AI Workstation Storage - Live Prices & Sizing

Two tiers, split by cost per terabyte

Storage for local AI is not one purchase, it is two, and conflating them is the most expensive mistake in this category. There is a working set — the handful of models you actually load this week — and there is the library, the accumulated zoo of everything you have downloaded, quantized, fine-tuned or kept around because you might come back to it. Those two roles have almost nothing in common except that both consume terabytes.

The working set belongs on NVMe, and it is small. Most people running local models keep between two and six models in active rotation, which at the approximate sizes below is somewhere between ~10GB and ~200GB. You do not need eight terabytes of flash for that. What you need is enough room that you are not deleting a model to make space for the next one, which in practice means a 2TB drive for a casual user and 4TB for someone who works across model families.

The library belongs on the cheapest reliable capacity you can buy, and that is spinning disk. The gap is not marginal: at today's tracked prices the cheapest in-stock CMR hard drive and the cheapest in-stock NVMe drive are separated by roughly an order of magnitude per terabyte, and the tables further down this page show both numbers live. For data you touch occasionally, paying flash prices to store it is simply a donation.

The reason this split works is that the two tiers are not competing for the same job. Archive capacity is where the zoo lives; the working drive is where this week's models live; and moving a model between them is a file copy you initiate deliberately, not something happening in your hot path. Buy the tiers for their roles and each one is cheap. Buy one tier to do both jobs and you will overspend on flash or run out of room, usually both.

Model libraries grow, and they grow more quickly than you plan for

Every sizing guide underestimates this, so budget generously. The approximate footprints at Q4-class quantization are the starting point: a 7B-class model is ~4GB, a 13B-class ~8GB, a 34B-class ~20GB and a 70B-class ~40GB. Image generation is chunkier per item than people expect — an SDXL-class checkpoint runs ~7GB — and audio models are comparatively small at ~1-3GB for Whisper-class variants. All of these vary by quantization and none of them should be treated as exact.

The arithmetic that catches people out is not the models, it is the multiplication. You rarely keep one copy of a model. You keep the Q4 you use, the Q5 you compared it against, the Q8 you downloaded before you understood quantization, and the original weights you never deleted. A single 13B family that looks like ~8GB on paper becomes ~30GB on disk once you have four variants of it, and the same pattern repeats across every family you experiment with.

Then there is everything that is not a model. Datasets for fine-tuning, LoRA adapters, generated image output, container images for the runtimes you tried, and the checkpoints your own training runs emit. Container layers in particular are quietly enormous and almost never counted in a storage plan. None of this is exotic usage — it is what a normal six months of experimenting looks like.

The practical implication is a headroom rule rather than a precise forecast: whatever you calculate, add at least half again. The model storage calculator defaults to a fifty percent headroom allowance for exactly this reason, and in our experience that is the floor rather than a generous cushion.

What storage does and does not decide

This is the honest scope note, and it matters because a lot of writing in this space blurs it. Storage determines which models you can keep locally and what that costs you. It is a capacity and budget question. Once a model is loaded, it is resident in VRAM or system RAM, and your storage is no longer part of the picture until you load something else.

So the questions this cluster answers are: how much do I need, what does that capacity cost today, which drives are actually in stock, and what warranty comes with them. Those are our data. We track live retail prices across NVMe, hard drives, NAS devices and ECC memory, and we can tell you what a decision costs at this hour rather than at whatever hour a static guide was written.

The questions this cluster does not answer are performance ones. We do not publish performance measurements, we do not rank drives by speed, and we do not tell you one drive will feel better than another. Manufacturer figures like PCIe generation and rated sequential read appear in the tables as specification columns because they are part of the product description, but we attach no editorial interpretation to them. If you want a performance opinion, this is the wrong site, and we would rather say so than pretend otherwise.

What we will say about capacity and money, we will say precisely, because it is computed from a database we refresh every few hours rather than recalled from whenever a guide was published.

The 2026 price context you are buying into

Buying storage in 2026 means buying into a constrained market, and it is worth understanding the shape of that constraint before you decide how much to buy at once. The pressure is real, documented and upstream of anything a retail buyer can influence.

Tom's Hardware, citing DigiTimes reporting from November 2025, described QLC NAND production as booked through 2026 on datacenter demand, with QLC bit output expected to overtake TLC by early 2027. The same reporting noted SanDisk raising NAND prices by approximately 50%. On the drive side, the picture is longer-dated still: enterprise hard drives on two-year backorders, and WD confirming it is effectively sold out of hard drive production for the year.

Those are the only external figures we use, and we use them because we have already sourced them elsewhere on this site rather than because they make a better story. Our QLC versus HDD economics report works through what the flash-versus-disk gap actually is at retail, and the storage buyer's reality report covers what buying into this market looks like in practice, including the counterfeit surge that follows every documented shortage.

The practical consequence for an AI builder is that the usual advice to buy incrementally is weaker than normal. In a market with tight allocation, the drive you decided to defer may not be available at the same price when you come back for it. That is not a scare tactic — it is the same reasoning that leads us to publish live prices instead of static recommendations. Check the number, then decide.

Who needs what

The casual 7B to 13B user is the largest group and the most over-sold to. If you run two or three models in the ~4GB to ~8GB range, experiment occasionally, and do not fine-tune, your entire library fits comfortably inside a few hundred gigabytes even after headroom. A 2TB NVMe drive is not a compromise for this profile, it is generous, and the money saved against a 4TB purchase is better spent on the archive tier or on nothing at all.

The 70B runner has a genuinely different problem. At ~40GB per model at Q4-class quantization, and with the near-certainty of keeping multiple quantizations of the same weights while deciding which to settle on, a handful of 70B-class families will consume most of a terabyte before anything else is counted. This profile wants 4TB on the working drive, and it is the profile that benefits most from an archive tier, because the models being compared and discarded are exactly the ones that should not live on flash.

The image-generation collector is the profile that runs out of space first and is usually the most surprised by it. At ~7GB per SDXL-class checkpoint, a collection of a few dozen checkpoints plus LoRAs plus the generated output itself is a multi-terabyte problem within a year. This is the profile for which the two-tier split pays for itself fastest: checkpoints in rotation on NVMe, the rest of the collection and all generated output on cheap disk.

The fine-tuner is a fourth case and deserves separate mention because their data is the only irreplaceable data in this whole category. Base models are re-downloadable; your own checkpoints are not. That distinction drives the backup guidance on the model library NAS page and it should drive yours.

Buying into each tier

For the working drive, the decision is capacity first and price second, and the sequence matters. Decide how much you need with the calculator, then buy the cheapest in-stock drive that clears it with headroom. The table below lists in-stock NVMe from 2TB upward ranked by cost per terabyte, so the top row is the cheapest way to buy a given capacity today rather than a recommendation we wrote last quarter.

For the archive tier, the ranking column is the same and the logic is simpler: cost per terabyte, in stock, CMR recording. We restrict archive picks to CMR because shingled drives behave badly in the rewrite-heavy patterns that a growing library produces, and because CMR is the right default for anything that might end up in a RAID array later. Condition is worth a look rather than an automatic new purchase — a refurbished enterprise drive with warranty remaining is a legitimate archive buy and frequently the cheapest terabyte on the page.

The two tables below are the whole point of this page. Everything above is reasoning that a static guide could also give you; the numbers below are the part that goes stale within hours anywhere else and is current here. If you want the ranked view across every capacity we track rather than the top rows, cheapest per TB has the full list, and enterprise NVMe covers the U.2 and U.3 side for buyers with server chassis rather than desktops.

Frequently asked questions