Memory price rises annoy anyone building a PC. Enterprise SSD price rises stop data center procurement cold. A 30TB enterprise TLC drive was quoted at US$3,062 in the second quarter of 2025 and US$10,950 to US$17,500 in the first quarter of 2026. Blocks&Files counts that as 257%, Forbes as 472%. The gap comes from different quote windows, not disagreement about direction: an increase of 257% to 472% in a year.
The damage in one table
| Metric | Figure | Source |
|---|---|---|
| Enterprise SSD contract price, overall | 2026 Q1 up 80% to 100% quarter on quarter; Q2 forecast +48% to +53% | TrendForce |
| NAND flash contract price | 2026 Q1 up 33% to 38% quarter on quarter; Q2 +70% to +75% | TrendForce |
| Enterprise SSD revenue | Q1 up 86.1% to US$18.46 billion | TrendForce |
| 30TB TLC enterprise drive, unit price | US$3,062 (2025 Q2) to US$10,950-17,500 (2026 Q1) | Vdura data, reported by Forbes and Blocks&Files |
| PCIe 4.0 mainstream enterprise drive, channel | Up 30% to 40% cumulative from January to June 2026 | China channel distributors |
| PCIe 5.0 AI-grade drive | Up 50% to 70% in the first half; the hot 3.84TB model up more than 10% in a month | Domestic channels |
| Quote validity | Some supplier quotes hold for one week only | Gartner / EE Times |
| SSD versus HDD cost per TB | A 30TB QLC SSD costs 22.6 times as much per TB as HDD, against 6.2 times nine months ago | Industry estimates |
"The quote is valid for one week" explains the market better than any percentage. A normal purchase runs two to four weeks through inquiry, comparison, approval and signature. By the time the paperwork clears, the price has moved, so buyers lock the number first and run approval afterwards.
Why: AI inference turned the SSD into part of the compute chain
This is not the DRAM story, where HBM took the wafers. AI inference is eating SSD supply.
- KV cache is the new bulk demand. Serving a large model means holding a key-value cache, which under long context can reach several to tens of gigabytes per request. HBM cannot hold it, so it spills to the next tier and the enterprise SSD becomes the offload layer. The drive stops being where data is stored and becomes part of the compute path.
- High-capacity QLC is the star. A KV cache tier wants capacity, good sequential reads and a moderate price per gigabyte, which is where QLC eSSD sits at 61TB and 122TB. Kioxia is validating a 245TB QLC enterprise drive, Samsung is pushing 236-layer NAND, and SK Hynix and Solidigm are expanding 176-layer TLC while accelerating 240 layers.
- AI's share of NAND demand is climbing fast. Industry estimates put AI at 18% of NAND demand in 2025 and 41% in 2027, with AI-related NAND demand up 60% year on year.
- Supply is contracting on purpose. Samsung, SK Hynix, Kioxia and Micron all cut NAND output in the second half of 2025, and 2026 capacity goes to the highest-margin high-capacity parts rather than expansion. Cloud providers lock output with three to five year agreements. Spot inventory has fallen to three to five weeks, a historic low against a normal eight to twelve.
- New capacity is years away. Meaningful new NAND output does not arrive until late 2027 or 2028: a multi-year constraint, not a quarterly wobble.
A new trend: SSDs start pretending to be memory
DRAM went up too. Gartner puts the rise at 50% to 200% for memory in the first half of 2026, with server costs up more than 125%. Designers started looking for ways to replace DRAM with high-capacity flash.
The clearest example is AMD's acquisition of MEXT, whose technology makes NAND look DRAM-like to the operating system: it predicts which data pages will be touched next and preloads them into memory, cutting AI deployments loose from expensive DRAM.
Worth watching, not planning around yet. TrendForce says fitting a startup's tiering platform into production takes heavy performance validation, and calls integration and validation the next major obstacle, with an optimistic 2027 to 2028 for commercial servers.
What works today is ordinary tiering: hot data on NVMe, warm on SATA SSD or QLC, cold on HDD, tape or cloud archive. No new technology required, and it cuts total cost now.
Four practical rules for storage procurement
1. Buy now if you have a real need, and only then
With budget and a live project (database growth, virtualized storage or an inference cluster), ordering now beats waiting. The third quarter adds another 10% to 40%: TrendForce says 10% to 15%, ADATA's chairman says suppliers have notified 35% to 40%, and risk planning should use the high end. Do not hoard. Leave SSDs on a shelf for two years and the cells can leak charge; you end up with dead stock.
2. Lock the price before you run approval
When a quote lives for a week, the old three-quotes-then-approve flow fails repeatedly. Ask the supplier for a price-lock letter, good for 7 to 30 days, and run approval against that. We issue 30-day locks to enterprise customers, which is one advantage of a wholesale channel.
3. Recalculate the SSD, HDD and cloud mix
SSD now costs 22.6 times as much per terabyte as HDD. Workloads that used to default to flash deserve a fresh look:
| Data type | Access pattern | Recommended media for 2026 |
|---|---|---|
| Database and virtualization system disks | High random IOPS, latency sensitive | NVMe SSD |
| AI inference KV cache | Large sequential reads | High-capacity QLC SSD |
| File sharing and office data | Medium to low frequency | SATA SSD or HDD, with HDD the better value |
| Backup and archive, 90 days or more untouched | Very low frequency, compliance retention | HDD plus cloud archive, or tape (LTO) |
| Surveillance video | Continuous writes, 30 to 90 days retention | Surveillance-grade HDD, not SSD, which costs about ten times more |
4. Consider used or refurbished storage, but keep it off the critical path
Servnet UK and others point out that refurbished storage sidesteps this NAND cycle entirely, because it comes from hardware that already exists rather than from new capacity. It suits test labs, sandboxes, non-critical file sharing and offline backup copies. Not for production databases, storage holding the only copy of your data, or anything under compliance.
Three things are mandatory. Check write-endurance headroom through SMART Percentage Used or the Media Wearout Indicator, and walk away below 80% remaining life. Run a full read and write test, not just a check that the drive enumerates. Insist on a one-year warranty. In a price spike the refurbished market gets murkier, with wiped SMART data more common.
The market is still growing, so nobody has stopped buying
IDC recorded US$9.2 billion in worldwide external OEM enterprise storage systems revenue for the first quarter of 2026, up 22.7% year on year, and all-flash arrays passed 50% of that revenue for the first time.
Prices are rising and companies are still buying more. AI storage demand is rigid: not "too expensive, so we wait" but "however expensive, it has to ship".
ACCPC supplies enterprise storage and wholesales drives and SSDs in Guangzhou, covering Kioxia, Samsung, Western Digital, Seagate, ZhiTai and Intel/Solidigm, including enterprise NVMe, SATA SSD and high-capacity QLC. We design tiered storage, issue 30-day price locks, inspect SMART and burn in every arrival, deploy RAID pools and migrate old data. We also wholesale servers from Dell, HP, Lenovo, Supermicro and Inspur, and quote storage with machines. Call 020-39029800.