The memory wall is now a booking calendar
The latest signal from South Korea is not simply that AI chips are hard to buy. It is that the memory sitting beside those accelerators is being reserved years before it ships. Samsung Electronics and SK Hynix are expanding long-term supply agreements, or LTAs, as cloud companies, AI developers and device makers try to secure high-bandwidth memory and advanced DRAM before the next wave of data-center builds arrives .
The headline number is 2029. SK Hynix says its DRAM and HBM output is effectively sold out through 2029 when signed LTAs are combined with customer demand now under negotiation . Samsung’s disclosed position is narrower but still powerful: it has sold out its 2027 HBM supply, has signed long-term agreements with 10 major customers and is negotiating with 10 more . Together, those statements show a market where supply is no longer allocated quarter by quarter. It is being claimed several product cycles in advance.
This matters because AI infrastructure is not only a GPU story. A server rack needs HBM attached to accelerators, large pools of DRAM, storage, substrates, networking and cooling. If memory is booked years out, the bottleneck moves from the accelerator alone to the full server bill of materials. That is the shift now visible in the Korean memory market.
What Samsung and SK Hynix actually said
The two companies are not making identical claims. SK Hynix has signed LTAs with 14 customers, including major cloud providers, and more than half of its 2027 and 2028 production capacity is already covered by long-term agreements . When volumes still being negotiated are added, the company describes DRAM and HBM as practically sold out through 2029 .
Samsung has signed long-term contracts with 10 customers and is in talks with 10 more, with demand expanding beyond AI companies into mobile, automotive and network customers . The company has said it is nearing its goal of allocating 60% to 70% of total production capacity to long-term contracts by 2030 . A separate current analysis emphasizes the distinction: SK Hynix’s 2029 figure includes negotiated demand, while Samsung’s confirmed sellout is focused on 2027 HBM, with pricing concluded .
That distinction is important. “Sold out” does not always mean every wafer is already covered by an unconditional purchase order. In SK Hynix’s case, the through-2029 framing includes signed LTAs plus specific customer discussions; confirmed LTAs exceed 50% of 2027–2028 capacity, but not every future unit is fully contracted yet . For Samsung, the public claim is more limited in time and product scope, but it still points to the same direction: HBM capacity is being committed early, and management wants a majority of output governed by long-term arrangements.
Why AI buyers are moving so early
AI companies used to compete mostly for accelerators. Now they are competing for the memory ecosystem that makes those accelerators useful. HBM is stacked next to AI processors and provides the bandwidth required to feed large models; advanced DRAM supports the rest of the server platform. If either piece is missing, expensive compute capacity can sit underused.
That is why LTAs are becoming strategic. These contracts typically set volumes, pricing mechanisms and sometimes deposits over a defined period, giving customers supply assurance while helping manufacturers plan investment and production . For chipmakers, LTAs reduce the classic memory-cycle risk of building too much capacity at the wrong moment. For buyers, they trade flexibility for survival: guaranteed supply may be more valuable than waiting for a lower price that never arrives.
The pressure is no longer limited to the largest U.S. hyperscalers. Samsung says demand for long-term contracts is spreading from AI companies to mobile, automotive and network firms . That broadening matters because memory capacity is fungible only up to a point. A wafer allocated to premium HBM or server DRAM cannot instantly solve shortages in every other category, and manufacturing transitions take time.
Pricing power, but not unlimited power
The booking wave strengthens memory producers’ hand. When customers are lining up years ahead, suppliers can negotiate not just price but duration, deposits and product mix. Korea JoongAng Daily reported that Samsung has LTAs with 10 major customers and that SK Hynix has LTAs with 14 customers, while both companies are benefiting from record semiconductor earnings .
Yet LTAs also complicate pricing. Aju Press noted that securing contracts does not automatically mean every price increase flows through at the market average, because some server DRAM contracts use pricing mechanisms that can limit upside . In other words, the industry has moved from spot-market scarcity to structured scarcity. That may smooth revenue for Samsung and SK Hynix, but it can also cap some of the upside if contracts are formula-based.
The broader supply view remains tight. Aju Press reported that market researchers expect shortages to persist into 2027, while Samsung has indicated tight supply could continue into 2028 because new fabs take time to build and equip . SK Hynix has gone further, saying customer demand exceeds its available HBM supply by more than double and expecting shortages through 2030 .
The risk shifts to cloud operators and chip designers
For cloud operators, the message is uncomfortable: securing GPUs is necessary but no longer sufficient. The memory stack must be booked early, and that means capital planning has to stretch deeper into the decade. If memory suppliers commit capacity to one group of customers, latecomers may face higher prices, fewer product choices or delayed deployments.
For chip designers, the lesson is equally sharp. AI accelerator roadmaps depend on memory availability. A faster compute die is less useful if HBM supply, packaging or server DRAM becomes the constraint. The most successful designs may be those that align early with memory partners rather than those that simply advertise peak compute.
This is why the “sold out through 2029” story is bigger than a supplier boast. It changes procurement behavior. Buyers will reserve earlier. Suppliers will demand clearer commitments. Smaller customers may be forced to accept whatever capacity remains. The memory wall has become a reservation system, and the best seats are being taken years before the show.
What to watch next
The first thing to watch is how much of SK Hynix’s through-2029 demand converts from negotiation into binding LTA volume. The second is whether Samsung reaches its target of placing 60% to 70% of capacity under long-term contracts by 2030 . The third is whether new capacity arriving after 2027 is absorbed by AI demand or triggers a price reset, a risk Aju Press highlighted through market-research commentary .
For now, the direction is clear. AI demand has pulled future memory supply into the present. Samsung and SK Hynix are not just selling chips; they are selling calendar years. In the AI infrastructure race, the scarce commodity is no longer only the accelerator. It is the memory allocation that lets the accelerator work.
Comments
Be the first to comment.