Best Drives for an AI Workstation - Live Prices

Sizing the working drive

The working set is the models you load in a given week, and for most people it is a small fraction of what they have downloaded. At the approximate published footprints — ~4GB for a 7B-class model, ~8GB for 13B-class, ~20GB for 34B-class and ~40GB for 70B-class, all at Q4-class quantization and all varying with it — a rotation of several models fits inside a couple of terabytes with room to spare.

That is why we push back on the instinct to buy the largest flash drive available. The capacity that actually grows is the archive, and archive capacity costs roughly an order of magnitude less per terabyte at the prices we track. Money spent on flash you will not fill is money not spent on the tier that fills up.

The exception is image generation, where SDXL-class checkpoints at approximately 7GB each accumulate quickly and generated output accumulates more quickly still. That profile justifies four terabytes on the working drive sooner than a language-model-only profile does.

Size it properly rather than by feel: the model storage calculator does the arithmetic against the approximate sizes and returns a capacity tier with a live price attached.

The archive tier and why it is CMR

Everything not in active rotation belongs on the cheapest reliable capacity you can buy. Superseded quantizations, model families you have moved on from, finished datasets, container images and generated output are all cold data, and paying flash prices to store cold data is the largest avoidable expense in this category.

We restrict archive picks to CMR recording for two reasons. A growing library is a rewrite-heavy pattern over time, which is where shingled drives behave worst, and CMR is the right default for anything that might end up in a RAID array later. Neither of those is a performance claim; both are about how the drive handles the write pattern a library produces.

Capacity planning here should follow your growth rate rather than your current usage. The reliable observation across every local-AI setup we have looked at is that libraries only grow, because deleting requires a decision and downloading does not.

Endurance, condition and warranty

Endurance matters for one specific profile: people writing large files repeatedly. Fine-tuning that emits checkpoints on a schedule, continuous image generation, and frequent re-downloading and re-quantizing of large model files all write far more than a typical desktop workload. The manufacturer figure for this is terabytes written, and we treat it as what it is — a warranty parameter stating how much writing the manufacturer will stand behind, not an indicator of how a drive behaves.

For read-heavy use, which describes most people running inference on models they downloaded once, endurance should not drive the purchase. Models are written once and read thereafter.

Condition guidance splits cleanly by tier. Buy the working drive new: it is the drive that holds work in progress and the price difference at these capacities is modest. For the archive tier, refurbished enterprise drives are a legitimate and frequently cheaper purchase, since the data on them is largely re-downloadable and the drives carry reseller warranties. Our recertified coverage explains what those warranties actually oblige, which is often narrower than the headline term suggests.

Two drives, or one drive and a plan

The question people ask before buying is usually whether to get one large drive or two smaller ones, and for an AI workstation the answer is shaped by the two-tier split rather than by a general rule.

If you are buying a single drive, buy it for the working set and accept that the library will need somewhere else to live before long. A 2TB or 4TB NVMe drive plus a plan for archive capacity is a better position than a single 8TB flash drive that has absorbed the whole budget, because the archive capacity you eventually need costs a fraction per terabyte of what you would have paid to pre-buy it in flash.

If you are buying two, the split is straightforward: NVMe sized to what you load, and the largest CMR drive that clears your cost-per-terabyte threshold for everything else. Most desktop boards have more than one M.2 slot and every one has SATA ports, so this is rarely a physical constraint.

There is one case where a single large flash drive is genuinely right, and it is worth naming so the advice is not dogmatic: if you curate aggressively, delete what you stop using, and your entire library is your working set, then a second tier saves you nothing and adds a step. That describes a minority of people, but it does describe some, and if it describes you then buy the flash and skip the rest of this.

For everyone else the archive tier is where the capacity goes, and it is the cheaper half of the purchase by a wide margin.

Capacity ladders and what they cost today

The two tables on this page are the part that a static guide cannot give you. The first ranks in-stock NVMe by cost per terabyte; the second does the same for CMR archive drives. Out-of-stock listings are excluded from both, because a ranking built on a price you cannot pay is not a recommendation.

Work the ladder rather than picking a number. Cost per terabyte does not rise smoothly with capacity — the next tier up is periodically cheaper per terabyte than the one below it, which changes the right answer from week to week. Checking the current figure takes a moment and occasionally saves a meaningful amount.

One piece of market context worth carrying into the decision: QLC production is booked through 2026 and SanDisk raised NAND prices approximately 50% (Tom's Hardware citing DigiTimes, November 2025), while enterprise hard drives sit on two-year backorders. In a rationed market, the usual advice to defer a purchase is weaker than normal, because availability is as much at risk as price.