Layer 2 — Memory & storage

Position in the die→grid hierarchy: inside the server.

Memory and storage attached to the accelerators and servers: HBM stacked on the package, DDR5 on the board, the interface silicon between them, and the NAND, SSDs and nearline drives that hold checkpoints and training data.

Parts in this layer

• 2.01 HBM (HBM3E, HBM4)constrained · 6 companies
  High-bandwidth memory stacked directly on the same package as an AI accelerator, giving the GPU or ASIC very fast, close-range access to the memory it needs to stay fed.
• 2.02 DRAM (DDR5 RDIMM/MRDIMM, LPDDR5X for Grace)constrained · 6 companies
  Off-package system memory — DDR5 modules for host CPUs and LPDDR5X soldered near Nvidia's Grace CPU — that holds the working data a server's processors operate on.
• 2.03 Memory modules · 7 companies
  The physical DIMM (or module) that packages raw DRAM die, substrate, and management circuitry into the pluggable unit a server actually sockets.
• 2.04 Memory interface / buffer chips (RCD, DB, MRCD/MDB) · 3 companies
  Small companion chips on a memory module that buffer and re-drive command, address, and data signals so many DRAM chips can be packed per module without losing signal integrity.
• 2.05 CXL controllers & memory expanders · 8 companies
  Chips that let a server attach an extra pool of DRAM over a CXL link rather than only through the CPU's own memory channels, expanding usable capacity beyond what fits in local DIMM slots.
• 2.06 PCIe/CXL switches & retimers · 5 companies
  Fabric silicon — switches that fan a single PCIe or CXL root out to many downstream devices, and retimers that clean up the signal over longer runs — connecting accelerators, SSDs, and pooled memory back to the host.
• 2.07 NAND flash · 6 companies
  Non-volatile flash memory die — the raw storage medium later packaged into SSDs — that holds data and model artifacts when the system is powered off.
• 2.08 Enterprise SSDs & controllers · 15 companies
  Solid-state drives built for data center duty cycles, and the controller chips that manage them — the fast local storage tier that feeds checkpoints, staged datasets, and cached weights to accelerators.
• 2.09 HDD (nearline / object tiers) · 3 companies
  Spinning hard disk drives used for the bulk, lower-cost nearline and object storage tier that holds data too large or too cold to justify keeping on flash.
• 2.10 AI storage systems (parallel file systems, object, GPU-direct) · 14 companies
  Purpose-built storage software and appliances — parallel file systems, object stores, and GPU-direct pipelines — that aggregate many drives into one system capable of feeding a large number of accelerators at once.

Public plays in this layer

Listed companies with the most part placements in layer 2. Placement count is breadth, not revenue exposure.