"Our clients' demand for HBM is reportedly more than double the volume we can supply," SK Hynix told analysts on its third-quarter call. The company reportedly printed an operating profit near 83 trillion won, roughly seven times the year-earlier figure, at an operating margin close to 80%, per a third-party secondary source.
Last week the story was that memory had become the binding constraint on AI. This week it became something more specific: a signed, multi-year contract architecture, with the bill already arriving at the consumer's door.
The tell is not the sold-out book. It is what SK Hynix said about its own pricing. CFO Kim Woo-hyun told the call that next year's HBM volumes and prices are fully negotiated with major customers, that HBM4's higher value was "faithfully reflected" — and that the profitability gap between HBM and conventional DRAM will therefore narrow meaningfully.
A seller of a product reportedly sold out into 2029 is guiding its premium down. In a normal shortage, that sentence does not exist.
Read that twice. A seller of a product that is reportedly sold out into 2029 is guiding its premium down. In a normal shortage, that sentence does not exist.
What replaces it is allocation. SK Hynix has reportedly expanded its long-term agreement roster from about ten clients to fourteen, and those contracts reportedly cover more than 50% of its total 2027–28 production capacity. Terms run as long as five years, with extensions under discussion.
Samsung is walking the same road faster. It has reportedly completed LTAs with ten major customers and is reportedly negotiating with ten more. Vice President Kim Jae-jun reportedly said 2027 HBM supply is sold out at target prices, with additional volume requests still arriving. The company reportedly plans to allocate 60–70% of total capacity to long-term contracts by 2030.
The customer list has escaped the AI bubble. Long-term demand is spreading from hyperscalers and server OEMs into mobile, automotive, and networking. Everyone who needs memory has decided the spot market is no longer a place to buy it.
This is the structural break. Memory used to be negotiated quarterly and priced off spot, which is why it was a commodity. It is now negotiated in multi-year tranches with advance deposits that secure fulfillment and help finance capacity — an order book closer to a foundry than to a DRAM vendor.
SK Hynix reportedly says that its LTA pricing operates within agreed ranges that reflect market conditions. That clause is the whole trade. It cushions the downside in a bust and caps the upside in a boom.
Now trace the wire from that contract to the phone in your hand. Producing HBM consumes roughly three times the wafer capacity per unit of memory as conventional DRAM — bigger dies, complex stacking, process waste. About 70% of advanced memory capacity has reportedly been absorbed by AI data centers, leaving consumer electronics to compete for the remainder.
When lines shift to HBM, general DRAM loses capacity. SK Hynix's DRAM average selling price rose in the high-20% range sequentially in the third quarter, and DDR4 pricing ended the quarter above four times its year-earlier level. NAND is up roughly eightfold.
That passes through to the bill of materials. Memory's share of a phone's component cost has reportedly climbed from 10–15% to over 30%, and reportedly approaches 50% on large-storage flagships. A 12GB mobile DRAM procurement that reportedly cost around 200 yuan now reportedly runs near 600; a 1TB flash module has reportedly gone from just over 200 yuan to 600.
Samsung's own mobile division is the cleanest proof. The DX division reportedly posted an 800 billion won operating loss in the third quarter, its second straight, with roughly 600 billion won of it reportedly coming from mobile — reportedly driven by memory costs.
Vice President Daniel Araujo of the MX business reportedly said volumes fell mainly in mid- and low-priced handsets because of memory price increases. The company's answer is to shift the product mix toward flagships. That is what cost-push looks like when it reaches the shelf: fewer cheap phones, more expensive ones.
The physical constraint is not going away on anyone's 2027 calendar. Samsung reportedly noted that the lead time from fab construction to wafer output means next year's higher capital spending cannot contribute meaningfully to 2028 bit supply. HBM4E and HBM5 have larger die sizes and substantially more metal process layers, which further limits how fast volume can grow.
HBM4E mass production is reportedly targeted for the first half of next year, with ramp in the second — reportedly the first concrete next-generation timeline any global memory maker has given. That is a capacity statement, not just a technology one.
Both companies reportedly now say tight supply persists into 2028, and SK Hynix reportedly extends the shortage view to 2030. Its CFO reportedly forecast next-year DRAM bit growth in the mid-20% range and NAND in the high-20% range, against demand he reportedly expects to outrun it.
So why cut the premium? Because the contract has already converted scarcity into something the seller values more than price: allocation security. Predetermined volumes keep utilization high, cut inventory risk, and finance expansion with customer deposits. HBM is designed and certified per customer, which makes supplier switching nearly impossible once a socket is won.
The cost of that certainty is that when demand explodes, you cannot reprice. The upside is capped by the very agreements that guarantee the volume. Record profit and shrinking premium are the same fact wearing two faces.
Market pricing already reflects the tension. Samsung and SK Hynix reportedly together posted nearly 190 trillion won in third-quarter semiconductor operating profit, and by October 30 their shares reportedly sat 39.9% and 52.5% below their 52-week highs. Investors are pricing the cycle turn before the shortage has ended.
That is the contradiction worth sitting with. The industry has stopped selling a commodity and started selling rationed capacity — which means the shortage is no longer the variable that sets your laptop's price. The contract is.
Consumer relief, when it comes, will not arrive because supply catches demand. It will arrive when the contracts roll.