Samsung Electronics has reportedly committed roughly 70% of its memory production capacity to long-term supply agreements running through 2031, a sign of how far AI infrastructure demand is reshaping the DRAM and HBM market. The report comes from South Korean newspaper Seoul Economic Daily, which names NVIDIA, Microsoft, and Google among the main counterparties to these deals.
Samsung’s memory story in 20 seconds
- Samsung has reportedly allocated around 70% of its memory capacity to long-term contracts running through 2031.
- NVIDIA, Microsoft, and Google are among the major customers cited.
- HBM3E bought on the spot market can cost four to five times more than under certain contracts.
- The shift to HBM4 is adding more pressure to DRAM capacity.
- Samsung and SK hynix are weighing further production expansions.
The 70% figure needs some context. During its second-quarter 2026 earnings presentation, Samsung had already said that if ongoing negotiations were completed, it expected its multi-year DRAM and NAND contracts to reach between 60% and 70% of planned capacity. The new report out of Korea now claims, citing industry sources, that roughly 70% of its memory division’s capacity is tied up in long-term agreements through 2031.
The shift has implications well beyond Samsung. For decades, memory has been one of the most cyclical segments of the semiconductor industry. When chips ran short, prices rose and manufacturers expanded capacity. When oversupply hit, prices could collapse quickly.
AI’s expansion is partly rewriting that model, because the biggest buyers want to lock in memory years ahead of time.
HBM is becoming a resource Big Tech wants to reserve
High Bandwidth Memory (HBM) has become one of the essential components in the accelerators used to train and run AI models.
A GPU or AI accelerator needs to move enormous amounts of data between the processor and memory. HBM stacks multiple layers of DRAM and connects them using advanced packaging technologies to deliver far higher bandwidth than conventional memory.
The problem is that producing HBM consumes industrial resources that could otherwise go toward other types of memory.
HBM4’s arrival is intensifying that pressure.
Seoul Economic Daily notes that the new generation is still in the early stages of volume production and has lower manufacturing yields than HBM3E. Ramping up HBM4 output can therefore consume a larger share of DRAM capacity while manufacturers gradually improve their processes — the same tight supply picture that has NVIDIA reportedly reconsidering how much HBM it packs into Rubin Ultra.
That trend already shows up in South Korean trade statistics cited by the newspaper.
Between May and July, the exported volume of DRAM used in AI applications, including HBM and LPDDR, reportedly fell 13.2%, from about 681.7 million to 591.74 million units. However, the total value of those exports rose 18.5%, to $13.55 billion.
The average price per unit climbed from $16.76 to $22.90, a 36.6% increase in two months.
It’s an unusual combination: fewer units exported, but considerably more revenue.
The spot market shows just how tight HBM supply is
The gap between long-term contracts and one-off purchases offers another snapshot of the current strain.
According to figures published by Seoul Economic Daily based on analyst firm MegaGrid Supply, a 36 GB HBM3E module bought on the spot market currently runs around 2.87 million South Korean won, about $2,100.
Similar products supplied under certain long-term contracts reportedly go for between 500,000 and 700,000 won.
That means buying HBM3E on the spot market can cost four to five times more than under those agreements.
The gap doesn’t mean any customer could sign a contract today and automatically buy HBM five times cheaper.
Multi-year contracts involve volume commitments, negotiated terms, possible upfront payments, and obligations spanning several years. Their specific terms also aren’t public.
The comparison mainly illustrates the price being paid for immediate availability.
| Memory | Situation cited | Approximate price |
|---|---|---|
| HBM3E 36 GB | Long-term contracts | 500,000-700,000 won |
| HBM3E 36 GB | Spot market | 2.87 million won |
| 16-layer HBM4 | Spot market | 4.8 million won |
| HBM3E gap | Spot vs. certain LTAs | 4-5x |
The situation is even more extreme with HBM4.
The Korean newspaper puts a 16-layer HBM4 product ready for volume production at around 4.8 million won, about $3,500, on the spot market.
That figure shouldn’t be read as a universal HBM4 price. Memory built for AI accelerators is manufactured and validated to each customer’s specific requirements, and prices can vary considerably by volume, generation, capacity, packaging, and contract terms.
Samsung is trying to reduce its exposure to the memory cycle
Long-term contracts also serve a clear purpose for manufacturers.
Samsung has said it’s moving toward multi-year agreements and a structure it calls five-year rolling contracts, whose base term can reach five years and can be extended through later negotiations. Some deals also include upfront payments.
That gives the manufacturer something especially valuable before it commits billions to new facilities: visibility into future demand.
Historically, one of the memory industry’s biggest challenges has been deciding how much to expand fabs by.
Building capacity when prices are soaring can look like an obvious call. But a fab takes years to reach full production. If demand has cooled by the time it does, the market can flip quickly from shortage to oversupply.
Multi-year contracts shift part of that risk.
Samsung gets earlier visibility into certain volumes its customers want to buy, while large tech companies can reserve capacity before they physically need it.
Samsung is already applying this kind of strategy beyond the customers named in the South Korean report. In July it announced a memorandum of understanding with Broadcom to expand their collaboration on memory, manufacturing, and advanced packaging for AI infrastructure over the coming years. Samsung said both companies estimated the scope of the collaboration at more than $200 billion through 2030, though that figure spans several areas and isn’t limited to HBM.
HBM’s expansion could squeeze the rest of the memory market
The most significant effect for the tech market is that the race for HBM doesn’t happen inside a fab that’s fully separate from the rest of a manufacturer’s product line.
Industrial capacity is limited.
Manufacturers have to decide where to put their investment, wafers, equipment, and packaging processes. If AI-bound products deliver higher margins and come backed by multi-year contracts, there’s a strong incentive to prioritize them.
That can keep supply of other memory types tight.
The pressure isn’t limited to the GPUs used in data centers. Conventional DRAM, server memory, and other products compete indirectly for investment and capacity within an industry where building new fabs takes years.
Samsung and SK hynix are, in fact, weighing new expansions.
According to Seoul Economic Daily, Samsung is considering converting S5, currently the only foundry line at its Pyeongtaek complex, to memory production starting next year. SK hynix, which is already mapping out its next HBM generations, is looking at various expansions, including a possible joint plant in Japan. These are reported possibilities based on industry sources and shouldn’t yet be treated as fully approved projects in all their details.
The message coming from the investments the industry has already announced is similar: manufacturers don’t expect demand for advanced AI memory to fade any time soon.
AI is also changing how memory gets bought
The 70% figure may end up mattering more for the commercial shift it represents than for the number itself.
Microsoft, Google, NVIDIA, and other large buyers need to plan data centers and accelerators years in advance. GPU availability also depends on having enough HBM to pair with them.
Reserving manufacturing capacity becomes a way to secure future infrastructure.
For smaller buyers, the opposite problem appears.
If a large share of future production is already locked up in multi-year contracts, less capacity is left exposed to short-term negotiations and the spot market.
The current gap between HBM3E’s contract and spot prices shows what can happen when demand far outstrips available supply.
None of this means the memory market’s historical cyclicality has disappeared.
Contracts can stabilize part of demand, but risks tied to overinvestment, technology shifts, manufacturing yields, and AI growth forecasts still remain. The exact terms of many long-term agreements also aren’t public, making it hard to tell how much risk each manufacturer takes on versus how much falls on the buyer.
Samsung and SK hynix also appear to be taking different contractual approaches. Reports published during August suggest Samsung is seeking to commit a larger share of its capacity, while SK hynix is reportedly keeping more production outside long-term contracts.
What does seem clear is that memory is no longer being bought with just next quarter in mind.
If Samsung ends up with around 70% of its capacity committed to agreements running through 2031, major AI projects will be reserving critical components years before the servers that will use them ever go online.
And that helps explain why AI’s pressure is no longer just about GPU prices.
It’s also shaping how much memory will cost, and how much will be available, for the rest of the industry.
Frequently asked questions
What share of Samsung’s memory production is reportedly reserved?
Seoul Economic Daily reports, citing industry sources, that around 70% of the memory division’s capacity is allocated to long-term supply agreements running through 2031. Samsung had previously said it expected to reach between 60% and 70% if ongoing negotiations were completed.
Which companies have long-term memory contracts with Samsung?
The South Korean report names NVIDIA, Microsoft, and Google among the main counterparties. Samsung hasn’t published the individual terms of those deals.
Why could HBM4 add more pressure to DRAM supply?
HBM uses multiple layers of DRAM and requires complex manufacturing and packaging processes. Ramping up production of generations like HBM4 consumes industrial capacity that could otherwise go to other memory products.
Why is HBM so much more expensive on the spot market?
The spot market reflects immediate availability at a time of severe shortage. Multi-year contracts secure large volumes over long periods under different negotiated terms, so their prices aren’t directly comparable to a one-off purchase.
