The lead that SK hynix built in high-bandwidth memory was supposed to be durable. It was the reward for moving first, for absorbing the yield risk on Nvidia’s earliest HBM orders, and for spending on capacity when rivals hesitated.
A Bernstein analysis published Oct. 26 concludes that the lead has been erased, at least on the revenue line. Working from Korean export data for July through September, the brokerage’s regression model estimates that Samsung Electronics booked about $8.7 billion of HBM revenue in the third quarter against roughly $7.6 billion at SK hynix. If the estimate holds, it is the first quarter in several years in which Samsung has sold more HBM than its smaller domestic rival.
The export numbers behind the model were striking on their own. HBM export value rose 45% month over month, 17% from the second quarter and 131% from a year earlier, according to the Bernstein note.
Neither company has reported yet. SK hynix publishes results on Wednesday, together with a shareholder return plan investors have been waiting for since its cash flow turned. Samsung discloses its memory figures later in the season. Until then the ranking rests on a model built from customs data, which records the value of goods leaving Korea rather than revenue recognised by the seller, and which can be moved by where a stack is packaged before it reaches the customer that designed it in.
The shift has a mechanism, and it is not technical. SK hynix won the first two generations of the AI memory cycle by qualifying early with Nvidia and holding the volume Nvidia needed. Samsung spent more than a year failing the qualification tests on HBM3E, a delay that cost it the opening phase of the boom and much of the credibility it had built as the largest memory maker in the world. It qualified eventually. Then it did what a company with the largest wafer capacity in the industry can do: it priced to take volume.
That is a different business from the one SK hynix has been running. Samsung’s HBM revenue includes stacks it assembles internally and logic dies it can build on its own foundry lines; SK hynix buys base dies from TSMC for the HBM4 generation and pays for the packaging step. Samsung can therefore accept a lower price per stack and still book a margin, while SK hynix has to defend a premium to cover the outside work.
Revenue leadership is not profitability, and the gap between the two is where this quarter gets interesting. SK hynix has run HBM margins far above its commodity DRAM business for two years, which is why its shares have outperformed Samsung’s so heavily. A quarter in which Samsung outsells it on revenue while SK hynix keeps the better margin would describe a market splitting into two roles: one supplier selling performance at a premium, the other selling capacity at a price.
There is a customer angle as well. Nvidia’s allocation decisions set the tone for the whole market, and the company has an interest in keeping at least two qualified HBM suppliers so that it is not dependent on one Korean factory for the memory inside its accelerators. Samsung’s return to volume fits that interest. It also makes Samsung a candidate for the HBM4 generation that enters production for Nvidia’s next platform, where the base die becomes a logic chip and the memory maker’s foundry capability starts to matter.
SK hynix has answered on the technology side. It has said its HBM4 samples run at the speeds the platform requires and that its packaging capacity is booked well into next year. The company’s argument is that share of the leading edge is worth more than share of total volume, and that its yields on the most difficult stacks remain the reason customers pay.
Both claims can be true in the same quarter. Total HBM revenue is growing fast enough, up 131% from a year earlier, that a supplier can lose the top spot while still setting records of its own. That is what the export data suggests is happening: a rising market in which the second supplier grows faster than the first.
The memory industry has a long record of revenue crowns that changed hands just before prices turned. What usually decides the next phase is not who shipped the most in the quarter when supply was tight, but who is holding inventory and who has the yield when the shortage ends. Both companies have spent the past two years adding capacity on the assumption that demand from AI data centres keeps compounding. The Bernstein estimate says nothing about whether that assumption is right. Wednesday’s numbers will say more about who is better paid for it.
For now the ranking is a model output, published by an analyst who covers both companies and built from trade statistics that both companies can see. The audited figures arrive within days. If they confirm the estimate, the memory business will have a new leader in the product that matters most to it, and a new argument about whether leadership in HBM is a technology position or a pricing one.
