Part of Running LLMs on Your Own Hardware: From “Will It Even Fit” to a Model That Does Your Work — running large language models on your own hardware.
Choosing SSD storage for local AI ultimately comes down to two numbers: how many gigabytes your models occupy, and how fast the drive can push them into memory. A single quantised 35B model needs about 20 GB, and reading that file took 44.3 seconds on SATA against 8.9 seconds on NVMe in direct measurement. This article covers what to buy, how much capacity to plan, and why storage prices turned hostile in 2026 exactly when AI workloads started demanding more of it.
- One quantised 35B model in GGUF format occupies 19.92 GiB; a working set of three or four models passes 60 GB.
- NVMe delivered 2,339.6 MB/s against 461.6 MB/s on SATA in the same machine — a measured 5.07× difference.
- Enterprise SSD contract prices rose roughly 80% in Q1 2026, and client SSDs are forecast to climb over 40%.
- Plan for 2 TB minimum if you run local models seriously; 1 TB fills within weeks of experimenting.
- HDDs still cost 7 to 8 times less per terabyte, which keeps them useful for datasets and archives but not for model files.
How much SSD storage for local AI you actually need
In practice, model files are larger than most people expect, and quantisation only softens the blow. The workstation behind this article runs Qwen3.6-35B-A3B, and its files measure exactly what the table below shows.
| Artefact | Size | Notes |
|---|---|---|
| Qwen3.6-35B-A3B-UD-Q4_K_S.gguf | 19.92 GiB | 4-bit quantisation, measured |
| mmproj-F32.gguf | 1.66 GiB | vision projector for the same model |
| Typical 7B model, Q4 | 4–5 GB | the smallest useful tier |
| Typical 70B model, Q4 | 38–42 GB | needs 48 GB VRAM or offloading |
| Whisper large-v3 | ~3 GB | speech transcription |
| Embedding model | 0.5–2 GB | required for any RAG setup |
However, those numbers describe one copy of each. In practice you keep several quantisations of the same model to compare quality, which multiplies the total quickly. Furthermore, tools such as Ollama and LM Studio maintain their own caches, so the same weights can exist twice on disk.
Moreover, add the surrounding infrastructure and the picture sharpens. Vector databases grow with every document you index. Fine-tuning checkpoints run into tens of gigabytes each. Container images for inference stacks routinely exceed 10 GB. Consequently a 1 TB drive that looked generous in 2023 now fills within a few weeks of serious experimentation.
Why the 2026 storage market changed the maths
Unfortunately, buying advice from last year no longer applies, and the reason is structural rather than seasonal. AI infrastructure has absorbed the available NAND supply.
Specifically, enterprise SSD contract prices rose about 80% during the first quarter of 2026, according to market reporting on the memory shortage. Client SSD prices are forecast to rise over 40% across the year. Some popular Gen4 drives already cost double or triple their late-2024 price.
In fact, the cause is a deliberate allocation decision. Samsung and SK Hynix have cut NAND output to prioritise high-bandwidth memory, because HBM serves AI accelerators at gross margins above 60%. As a result the same technology wave that makes you want more local storage is the one making it expensive.
Meanwhile, hard drives offer no escape either. Enterprise drives are reportedly backordered beyond 52 weeks, and helium-sealed models saw 20 to 30% increases in March 2026. No meaningful new NAND capacity is expected before 2027 or 2028, so relief is not close.
SATA, NVMe and HDD compared for AI workloads
In short, the three storage types differ by roughly an order of magnitude at each step, and the practical consequences are easy to state.
| Type | Sequential read | Time to load 20 GB | Price per TB | Best use |
|---|---|---|---|---|
| NVMe PCIe 4.0 | 2,340 MB/s measured on a Gen3 link | 8.9 s | $90–200 | active model files, OS |
| SATA SSD | 461.6 MB/s measured | 44.3 s | $70–110 | secondary storage, datasets |
| HDD 7200 rpm | 150–280 MB/s typical | ~90 s | $12–22 | archives, backups, cold data |
However, two caveats matter. First, the NVMe figure came from a PCIe 3.0 workstation, so the drive was operating well below its rating. Second, the measurement used a single-threaded queue depth of one, which no NVMe drive can saturate. A modern desktop with a Gen4 link and parallel reads would show a wider gap still.
Free ebook
Free AI Video, Generated Locally
Working scripts and measured benchmarks. Free.
No spam. Unsubscribe at any time.
Above all, notice what the middle column means in daily use. Loading a model from HDD takes about a minute and a half every time the server restarts. On NVMe it takes under ten seconds. Over a working day of experimentation that difference compounds into real minutes lost.
What to buy for local AI storage in August 2026
Accordingly, the table below records street prices as gathered on 10 August 2026. Prices vary widely between retailers right now, so treat these as a snapshot rather than a quote.
| Drive | Capacity | Price (USD) | Approx. $/TB | Source |
|---|---|---|---|---|
| WD Black SN850X | 2 TB | 124–156 | 62–78 | Amazon listings, price trackers |
| Kingston KC3000 | 2 TB | ~298 (sale 246) | ~149 | Pangoly price history |
| Samsung 990 Pro | 2 TB | 165–385 | 83–193 | varies sharply by retailer |
| Samsung 990 EVO Plus | 4 TB | ~230 (promo) | ~58 | promotional listing |
| Crucial T500 | 4 TB | 310–320 | 78–80 | retail listings |
| 4 TB SSD, market average | 4 TB | ~374 | ~94 | price tracking aggregate |
Incidentally, the spread on the Samsung 990 Pro deserves comment, because a range from $165 to $385 for the same product is not a rounding error. It reflects a market where quotes stay valid for a week and stock moves unpredictably. Therefore check two or three retailers on the day you buy.
Overall, for a local AI workstation the sensible pick is a 2 TB TLC drive with DRAM cache from a tier-one brand. DRAM-less designs save money but stumble on sustained writes, which matters when you download a 40 GB model. The drive used in this migration, a Kingston KC3000, is a TLC design with DRAM and performed exactly as expected.
Sizing rules for AI storage that survive a year
Fortunately, capacity planning for AI storage follows a simple rule: estimate your working set, then double it. Models arrive faster than you delete them.
1 TB suits occasional use — one or two models, no fine-tuning, no large datasets. It fills quickly once you start comparing quantisations. 2 TB is the practical baseline for regular local inference, leaving room for four or five models plus the operating system. 4 TB makes sense if you fine-tune, keep datasets locally, or run several model families side by side.
In addition, a split configuration works better than one large drive. Keep the operating system and active model files on NVMe, and put datasets, checkpoints and archives on cheaper storage. The migration described in this series ended with 1.9 TB of NVMe for the system and models, plus a repurposed 447 GB SATA SSD for data.
Finally, one more consideration is often overlooked. SSD endurance is rated in terabytes written, and downloading models is write-heavy. After several years the source drive in this project showed 3% wear, which is entirely healthy. Nevertheless a heavily used AI workstation writes far more than a typical office machine.
Backups grow with the models
Capacity planning usually stops at the working drive, yet backups scale with it. Consequently a 2 TB model library needs somewhere to be copied, and that somewhere costs money too.
Fortunately model weights are the easiest data to lose safely. Because published models can be downloaded again, they rarely justify backup space. Instead, protect what cannot be regenerated: fine-tuning checkpoints, curated datasets, prompt libraries and evaluation results. In other words, back up your work rather than someone else’s weights.
That distinction changes the arithmetic considerably. A workstation holding 1.5 TB of models might need only 200 GB of genuine backup coverage. As a result an external HDD at roughly $15 per terabyte covers it comfortably, while an equivalent SSD would cost several times more for no practical gain.
Nevertheless one rule still applies. A second copy inside the same machine is not a backup, since it shares a power supply, a filesystem and a thief. Therefore keep at least one copy on a separate device, and ideally one off-site.
The upgrade path from an existing machine
Generally, adding storage is straightforward. Replacing the system drive is not, mainly because Windows makes assumptions about disk identity that cloning breaks.
Therefore, if your workstation still boots from a SATA SSD, the migration is worth doing for the load-time gain alone. The full procedure, including rescue-medium verification and drive identification, is documented in how to clone Windows to an NVMe SSD. The failures that follow a clone — and there were five of them — are covered in what breaks after cloning.
For the measured effect on actual model loading rather than raw throughput, see the NVMe versus SATA model loading benchmark.
Frequently asked questions
How much SSD storage do I need for local AI?
As a rule, plan 2 TB as a baseline. One 35B model occupies about 20 GB, and a realistic working set of four models with tooling and caches passes 150 GB quickly.
Does faster SSD storage for local AI make models run faster?
Rather, it makes them load faster, not generate faster. Once weights sit in VRAM, inference speed depends on the GPU. Loading a 20 GB model took 8.9 seconds from NVMe against 44.3 seconds from SATA.
Is a PCIe 5.0 SSD worth it for AI work?
Rarely. Indeed, model loading is sequential and happens once per session, so a Gen4 drive already removes the bottleneck. Spend the difference on capacity instead.
Can I keep models on an external USB drive?
Technically you can, though USB 3.2 caps at roughly 1,000 MB/s in the best case and far less in practice. It works for archives; it is frustrating for models you load daily.
Why are SSD prices rising in 2026?
AI datacentre demand consumed NAND supply while manufacturers shifted capacity toward high-bandwidth memory. Enterprise SSD prices rose around 80% in Q1 2026, and consumer drives followed.
Should I buy an HDD for AI datasets?
Yes, for cold data. Specifically, at $12 to $22 per terabyte HDDs remain seven to eight times cheaper than SSDs, which suits archives, backups and datasets you stream rather than random-access.
Does DRAM-less matter for a model drive?
Certainly it matters for sustained writes. DRAM-less drives slow noticeably once their SLC cache fills, which happens while downloading a large model. Choose a design with DRAM if the budget allows.
Summary
SSD storage for local AI is now a capacity problem and a market problem at the same time. A single quantised 35B model occupies 19.92 GiB, a serious working set passes 150 GB, and NVMe reads that data 5.07× faster than SATA on identical hardware — 2,339.6 MB/s against 461.6 MB/s in five measured runs each.
Buy 2 TB of TLC NVMe with DRAM if you run models regularly, keep cold data on cheaper storage, and buy sooner rather than later. Enterprise prices climbed roughly 80% in the first quarter of 2026, client drives are forecast above 40% for the year, and no new NAND capacity is expected before 2027. The prices in this article were read on 10 August 2026 and are recorded here precisely so they can be compared later.
Free ebook
Free AI Video, Generated Locally
Run Wan 2.1 in ComfyUI on your own GPU — the scripts I use, measured times, sample clips. No cloud, no API keys.
No spam. Unsubscribe at any time.
Leave a Reply