TL;DR
Bernstein uses South Korean export data as a leading indicator for HBM revenue, estimating Samsung's Q3 HBM revenue could reach $12 billion, about 30% higher than its original forecast.
Samsung's exports from the relevant regions in July increased by 122% compared to April, with unit values doubling, indicating that growth is likely driven mainly by the high-priced HBM4 volume.
SK Hynix's related exports decreased by about 27% compared to April, and a baseline regression model points to a 20% quarter-on-quarter decline in Q3 HBM revenue, but monthly data and seasonal differences introduce significant uncertainty.
Bernstein attributes SK Hynix's weakness to delays in Rubin-related HBM4 shipments, but this remains speculation by analysts and is not a confirmed causal relationship by Nvidia or SK Hynix.
HBM prices have not surged in tandem with traditional DRAM, and Samsung's unit value increase reflects more of a shift in product mix towards HBM4 rather than significant price hikes for products of the same specifications.
Bernstein remains bullish on Samsung, SK Hynix, and Micron, believing that the next round of profit upgrades is more likely to come from 2027 HBM contract prices rather than solely relying on this year's share changes.
South Korea's July memory export data provides the first set of leading signals for Samsung and SK Hynix's Q3 HBM business.
Bernstein has been tracking South Korea's multi-chip memory exports to Taiwan and Malaysia. The report suggests that these exports have a strong correlation with Samsung and SK Hynix's quarterly HBM revenue: Taiwan is a key location for CoWoS advanced packaging, while Malaysia hosts Intel's EMIB packaging facilities.
In July, South Korea's multi-chip memory exports to Taiwan and Malaysia decreased by 32% from the historical high in June. However, the report attributes this change mainly to seasonal factors within the quarter. Compared to the first month of the previous quarter in April, July's export value still grew by 13% and increased by 64% year-on-year, indicating that overall HBM demand has not significantly weakened.
While the total remains strong, the export trends of the two South Korean memory manufacturers have diverged: Samsung's related exports continue to rise, while SK Hynix's data is significantly below Bernstein's expectations.
Bernstein uses Chungcheongnam-do's exports as a proxy indicator for Samsung's HBM shipments, as Samsung's related backend production and packaging facilities are mainly located in this region.
In July, Chungcheongnam-do's multi-chip memory exports to Taiwan and Malaysia reached $2.2 billion. Although this is a 35% decrease from the seasonal peak in June, it represents a 122% increase from April. Samsung's related export value has also exceeded SK Hynix's for two consecutive months.
Based on the regression relationship between historical export data and HBM revenue, Bernstein estimates that Samsung's Q3 HBM revenue could reach approximately $12 billion, representing a quarter-on-quarter increase of about 80%, which is about 30% higher than its previous estimate of approximately $9.3 billion.
If only recent data since Q1 2025 is used for regression, the forecast value would rise further to $12.6 billion, corresponding to a quarter-on-quarter increase of about 90%.
However, this figure is still a model estimate and not company guidance. Samsung's exports are typically concentrated in the last two months of the quarter, and since 2025, the average share of the first month of the quarter accounts for less than 20% of total quarterly exports. Therefore, whether Q3 revenue will ultimately reach the model's forecast depends on actual shipments in August and September.
Bernstein believes that the July data at least aligns with the direction previously indicated by Samsung: the company expects Q3 HBM4 sales to increase more than threefold quarter-on-quarter and to account for over 60% of HBM sales in the second half of 2026.
In addition to export scale, the export unit value of Samsung's related products has also significantly increased.
In July, Chungcheongnam-do's multi-chip memory export unit value increased by another 21% quarter-on-quarter, more than doubling from April levels, approaching four times that of SK Hynix's related regions. Due to the higher capacity, stacking layers, and technical complexity of HBM4, Bernstein believes this change likely reflects a rapid increase in the share of Samsung's HBM4 shipments.
This does not mean that Samsung's HBM prices of the same specifications have doubled within months. Export unit values are also influenced by product structure, capacity specifications, and packaging combinations, and can only serve as directional indicators for average selling prices. A more reasonable explanation is that Samsung is replacing some HBM3 and HBM3E with higher-priced HBM4, thereby raising the overall export value.
This is also key to Samsung potentially regaining market share. Previously, SK Hynix held a leading position with HBM3E and a closer supply relationship with Nvidia; as the HBM4 cycle begins, the focus of competition is shifting to who can complete validation, improve yield, and achieve scale delivery first.
In contrast to Samsung, SK Hynix's related exports significantly weakened in July.
Bernstein uses Chungcheongbuk-do and Icheon's exports as proxy indicators for SK Hynix's HBM shipments. In July, the export value from these two regions decreased by 28% compared to June and by about 27% compared to April.
Based on this, the baseline regression model estimates that SK Hynix's Q3 HBM revenue could be only $5.6 billion, a quarter-on-quarter decline of about 20%, which is 55% lower than Bernstein's original forecast. If this result materializes, Samsung's Q3 HBM revenue will significantly surpass SK Hynix's.
However, the uncertainty of this prediction is much higher than for Samsung. The report also notes that if SK Hynix's historical shipping patterns, which are most concentrated in the latter part of the quarter, are applied, its Q3 HBM revenue could still reach $12 billion, close to Bernstein's original forecast.
In other words, the same set of July data under different seasonal assumptions could correspond to a huge range from $5.6 billion to $12 billion. At this stage, a more prudent judgment is not that "SK Hynix has lost its lead," but rather that its Q3 start is weaker than expected, and whether it can catch up in August and September will determine the final share.
Bernstein speculates that the weakness may stem from delays in Rubin-related HBM4 shipments. SK Hynix previously stated that HBM4 began mass production in Q2 and will ramp up fully in the second half of the year. However, directly attributing the July export decline to Nvidia's Rubin schedule remains an inference made by analysts based on industry rhythms and has not been clearly confirmed by the relevant companies.
The report also points out that the price trends of HBM and traditional memory are diverging.
Since Q3 2025, traditional memory prices have increased about fivefold, but the export unit values of SK Hynix's related products have remained generally stable. Samsung's unit value has approximately doubled, mainly due to the increased share of HBM4, rather than significant price hikes for existing products like HBM3E.
This indicates that the contract pricing, supply relationships, and product iteration cycles of HBM are causing it to partially detach from the fluctuations of traditional DRAM spot prices. Short-term price increases in traditional DRAM do not necessarily translate proportionately to HBM, whose profit changes depend more on the share of new generation products and annual customer contracts.
Bernstein believes that suppliers and customers have already begun negotiating HBM contracts for 2027, and expected price increases will drive Samsung and SK Hynix's profit expectations for next year to continue to rise. Compared to the single-quarter share changes in Q3, this could be a more important pricing variable for the memory sector in the next phase.
In July, South Korea's multi-chip memory exports to Malaysia increased by 20% month-on-month to about $1.3 billion, continuing the rapid growth of the past few quarters. Samsung still accounts for the majority, but SK Hynix's exports have also significantly increased.
Bernstein speculates that this HBM may be flowing to Intel's advanced packaging facilities in Malaysia for server chips equipped with HBM. However, the report also acknowledges that it is currently difficult to explain why related exports have grown so early before product mass production.
Therefore, exports to Malaysia are more suitable as subsequent supply chain clues and cannot be directly equated to Intel customer orders or specific product volume.
Overall, the July data reinforces the logic of Samsung's accelerated catch-up in the HBM4 cycle, but it is still insufficient to confirm that the long-term share pattern has reversed. What needs to be observed next is whether Samsung can maintain export growth in August and September, and whether SK Hynix can catch up with the shipment gap at the beginning of Q3 as HBM4 ramps up.
The July export data reinforces the logic of Samsung's accelerated catch-up in the HBM4 cycle: related export scale has significantly increased, unit values have rapidly risen, and they corroborate the HBM4 sales guidance provided by the company.
However, these data are still insufficient to confirm that Samsung has regained its leading position in HBM on an annual basis.
First, Bernstein uses a regression model of regional exports and company revenues, not actual order data from the companies. Second, if HBM packaging is transferred outside South Korea, or if some products are further packaged within South Korea, customs data may not fully capture this. Finally, monthly exports can also be affected by customer acceptance, shipping arrangements, and quarterly seasonality.
Next, three variables need to be observed: whether Samsung's exports in August and September can continue to grow, whether SK Hynix's exports and unit values can rebound as HBM4 ramps up, and whether the two companies can truly raise HBM contract prices for 2027.
If Samsung's exports continue to grow and SK Hynix fails to catch up, HBM4 may drive more substantial changes in market share; if SK Hynix concentrates deliveries in the last two months of the quarter, the divergence in July may be more likely just a timing mismatch in shipments.
Therefore, the current data truly changes not that the HBM competition has yielded results, but that Samsung's speed of re-entering the share contest may be faster than the market previously expected.
This content is provided for general informational purposes only and doesn't constitute financial, investment, legal, or tax advice. Any events, rewards, online promotions, or related information mentioned herein should not be considered a recommendation, solicitation, or invitation to purchase, sell, trade, or otherwise deal in any crypto assets. Crypto assets are highly volatile and may result in loss. The availability of WEEX services, products, and related events may vary by region. You are responsible for ensuring that your participation is in accordance with applicable local laws and regulations.





























