Your SSD Is Now Part of Your AI Memory

November 16, 2026
Published on  Updated on  
Local AI · SSD Guide

Your SSD Is Now Part of Your AI Memory

For years, the SSD's job in local AI was simple: store models and load them quickly. A new technique called SSD expert streaming turns fast NVMe storage into an extra memory tier for very large models.

Quick Answer

For most users, the SSD stores models and decides how fast they load. SSD expert streaming goes further: it reads parts of a huge Mixture-of-Experts model straight from the NVMe drive while the model runs, so models larger than your RAM can still work. For this, read performance matters. PCIe Gen4 is the practical minimum and PCIe Gen5 gives the most headroom.

Load times are theoretical, based on maximum sequential read speeds. Real-world results depend on CPU, file system and software.
~2 sec
20GB Model Load on Gen5
397B
Params Streamed on a Laptop
~27 MB
Expert Data Read per Token

The SSD's Traditional Job: Storage and Load Time

In a typical setup, the model sits on your SSD until you start it. Then it is copied into RAM or VRAM, and the SSD goes quiet.

Here the SSD affects two things: capacity, meaning how many models, adapters and datasets you can keep, and load time, meaning how long you wait when you switch models.

Drive Type Max Read 20GB Model (theoretical)
2.5" external hard drive ~120–150 MB/s Over 2 minutes
USB 3.2 Gen 2 portable SSD ~1,050 MB/s ~20 seconds
PCIe Gen4 NVMe (UD90) Up to 5,000 MB/s ~4 seconds
PCIe Gen4 NVMe (US75) Up to 7,000 MB/s ~3 seconds
PCIe Gen5 NVMe (US85) Up to 10,300 MB/s ~2 seconds

The New Job: SSD as Part of AI Memory

Mixture-of-Experts (MoE) models activate only a few "experts" for each token. Open-source developers realized this means the whole model does not need to be in RAM: the experts can stay on the SSD, and only the ones each token needs are read.

A real-world example: a widely shared open-source demo in 2026 ran a 397-billion-parameter MoE model on a laptop with 48GB of unified memory. The 4-bit expert weights, about 209GB, stayed on the internal NVMe SSD. For each token, the engine read only four small experts per layer, around 27MB in total, and generated roughly 4 tokens per second. That is slow for chat, but it proves a fast SSD can now act as an extra memory tier.

Research projects are pushing the idea further by predicting which experts will be needed next and reading them from the SSD ahead of time, which greatly improves speed.

What This Means When You Choose an SSD

  • Read performance matters more than ever. Expert streaming reads many small blocks repeatedly, so both sequential throughput and random-read performance count.
  • Capacity matters too. Very large MoE models take 200GB or more on disk. Plan 2–4TB for a dedicated model drive.
  • Cooling helps. Long streaming sessions keep the drive busy. A heatsink or good M.2 cooling helps it hold peak speed.
  • RAM is still faster. Streaming makes huge models possible, not fast. For daily use, keep your main model in RAM. See Run Bigger AI Models with 64GB DDR5.

Recommended Silicon Power SSDs for Local AI

Silicon Power NVMe SSDs
Gen5 · Best Performance

US85 NVMe SSD (1TB / 2TB)

The top choice for SSD expert streaming and frequent model switching.

PCIe Gen5 x4Up to 10,300 MB/s readUp to 8,600 MB/s write
View product →
Gen4 · Large Library

US75 4TB with Heatsink

High-speed Gen4 in a large capacity, with a heatsink for sustained reads.

PCIe Gen4 x4Up to 7,000 MB/s4TB
View product →
Gen4 · Fast Loading

US75 NVMe SSD (1TB / 2TB)

A high-bandwidth Gen4 drive for quick model loads.

PCIe Gen4 x4Up to 7,000 MB/s readUp to 6,500 MB/s write
View product →
Gen4 · Best Value

UD90 NVMe SSD (1TB / 2TB / 4TB)

Great value per terabyte for storing large model collections.

PCIe Gen4 x4Up to 5,000 MB/s
View product →

Suggested Drive Layout

Drive 1

System Drive

1–2 TBGen4 NVMe
OS+ Apps
  • US75 1TB or UD90 2TB
  • Operating system, apps and tools
  • Keeps downloads and model reads from slowing the system
Drive 2

Model Drive

2–4 TBGen4 / Gen5
Models+ Datasets
  • US85 2TB or US75 4TB
  • Active models, MoE experts and datasets
  • Dedicated bandwidth for loading and streaming
Key Takeaways
  • A Gen4 NVMe SSD loads a 20GB model in a few seconds.
  • SSD expert streaming lets very large MoE models run even when they do not fit in RAM.
  • Read performance, capacity and cooling all matter for streaming workloads.
  • Use a separate NVMe drive for models to keep the system responsive.

Frequently Asked Questions

Will SSD expert streaming wear out my SSD?

Streaming is almost entirely reads, and reads cause very little NAND wear. Wear mostly comes from writes, such as downloading new models.

Do I need a PCIe Gen5 SSD for local AI?

Not for ordinary use. A Gen4 drive loads models in seconds. Gen5 is worth it if you switch models often or want to experiment with SSD expert streaming.

Can I stream models from an external SSD?

It works, but USB is much slower than internal NVMe. Use external SSDs for backups and moving models between machines, and an internal NVMe SSD for active models.

Does a faster SSD make the model answer faster?

Once a model is fully loaded into memory, no. Generation speed then depends on memory bandwidth. A faster SSD shortens load times, and matters during generation only when you stream experts from the drive.

Local AINVMe SSDPCIe Gen5MoEExpert StreamingModel Loading

← Back to the Local AI Hardware Guide

Published on  Updated on