The most profitable memory company in the world just decided the least glamorous product in its catalog is worth more than the one everybody is fighting over. SK hynix is dialing back the pace at which it converts HBM3E lines to HBM4 — the memory riding Nvidia's next-generation Rubin accelerators — and pointing that capacity at ordinary DDR5 server DRAM, where margins are approaching levels memory executives have historically only fantasized about.
Inside the company, the logic is arithmetic. Outside it, the move landed closer to a confession. Within two weeks Korean chip stocks had given up a chunk of a year's gains, and the question on Seoul trading desks was no longer whether AI memory demand is strong, but whether SK hynix knows something the market does not.
What SK hynix Actually Did
The reporting originated with Chosun Biz on June 23 and was extended by TrendForce the same day. SK hynix is slowing its HBM4 ramp and pulling back on the planned conversion of selected HBM3E lines earmarked for HBM4 upgrade. The reason is margin: analysts cited in those reports project conventional DRAM operating margins approaching a theoretical peak of 90 percent within the year, making commodity DDR5 more immediately lucrative than the next generation of high-bandwidth memory.
SK hynix has room to make that trade because it is already winning. HBM accounts for more than 40 percent of its revenue, and it holds roughly 56 percent of global high-bandwidth memory revenue. The commodity side is running hot — DRAM average selling price rose in the mid-60-percent range in the first quarter, and a three-year DDR5 supply deal with Microsoft gives visibility into demand that does not depend on the accelerator cycle.
TrendForce quantified the shift: HBM4 mass production has been pushed back, with meaningful volume now expected in 3Q26, and the full-year HBM4 shipment outlook cut from 4.5 billion Gb to 4 billion Gb. Samsung, running the opposite play, saw its full-year estimate revised up from 3.5 billion Gb to 4 billion Gb after crossing 1 billion dollars in cumulative HBM4 revenue just four months after February mass production.
This is a slowdown, not a stop. Digitimes reported on July 13, citing The Bell and industry sources, that SK hynix began mass-production shipments of 12-layer HBM4 to Nvidia at the end of June and is increasing output ahead of a broader ramp in September. Both things are true at once: the product is shipping to the industry's most important customer, and the capacity behind it is growing more slowly than planned.
The Market Read
Investors did not parse the nuance. On July 2, a chip rout that began on Wall Street spread to Seoul, with Samsung falling around 9 percent and SK hynix considerably further. After Samsung's July 7 preliminary results showed a 19-fold jump in quarterly operating profit — and still disappointed — its shares fell 6.9 percent, dragging the Philadelphia semiconductor index down 4.7 percent overnight. By July 8, Samsung slid as much as 7.6 percent and SK hynix 5.2 percent intraday. Micron, Samsung, SK hynix and the Roundhill Memory ETF all fell more than 20 percent from recent highs.
The bear case is not about AI. It is about the second derivative of memory pricing. Park Yuak of Kiwoom Securities cut his Samsung target price about 9 percent to 390,000 won, warning that rising component prices for memory and substrates are driving up PC and smartphone selling prices. Concerns over weakening demand are making customers more conservative in additional purchases, he wrote, limiting the likelihood that memory price increases will exceed market expectations.
Analysis: Two Stories That Cannot Both Be Right
The elegant reading is that this is a demand signal — that SK hynix sees HBM4 orders softening and is redeploying before the market catches up. The awkward problem is that SK hynix's own chief executive is saying the opposite, publicly, with maximum legal consequence attached.
On July 10, the day the company completed the largest-ever U.S. listing by a foreign issuer, raising 26.5 billion dollars on Nasdaq, CEO Kwak Noh-jung told Reuters: We forecast that next year will be the worst year in the industry's history from the supply perspective. He expects demand to outstrip supply beyond 2030. That is not the language of an executive hedging against a demand cliff.
A more mundane explanation fits better. HBM4 is capital-intensive, yield-sensitive and priced under long negotiated supply agreements. Commodity DDR5, in a genuine shortage, prices off spot and contract markets moving violently in suppliers' favor. If you already own 56 percent of the HBM market and commodity margins approach 90 percent, deferring conversion capex by a quarter is not a bearish call on AI. It is a bet that the scarce resource is wafer starts, not customers.
What makes the selloff rational anyway is timing risk. SK hynix has ceded near-term HBM4 share to a Samsung that qualified early on a 4nm FinFET base die. If Nvidia's Rubin ramp pulls harder than expected, SK hynix has less slack to answer it — and an incumbent optimizing for margin while a challenger optimizes for share is a pattern with an unhappy history in this industry.
What to Watch
Three markers: whether HBM4 volume materializes in 3Q26 and the September Nvidia ramp holds schedule; whether fourth-quarter DDR5 contract pricing decelerates, eroding the arbitrage that justified this decision; and Samsung's share trajectory into HBM4E, where SK hynix shipped 12-layer samples on June 18 at 48GB per stack and 16Gbps per pin.
Treat the specifics with caution. The core reporting rests on Korean-language trade coverage citing industry sources, not company disclosure, and SK hynix has not confirmed the line-conversion details. What is confirmed is the tell that matters: a company that says 2027 will be the worst supply year in history is, right now, allocating its scarcest capacity to the cheaper chip.
“We forecast that next year will be the worst year in the industry's history from the supply perspective.”— Kwak Noh-jung, Chief Executive Officer, SK hynix