TL;DR
Samsung pre-allocated 60-70% of its future memory capacity into five-year supply agreements with major data center companies, part of roughly $950 billion in deals across Korean chipmakers. The move signals a coming memory shortage through 2027-2028 and gives Samsung guaranteed revenue while locking competitors out of supply.
The Deals
Samsung’s supply agreements represent the largest memory chip deals in history:
- Duration: Five-year agreements with major cloud providers and AI companies
- Capacity: 60-70% of Samsung’s future memory production
- Total value: Part of $950 billion in combined deals across Samsung, SK Hynix, and other Korean chipmakers
- Focus: High Bandwidth Memory (HBM) and advanced DRAM for AI training
The agreements include provisions for capacity expansion, ensuring that Samsung can meet growing demand for AI memory while guaranteeing buyers access to supply.
Samsung’s foundry AI-revenue share is expected to exceed 30% this year, reflecting the growing importance of AI to its business.
Why Memory Matters
AI training requires massive amounts of memory:
- GPU memory: Each Nvidia H100 GPU has 80GB of HBM3 memory
- Training clusters: Large language model training requires thousands of GPUs with pooled memory
- Inference: Memory bandwidth is the primary bottleneck for AI inference performance
- Growth: AI memory demand is growing 3-4x faster than supply
The memory shortage is driven by several factors:
- HBM demand: AI chips require specialized HBM memory that has limited manufacturing capacity
- DRAM conversion: Converting existing DRAM lines to HBM reduces standard memory production
- New fabs: New memory fabrication facilities take 2-3 years to build and equip
- Demand growth: AI training and inference workloads are growing exponentially
Market Impact
Samsung’s pre-allocation has several implications:
- Competitor access: Smaller companies may struggle to secure memory supply at competitive prices
- Price increases: Memory prices are expected to rise through 2027-2028
- Supply chain risk: Dependence on Korean chipmakers for AI memory creates geopolitical risk
- Innovation pressure: Companies may develop more memory-efficient AI architectures to reduce demand
For the AI industry, the memory shortage is a critical constraint. Training large language models requires enormous amounts of memory, and limited supply could slow the pace of AI development.
Samsung’s strategy of locking in long-term agreements ensures revenue stability while positioning the company as the dominant supplier of AI memory. Whether this benefits the broader AI ecosystem — or simply concentrates power in the hands of a few chipmakers — remains to be seen.