Best NVMe Drives for Local AI - Live $/TB

What we rank on, and what we do not

Every other guide to NVMe drives for AI work ranks on speed. We do not, and the reason is not modesty about our data — it is that we have genuinely good data on price, capacity, availability and warranty, and no data at all on performance. Publishing a speed ranking would mean repeating someone else's published numbers and dressing them as our own analysis.

So the ranking column on this page is cost per terabyte, computed from live in-stock listings. The manufacturer's rated sequential read and the PCIe generation appear in the table because they are part of the product specification and you are entitled to see them, but we attach no interpretation to them. If a drive is cheaper per terabyte and in stock, it is higher in our table. That is the whole method.

Availability deserves more weight than it usually gets in 2026. A drive that is out of stock is not a recommendation regardless of how well it reviews, and this page excludes out-of-stock listings from the rankings entirely rather than showing them greyed out with a stale price. What you see is what you can buy at the moment the page loaded.

Warranty is the fourth axis and the one buyers most often skip. It is a real difference between products, it is documented by the manufacturer rather than inferred, and it matters more in a market where replacement stock is tight. Our warranty coverage explains what the terms actually oblige, which is frequently narrower than the headline year count implies.

Gen4 or Gen5, decided by the price difference

The Gen4 versus Gen5 question is usually argued on performance, which puts it outside what we will claim. It can be argued entirely on price, which is inside it, and the price argument turns out to be the more decisive one for most buyers anyway.

The comparison below is computed live and like-for-like: the same M.2 2280 form factor, the same capacity, both generations in stock, and a minimum number of listings on each side before we will quote a premium at all. Comparing a Gen5 drive at one capacity against a Gen4 drive at another, or a client drive against an enterprise one, produces a number that describes product mix rather than generation, and we would rather show nothing than show that.

The way to read the resulting premium is as a straightforward question: is what Gen5 offers you worth that percentage of your storage budget? We are not going to answer it for you, because answering it requires exactly the performance judgment we do not make. What we will point out is that the premium is a percentage of a purchase you are making today, in a market where the same money spent on capacity buys a definite and measurable amount of additional space.

For buyers without a specific reason to need the newer generation, the difference is money kept. For buyers who do have that reason, they generally know what it is, and the premium below tells them what it costs. Both tables are current as of page load.

Capacity guidance against real library sizes

Two terabytes is the entry point and it is a genuinely sufficient one for a large share of local AI users. At the approximate footprints — ~4GB for a 7B-class model, ~8GB for 13B-class — a working set of several models plus room to experiment fits inside 2TB with comfortable headroom. Buying more than this because a guide implied AI needs enormous storage is the most common overspend in the category.

Four terabytes is the standard recommendation for people working across model families or running 70B-class models, where a single family at ~40GB plus its quantization variants consumes a meaningful fraction of a terabyte on its own. It is also the right tier for image generation work, where SDXL-class checkpoints at ~7GB each accumulate more quickly than language models do and generated output accumulates more quickly still.

Eight terabytes is a hoarder tier and we mean that descriptively rather than pejoratively. If you keep everything, work across many families, generate a lot of image output and would rather not think about storage again for two years, it is a rational purchase. Check the cost per terabyte before assuming it is the expensive option — capacity tiers do not price linearly, and the largest tier is periodically the cheapest per terabyte on the page.

Below two terabytes we would generally rather you bought archive capacity. A 1TB working drive plus cheap disk for the library will serve most people better than a 2TB drive alone, at a similar total price. The calculator will tell you which side of that line you fall on.

Endurance as a specification, for people who write a lot

Endurance is worth attention for one specific profile: people who write large files repeatedly. Fine-tuning that emits checkpoints on a schedule, image generation pipelines that write continuously, and anyone regularly re-downloading and re-quantizing large model files are all writing far more than a typical desktop workload.

The manufacturer figure for this is terabytes written, and it is a warranty parameter rather than a performance one. It states how much writing the manufacturer will stand behind over the warranty period. We treat it exactly that way — as a specification and a coverage question, not as an indicator of how a drive will behave day to day.

For a read-heavy user, which describes most people running inference on models they downloaded once, endurance is close to irrelevant and should not drive the purchase. The models are written once and read thereafter. Spending extra for a higher endurance rating on a drive that will spend its life being read is spending for a specification you will not consume.

Where it does matter, it interacts with warranty in a way worth understanding before you buy rather than after: exceeding the rated writes can affect coverage. Our warranty section covers what the terms actually say, and the drive lifespan page covers replacement economics for people planning a multi-year workload.

What to skip

The first thing to skip is unbranded flash at prices that do not make sense. Every documented shortage in storage history has pulled counterfeits and relabels into the market, and 2026's price levels make that incentive stronger than it has been in a decade. A drive priced far below every other listing at the same capacity is a claim that deserves scepticism rather than a bargain that deserves a click. Our storage buyer's reality report covers the counterfeit surge in detail.

The second is treating the NAND type as a verdict rather than a disclosure. QLC and TLC are different products with different specifications, and which one a drive uses is information you are entitled to have before buying. It is not, by itself, a reason to buy or not buy, and the confident declarations in either direction that circulate online generally rest on performance claims rather than on anything a specification sheet says. Our QLC versus HDD economics report works through what the QLC transition actually means at retail pricing.

The third is buying flash for cold data. It is the largest avoidable expense in this category and it is entirely a habit rather than a requirement. If the file is not in your working set, it does not need to be on a flash drive, and the cost per terabyte gap between the two tiers is roughly an order of magnitude at the prices we are tracking today.

The fourth is deferring on the assumption that prices will improve. With QLC production booked through 2026 and SanDisk raising NAND prices approximately 50% (Tom's Hardware citing DigiTimes, November 2025), the sourced picture does not support waiting. That is not urgency for its own sake — it is why we publish live numbers and invite you to check them rather than trust a static recommendation.

Frequently asked questions