Enterprise SSD rose 80% in a single quarter, 30TB disk rose from 3,000 to $17,500: AI took all storage prices flying

If the memory price increase makes PC build users feel pain, then the price increase of enterprise-level SSD is to make data center procurement directly dumbfounded.A 30TB enterprise TLC SSD, quoted at $3062 in Q2 2025, rising to $11,000 - 17,500 in Q1 2026- The calibre of different institutions is 257%(Blocks&Files) and 472%(Forbes) respectively. The difference comes from different quote windows, but the direction is the same: 2.5 to 4.7 times a year.

Increase data summary

"Quote is only valid for one week" This sentence can best explain the market status. Normal enterprise procurement process (inquiry, price comparison, approval, contract signing) usually takes 2-4 Zhou, now the price has changed after the process, many enterprises are forced to change to "lock the price before going through the process".

Why: AI reasoning turns SSD into part of computing power

The demand-side logic of this round of price increases is not exactly the same as that of memory. Memory is being robbed of production capacity by HBM, SSD isAI reasoning directly eats demand:

1. KV cache is a huge new requirement.KV Cache (Key Value Cache) needs to be maintained during large model reasoning. Under long context scenarios, the KV cache of a single request can reach several GB to tens of GB. GPU memory (HBM) cannot be loaded, so it must be placed on the next layer-enterprise SSD becomes the unloading layer of KV cache. This means SSD has changed from "where data is stored" to "part of the computing link", and the nature of requirements has completely changed.
2. The high-volume QLC is the protagonist.KV cache layering requires large-capacity, high-order read and relatively low-cost disks, which is exactly the positioning of QLC eSSD (61TB, 122TB). Armor Xia is verifying 245TB QLC enterprise disks, Samsung is pushing 236 layers NAND, SK Hynix/Solidigm is expanding 176 layers TLC and accelerating 240 layers--the original product focus is fully shifted to high-capacity enterprise level.
3. AI's share of NAND demand is rising rapidly.The industry estimates that the proportion of AI in NAND demand will rise from 18% in 2025 to 41% in 2027, and AI-related NAND demand will increase by 60% year-on-year.
4. The supply side actively contracts.Samsung, SK Hynix, Kaixia and Micron will collectively reduce NAND production in the second half of 2025. Instead of expanding production in 2026, they will give priority to existing production capacity to high-capacity enterprises with the highest profits. Cloud manufacturers will lock production capacity with 3-5 year long-term agreements.Spot inventories fell to a historical low of 3-5 weeks(The normal level is 8-12 Zhou).
5. New capacity has to wait.Meaningful new NAND capacity will not be released until the end of 2027 or 2028, which is a multi-year constraint, not a quarterly fluctuation.

A new trend: SSDs begin to "impersonate" memory

Because DRAM has also gone crazy (Gartner caliber: memory prices rose 50%-200% in the first half of 2026, server costs rose more than 125%), system designers began to find ways to replace some DRAM with high-capacity SSDs.

The most representative action isAMD acquires MEXT--The company's technology makes NAND flash look "DRAM like" to the operating system: preloading them into main memory in advance by predicting which data pages will be accessed next. The goal is to reduce the dependence of large-scale AI deployments on expensive DRAM.

This direction deserves enterprise IT attention, but don't expect to land in the short term: TrendForce clearly points out that integrating a startup's memory tiered platform into a production data center requires a lot of performance validation work, and "integration and validation are the next major obstacles." Optimistic estimates are 2027-2028 will appear on commercial servers.

A more realistic landing would beTiering Storage: Tier data by frequency of access, hot data on NVMe SSD, warm data on SATA SSD or high-capacity QLC, cold data on HDD or tape/cloud archive. This approach does not rely on new technology, can be done now, and can significantly reduce the total cost of storage in the price-increasing environment.

Four practical suggestions for enterprise storage procurement

1. Buy now what you should buy, but only what you have a clear need for.

If there is budget and key projects (database expansion, virtualized storage, AI reasoning cluster), it is better to place an order now than to wait_Q3 forecast will increase by 10%~40%(TrendForce said 10-15%, Weigang Chairman said that the supplier has notified 35-40%, take the high value to make risk plan).But don't hoard goods without clear demand: Storage planning should follow business, not market sentiment. Store a batch of SSD for two years, flash particles long-term power failure has the risk of charge leakage, but may be bad in hand.

2. The quote process is changed to "lock the price before approval"

In the market environment where the quote is valid for one week, the traditional process of "three price comparison → approval → order again" will fail repeatedly. It is suggested to change: apply for a price lock form from the supplier first (usually 7-30 days), and then go through internal approval after obtaining the price lock form. We can provide 30 days of price lock form to enterprise customers, which is the advantage of wholesale channels.

3. Re-evaluate SSD/HDD/Cloud mix

SSD costs 22.6 times more per terabyte than HDD, a multiple that expanded from 6.2 times in nine months. This means that many of the scenarios that used to be "easy on SSD" should now be re-calculated:

4. Consider used/refurbished storage (but only for non-critical scenarios)

Servnet UK and others suggest that refurbishing storage for non-performance critical workloads can bypass this NAND hike entirely because it comes from existing hardware rather than new capacity. Applicable scenarios: test environments, development sandboxes, non-critical file sharing, offline backups.Not Applicable: Production database, storage of unique data, system with compliance requirements.

There are three things you must do to buy a refurbished disk: check the SSD's write life margin (SMART Percentage Used / Media Wearout Indicator, do not buy less than 80% of the remaining life), a full read and write test (not just to see if it can be identified), and ask for at least a 1-year warranty.

The market as a whole is still growing: not "no one bought"

An easily overlooked data:IDC recorded 2026 Q1 global external OEM enterprise storage systems revenue of $9.2 billion, up 22.7% year-over-year, and all-flash array (AFA) exceeded 50% of this revenue for the first time.

That is to say, although prices are rising, companies are still buying, and buying more. This shows that the storage demand brought by AI is rigid-not "expensive and so on", but "more expensive and more expensive". Under this demand structure, it is unrealistic to expect prices to fall naturally.

ACCPC provides enterprise storage solutions and wholesale of hard drive/SSD in Guangzhou, including the full series of Armada, Samsung, Western Digital, Seagate, Titanium, Intel/Solidigm, including enterprise-class NVMe, SATA SSD, QLC large-capacity disk. Services include: tiered storage solution design, price locking procurement (Enterprise can lock for 30 days), SMART inspection and burn-in test of arrival, RAID/storage pool deployment, old data migration. Also do Guangzhou server wholesale (Dell/HP/Lenovo/ultramicro/Inspur), storage + whole machine package to discuss the price. Phone 020-39029800.