Buying an SSD Now That Prices Have Climbed — What Sets the Price and the Lifespan

2026年8月17日2026年8月20日

6 August 2026

I was down to 211GB of free disk.

1.2TB of 1.5TB used up. Writing local AI articles for this blog means models accumulate. Download one to measure it, keep it to compare against, then a new one appears and you download that. I could delete them, but I often end up re-measuring something, so I never quite do.

Time to buy another drive, I thought, and went to look at prices. It was not what I expected. (Prices are Japanese retail as of August 2026; ¥1,000 is roughly $6.70.)

While I was at it I measured what my own machine actually does, and that was the more surprising result. Does buying a fast SSD shorten the waiting?

What is taking up all the space?

First, I counted what is in there.

Here is what a single model weighs, from the larger ones I have on hand.

Keep three of the big ones and that is 200GB. I keep a spread of sizes for comparison, which is how it piles up here.

How far have SSD prices climbed?

I went to buy another one and stopped.

Nearly ¥37,000 for 2TB, and that is the cheap end, checked on Amazon on 6 August 2026. Go to a name-brand drive and 2TB runs ¥60,000–80,000.

Looking into why led to the same place as the graphics card increases.

So “wait until it gets cheaper" is not much of a plan right now. But it turns out that buying an expensive one does not straightforwardly solve things either.

How many bits go in one cell: what changes between TLC and QLC

This is about what is inside the drive, and it ties directly to price and lifespan.

An SSD remembers data by holding charge in tiny containers (cells). How many bits you cram into one cell is what the names refer to.

Cram in more and it gets cheaper and larger, and the life gets shorter. QLC has to distinguish 16 voltage levels, which takes precise control on both writing and reading, and that burden accumulates.

The important part here is that reading and writing are completely different stories.

Checking the 2TB drives that are actually selling, nearly all of them are TLC. QLC has dropped out of the mainstream shelf. Perhaps because the price rises narrowed the gap and being cheap was its whole argument.

Both of my drives, incidentally, are QLC. I am running on the minority side these days. Loading models still behaves perfectly well.

The 2TB drives that are selling

Checked on Amazon on 6 August 2026. TBW is a guide to “how many terabytes you can write in total."

TBW came out backwards against price. The cheapest, the Ediloca, is rated 1,400TB; the ¥68,000 WD is 1,200TB. Expensive does not line up with long-lived.

There was also a case of a listing contradicting itself. The Silicon Power page says “TBW: 1200" in the description and “warranty valid within the TBW value (300TB)" in the warranty section. A factor of four apart, on the same page. Reading the warranty figure is the safer bet.

As an Amazon Associate we earn from qualifying purchases.

What actually wears a drive out: how to read a TBW figure

TBW is a total of writes. This was the most reassuring part of the whole exercise.

Reading barely wears anything. However many times you load a model, that does not come off the total.

Storing local AI models is a write-once, read-many pattern. Download a hundred 40GB models and that is 4TB of writes. Against 1,400TB, that is 0.3%.

I also looked at how much my SSD had written since boot, over 16 days.

I could find no reason to buy an expensive drive out of concern for lifespan, for this use.

There are uses that do eat the lifespan

• Using it as the spillover from memory. Small writes, endlessly
• Producing intermediate files for fine-tuning (LoRA and friends)
• Repeatedly downloading and deleting models. At 40GB each, every swap adds up
• A resident tool writing logs continuously

The last one has a real example.

A command-line AI coding tool was found to have a bug that kept writing logs at the most verbose setting. On the machine of the person who reported it, 37TB had been written in 21 days. Annualised, about 640TB. That would exhaust a 1TB-class drive’s rating (around 600TB) in under a year.

I checked my own setup: the relevant log was 48KB and the bug was not occurring. That is consistent with the 7TB a year figure.

So when people worry that local AI will wear out an SSD, the thing to suspect is not the models but whatever is running all the time. On Linux you can see your own write total with this.

cat /sys/block/nvme0n1/stat | awk '{print $7*512/1024/1024/1024, "GB"}'

That is the total since boot. Divide by how many days it has been up and you have a rough annual figure.

What I measured, and how: read speed and lane width

Two things. How many gigabytes per second the SSD in this machine can actually read, and how many PCIe lanes it is connected on.

Reads were measured past the cache (O_DIRECT). Read a file that is already sitting in memory and you measure the memory, not the drive. Four conditions:

• One file at a time
• Four at once
• Eight at once
• A 39.6GB straight read after dropping the cache

Lane width is read straight out of what Linux already exposes. /sys/class/nvme/nvme0/device/current_link_width is how many lanes it is on right now, max_link_width is how many it could use.

Test machine: desktop (GeForce RTX 3090 24GB + GeForce RTX 3060 12GB, 62GB system memory, two NVMe SSDs and two HDDs), August 2026.

Results

Reads came in at about 1GB/s, no matter how many at once

From here on, this is measured on my own machine.

The shops advertise numbers like “up to 7,400MB/s read." That means 7.4GB every second. Buy the drive with the big number and model loading ought to get faster.

So I measured what the SSD I already have actually reads at.

About 1GB per second. Reading more files in parallel changes nothing.

That is one seventh of the 7,400MB/s on the shelf. Mine is a drive from a few years ago, of course, and a new one would do better. But what sets the waiting on this machine is the 1GB/s, and fitting a 7,400MB/s drive does not guarantee you get 7,400MB/s. The reason follows.

It was connected on two lanes, on a drive rated for four

Something I was not expecting turned up.

An M.2 SSD connects to the PC over PCIe. That route has lanes, and normally there are four (x4). Looking at mine, it is on two lanes (x2). The drive supports four; half of them are going unused.

M.2 slots on a motherboard differ in lane count by position, and some of them share lanes with the SATA ports. It is in the manual, but it is not something you think about while building.

Before buying a fast SSD, it is quicker to check how many lanes the one you have is running on. That was the main thing I got out of this. Moving it to another slot costs nothing.

On Linux, these will tell you.

cat /sys/class/nvme/nvme0/device/current_link_width
cat /sys/class/nvme/nvme0/device/max_link_width

The first is how many lanes right now, the second is the most it could use. If they differ, you are leaving lanes on the table.

At what point do you run out of space?

I am using 742GB, but that is because I keep a spread of sizes for comparison. It is not normal usage.

I counted how many models fit in 1TB, assuming about 900GB is actually usable.

1TB is plenty to start with. Up to the 30B class you can hold dozens.

And the size you should be storing is decided by the memory you have. On a machine with 12GB or 24GB of VRAM, keeping an 86GB model around does not make it usable. There is no reason to store what you cannot load, so the capacity you need stays naturally in check.

You run short once you want to keep several 70B-plus models. Even then, there is an option before buying: split where things live.

I have not measured the external yet, so that figure is the specification rather than a result. Homework for next time.

Bigger drives cost less per gigabyte

The per-gigabyte figures separate clearly: ¥20.0 at 1TB, ¥18.5 at 2TB, ¥16.8 at 4TB.

Buying in bulk is certainly better value. But 1TB is ¥20,000 and 4TB is ¥67,000, so whether you can pay three times as much today is another question. Adding capacity later costs you the swap, so take this as no more than “go bigger if the budget allows."

What I could not put a number on

The write cliff is not measured here. That is the slowdown after the SLC cache is exhausted. Measuring it properly takes close to 100GB of writes, which means spending lifespan to measure lifespan. Those figures come from the manufacturers’ own explanations.

Prices, ratings and TBW figures were only looked up. Checked on Amazon product pages on 6 August 2026, on the Japanese store; availability differs by country, so the links point to equivalent products. Prices are still climbing and may well have moved by the time you read this.

NAND trading prices and the manufacturers’ increases are quoted from reporting, with sources given. They are not my measurements.

What a newer drive would actually do in this machine is still unknown. There is one drive here, several years old, and nothing to compare it against.

In summary: look at what is already in the machine before you buy

Going in, I assumed a faster drive would mean faster loading. Measuring said otherwise.

• Before buying, check how many lanes the drive you have is on. Mine was on half. Moving it costs nothing
• TLC is enough. In practice almost every 2TB drive that sells is TLC, so there is not much to decide
• TBW does not track price. The cheap one can have the bigger number, and a single page can give two different figures depending on where you read
• Reading does not consume the lifespan. Used as a model store, this will not become a problem
• Capacity is decided by your memory. There is no point storing what will not load
• 1TB gets you started. When it fills, external drives and HDDs are somewhere to move things
• Bigger is cheaper per gigabyte. Whether to pay three times as much today is a separate call

What I am going to do is move the drive to another slot first and see whether the lane count goes up. If that is not enough I will buy, but the order should start with the option that costs nothing.

With prices climbing, not buying at all is the best outcome available.

Products covered here

If you just want one drive.

2TB as a model store, where the TBW figure is actually stated.

4TB, the cheapest per gigabyte.

External drives, for when there is no internal slot free.