Types of memory

1) DRAM / HBM

Role in AI Compute:

DRAM (Dynamic RAM): Main system memory for CPUs/GPUs/accelerators. Holds working datasets, model parameters, intermediate activations, and general compute state while processing.
HBM (High-Bandwidth Memory): Specialized DRAM tightly stacked near GPU/accelerator cores. Provides very high throughput at low power compared to DDR DRAM, critical for large matrix multiplications and transformer layers in modern AI.
In AI contexts, DRAM is about capacity + latency, while HBM is about bandwidth.

Main Suppliers (Public + Tickers):

Micron Technology — MU (NYSE) — DRAM & HBM supplier via ecosystem partners.
SK Hynix — 000660.KS (Korea) — DRAM & HBM; one of the largest memory producers globally.
Samsung Electronics — 005930.KS (Korea) — #1 in DRAM and DRAM-derived products (including HBM).
Nanya Technology — 2408.TW (Taiwan) — DRAM maker (smaller share).

Role in AI Compute:

Primary non-volatile storage for SSDs and persistent buffers.
Used for model storage, weights caching, checkpoints, datasets that don’t fit in RAM and need persistent backing.
Helps enable fast I/O to feed data into faster DRAM/HBM layers.

Main Suppliers (Public + Tickers):

Sandisk – SNDK
Western Digital — WDC (NASDAQ) — NAND flash via joint ventures and SSDs.
Micron Technology — MU (NYSE) — Major NAND supplier.
Kioxia (private) — large NAND supplier (not publicly traded as of 2026).
SK Hynix — 000660.KS (Korea) — NAND products after acquiring Toshiba’s memory business stake.
Samsung Electronics — 005930.KS (Korea) — Largest NAND flash producer.

Role in AI Compute:

Low-cost, high-capacity bulk storage for datasets, archives, logs, backups.
Not used for active model training but essential for data lakes, long-term dataset retention, backup copies of models, and cold storage in AI pipelines.

Main Suppliers (Public + Tickers):

Seagate Technology — STX (NASDAQ) — One of the largest HDD makers for enterprise & cloud.
Western Digital — WDC (NASDAQ) — Major HDD maker alongside Seagate.

Role:

Ultra-fast memory used on-chip in CPUs/GPUs/accelerators for caches, register files, and local buffers.
Critical for low-latency access during computation.

Suppliers:

SRAM is built into chips; there aren’t standalone SRAM producers with significant public tickers. It’s designed inside logic chips by:
- Nvidia — NVDA (NASDAQ)
- AMD — AMD (NASDAQ)
- Intel — INTC (NASDAQ)
- AI ASIC firms (e.g., Google/Alphabet GOOGL, Apple AAPL)

Role:

Uses NAND flash + controller to provide high-speed persistent storage for datasets and model artifacts.
Often tiered with DRAM caches to speed I/O for AI training pipelines.

Suppliers:

These aren’t yet mainstream in AI compute but are relevant in research and low-power AI systems:

Memory Type	Role	Notes
MRAM (Magneto-Resistive RAM)	Non-volatile, faster than NAND	Early stage; used in embedded systems
ReRAM / PCM	Persistent, low power	Research for in-memory compute
Optane (3D XPoint)	Persistent memory between DRAM & NAND	Intel inventor (legacy products)

Suppliers:

MRAM & emerging:
- Everspin Technologies — MRAM (MRAM) (NASDAQ)
- Intel — INTC (NASDAQ) (historical with Optane/3D XPoint)
- Samsung / SK Hynix / Micron are researching next-gen memory but not always productized.

Layer	Typical Tech	Role in AI
On-chip registers & cache	SRAM	Fastest access for computation
High-bandwidth pool	HBM	Training/inference compute throughput
System memory	DRAM	Working set + dataset buffering
Persistent fast storage	SSD (NAND flash)	Model/dataset I/O
Bulk/archive storage	HDD	Long-term dataset/backup