SK Hynix Doubles Sales to Its Biggest Customer

The customer does not appear by name anywhere in the filing. It is listed as a single external buyer, and its purchases doubled in nine months.

SK Hynix said in its third-quarter report released Tuesday that sales to one customer reached 34.98 trillion won in the first three quarters, up 101.5% from a year earlier. That buyer accounted for 14.79% of total revenue, which implies company-wide sales of roughly 236 trillion won. The industry reads the customer as Nvidia Corp., the largest purchaser of high-bandwidth memory and the reason the market for it exists. SK Hynix does not identify customers in these disclosures.

The pace is what stands out. The company booked more than 17 trillion won of sales to that one buyer in the third quarter alone, close to everything it sold that customer in the first six months of the year.

High-bandwidth memory is the part of the AI supply chain where supply has stayed short. Each stack is a tower of DRAM dies wired to a logic base and placed beside a processor, and the manufacturing yield is lower than for conventional memory because a single bad die ruins the stack. SK Hynix began full mass production of HBM4, the fourth generation of the standard, in the third quarter.

The price effect has spread into ordinary memory. Average selling prices for DRAM rose about 20% from the previous quarter, and shipments of server-grade DDR5 rose more than 40%. That combination is unusual: DDR5 is the commodity product, and it tightened because wafer capacity keeps getting converted to HBM, which consumes several times the die area per bit of data stored.

Capital spending reached a record 31 trillion won. Management told analysts that DRAM and HBM capacity is sold out through 2029.

That last sentence deserves a second reading. Sold-out capacity is a statement about contracts, not about demand. Customers have signed for volumes years ahead, which gives the supplier visibility no memory maker has had in previous cycles and hands it the pricing power that comes with a queue.

Geography tells the same story as the customer line. Revenue from the United States was 147.7 trillion won, or 62.5% of the total. A company selling most of its output to one American chip designer books most of its revenue in America, and the remainder is spread across module makers and device manufacturers in Asia.

Memory has been the most cyclical business in semiconductors for four decades, and the shape of the cycle has not changed: shortage, price spike, capacity addition, glut, losses, consolidation. What has changed is the customer base. When DRAM went into personal computers, demand was spread across hundreds of buyers and prices fell as soon as anyone added a fab. When it goes into accelerators, demand concentrates in a handful of designers who plan their roadmaps years out and will pay for guaranteed supply.

Concentration is the price of that stability. One customer at 14.79% of revenue, and a customer whose own order book depends on cloud budgets, means SK Hynix’s results now move with a single procurement decision made in Santa Clara.

The competitive picture is more contested than the numbers suggest. Samsung Electronics is shipping HBM4 and has been working to qualify its stacks with the same customer. Micron Technology has taken share in HBM3E and is building capacity in the United States. Neither has displaced SK Hynix at the top of the stack, and both have spent two years trying.

SK Hynix’s advantage came from timing. It qualified HBM3 and then HBM3E first, which meant it took the volumes when the accelerator ramp arrived. The lead is now measured in quarters rather than years, and each generation resets the qualification race.

The record capital budget is the decision that will be judged in 2029. Memory makers have historically destroyed their own pricing by building through the top of a cycle, and 31 trillion won is a large number even for a company of this size. The defense is that HBM capacity is committed by contract before the equipment is ordered, which was not true in previous expansions.

There is a second defense, which is that the technology itself limits how fast capacity can be added. HBM4 stacks require advanced packaging and known-good-die testing, and the base die is now made on a foundry process. Adding capacity means competing for packaging lines and foundry wafers, not just for floor space.

What investors will watch next is the fourth quarter, when HBM4 pricing for 2027 is negotiated. A price increase confirms the shortage is structural. Flat pricing suggests the sold-out language covers a period the customers have already hedged.

The risk sits in that same concentration. If the accelerator market slows, the customer that doubled its purchases in nine months can halve them in two quarters, and a fab cannot be turned off. The memory industry has run this sequence many times. This time the buyer is one company, and it has its own reasons to want a second supplier.

For now, though, the arithmetic is simple. One customer, 35 trillion won, and a queue that runs three years past the current quarter.

Related Posts

  • November 20, 2026
  • 13 views
Sandisk Completes Eight-for-One Split as NAND Outlook Darkens

Eight for one. Sandisk’s stock split took effect on Nov. 19, and the shares began trading on the adjusted basis on Nov. 20, a change that divides the price by…

  • November 20, 2026
  • 13 views
A Lawmaker’s Privacy Suit Puts Model Design on Trial

For three years British courts have sidestepped the question of whether the company that builds an AI model answers for what its users produce with it. A written defence filed…