Samsung warns AI memory shortage will worsen into 2027 as HBM supply sells out

HBM supply being “sold out” for 2027 does not mean Samsung has stopped producing it. It means the memory Samsung expects to make that year has already been allocated through long-term customer contracts. Most of the available output is therefore spoken for before 2027 begins.

Samsung said ten major customers have committed to its 2027 HBM supply. Negotiations with ten additional customers are underway. These agreements give buyers more certainty, while they reduce the amount of uncommitted capacity available to new or smaller purchasers.

The shortage reflects demand from AI infrastructure, not merely a Samsung-specific backlog. Samsung estimates that AI-driven demand will significantly exceed industry supply in 2027. For organizations planning AI servers, securing HBM may require commitments well before delivery. South Korean business outlets report that tightness could extend into 2027–2028, although company guidance can change.

High-bandwidth memory, or HBM, is a type of memory built from stacked layers. The article identifies it as one of the most constrained components in AI infrastructure. Unlike ordinary memory, HBM is placed next to the AI accelerator, the chip that performs much of the model computation.

That nearby placement is important because AI accelerators constantly move large amounts of data between memory and the computing chip. Shorter connections and a wide data interface allow HBM to feed information rapidly. This helps the accelerator spend more time processing rather than waiting for data. The article does not provide technical bandwidth figures.

HBM has become central to the AI buildout, and its supply is already heavily committed. Samsung says its 2027 output is sold out, while SK Hynix says its DRAM and HBM are effectively sold out through 2029. This makes HBM availability a major constraint for future AI infrastructure purchases.

Samsung does not quantify the expected shortfall in units, percentage points, or dollars. Its stated estimate is that AI-driven memory demand in 2027 will significantly exceed total industry supply. The important point is the direction and scale of the imbalance: demand is expected to be substantially larger than what manufacturers can provide.

This is broader than Samsung’s own customer queue. Samsung frames the gap as an industry-wide imbalance. SK Hynix has separately said its DRAM and HBM are effectively sold out through 2029. Together, the statements indicate that much of the industry’s near-term capacity is already committed.

The available evidence therefore supports a clear conclusion, but not a precise forecast. Buyers should expect tight memory pricing and availability into 2027. The article identifies contract coverage, new fabrication capacity, and quarterly pricing trends as the main indicators to watch. Guidance can still change.

When demand remains above supply, scarce memory usually becomes harder to obtain and more expensive. The article specifically says memory pricing and availability are likely to remain tight into 2027. This pressure affects HBM directly and also reaches conventional DRAM used in servers and workstations.

AI servers depend on accelerators paired with HBM. If manufacturers cannot secure enough HBM, they may be unable to complete or expand server systems at the desired pace. Buyers could therefore encounter longer lead times, limited configurations, or higher purchase costs. The article does not quantify any expected price increase or delivery delay.

Long-term contracts may protect customers that secured capacity early, while leaving less committed supply for others. Samsung says its 2027 HBM output is already sold out, and SK Hynix reports effective sellouts through 2029. New fabrication capacity could ease pressure, but the article gives no timetable for how much relief it would provide.

The article identifies buyers only by category, not by company name. Samsung says ten major customers have committed to its 2027 HBM supply, while negotiations with ten additional customers are underway. The buyers are presumably organizations building AI infrastructure, but the article does not specify their identities.

Long-term contracts change the market by allocating memory before it is produced. A customer that signs early gains a clearer path to future deliveries. At the same time, more of the manufacturer’s planned output becomes unavailable to buyers that have not secured agreements. This makes timing and purchasing commitments increasingly important.

Samsung’s supply is sold out for 2027, and SK Hynix says its DRAM and HBM are effectively sold out through 2029. These announcements suggest that major memory capacity is being reserved years ahead. Smaller or later buyers may face tighter choices, although the article does not describe specific allocation rules or identify affected companies.

Manufacturers can potentially reduce a shortage by bringing additional fabrication capacity online. The article lists new capacity as a key factor to track, but it does not say how much Samsung, SK Hynix, or other manufacturers will add, or when that capacity will become available. Any relief is therefore uncertain from the information provided.

Other memory types may serve systems that do not require HBM, especially conventional DRAM in servers and workstations. However, HBM is the stacked memory placed next to AI accelerators, and the article treats it as a distinct bottleneck. It does not identify a substitute that can provide the same role or performance in AI infrastructure.

For now, production expansion and contract coverage are more concrete responses than substitution. Samsung expects AI demand to exceed industry supply in 2027, while SK Hynix reports effective sellouts through 2029. Pricing and capacity data in future quarterly reports will show whether the shortage is easing.

AI computing requires substantial memory bandwidth because accelerators repeatedly move model data and intermediate results while performing calculations. HBM is designed for this high-throughput role and is placed next to the accelerator. That arrangement makes HBM a central part of modern AI infrastructure, rather than a minor supporting component.

The key mechanism is data supply. An accelerator can perform calculations only as quickly as the required information reaches it. If memory cannot provide enough bandwidth, the accelerator’s potential performance is harder to use fully. If there is not enough HBM at all, a manufacturer may also be unable to complete additional AI servers or expand capacity.

Samsung says AI-driven demand will significantly exceed industry supply in 2027. Its HBM output for that year is already sold out, while SK Hynix reports effective sellouts through 2029. Thus memory availability can limit both system performance and the pace of AI infrastructure growth.