Korea, July 2026 export and import price and trade indices (preliminary); Korea Customs Service trade statistics (tradedata.go.kr); Counterpoint Research; TrendForce; Samsung Newsroom; SK hynix Newsroom and Q2 2026 results; Micron FY26 Q3 investor materials.
Disclaimer: This post is for informational and educational purposes only. It does not constitute investment advice or a recommendation to buy or sell any security. All investment decisions are your own responsibility.
Comments
Post a Comment