VDURA's V12 puts about 90% of AI capacity back on hard drives. Flash prices are why.

For two years the advice to anyone building an AI data center was simple. Put everything on flash.

Flash is the fast storage inside an SSD. No moving parts, very quick, and expensive per terabyte. Hard drives are the old spinning kind. Slower, but far cheaper for the same amount of data.

On September 30, 2026, VDURA, a company that has long sold hybrid flash-and-disk systems, released a new version of its storage software that makes the mix the headline. Keep the files you are actively using on flash. Put most of the data on hard drives. By VDURA's own account, about 90% of files stay on flash while about 90% of the total capacity sits on disk.

That sounds like a step backwards. It is not. It is what happens when flash gets dramatically more expensive.

Why storage matters to an AI cluster

A GPU is only useful when it has data to work on. If storage cannot feed it fast enough, a chip that cost tens of thousands of dollars sits idle. That is why AI data centers went all-flash in the first place. They were buying speed, and flash was cheap enough that nobody had to think hard about it.

Two things changed. Flash stopped being cheap, and power stopped being easy to get.

What VDURA shipped

The new version is called V12. It is available now for VDURA's V5000 systems and as an upgrade for existing customers. It runs on Supermicro hardware, specifically three models the company has qualified: the AS-1116CS-TN, the AS-2015HS-TNR and a 90-bay disk shelf called the CSE-947HE2C.

Two features are genuinely new. The first is multi-tenancy, which lets one storage system serve several customers who cannot see each other's data. That matters to neoclouds, the companies that rent out GPUs by the hour and need to keep one client's work separate from another's.

The second is what VDURA calls Context-Aware Tiering. Software watches which data is being used and moves it between flash and disk automatically, instead of an administrator deciding in advance.

VDURA's claims for all this: more than twice the performance per watt, more than 60% lower total cost, and one system that works from 8 GPUs up to 100,000. It says it has more than 1,000 deployments in over 50 countries.

The price shock

Here is the number that explains the design.

TrendForce projected that NAND flash contract prices would rise 70% to 75% in the second quarter of 2026 compared with the quarter before. NAND is the chip technology inside flash storage. For the first time in this cycle it was rising faster than DRAM, the memory that sits next to a processor.

Digitimes flagged the squeeze back in January. Avnet's analysts estimate AI data centers will take roughly 70% of all high-end DRAM in 2026, which is the reverse of every previous cycle.

We covered the memory side of this in DRAM prices and AI. Storage has had less attention and has ended up in the same place. Compounding TrendForce's first and second quarter forecasts, flash has roughly tripled in price inside six months. When the thing you store bytes on moves like that, putting most of those bytes somewhere cheaper stops being a compromise and starts being arithmetic.

Power is the real currency

The performance-per-watt claim deserves attention, though not for the usual reason.

A data center has a fixed amount of electricity. Every watt the storage system does not use is a watt available for a GPU. So power efficiency is not an operating cost footnote. It is a direct trade against how much compute you can run in the building.

This connects to something we have reported before. If the grid connection is the bottleneck, then the fight inside the building is about how to spend a fixed allowance. VDURA's argument is that hard drives draw less power per terabyte than flash for data that is not being read. Flash vendors dispute the comparison, pointing out that disks keep spinning.

Nobody has checked the numbers

Now the part the rest of the coverage left out.

The announcement ran almost word for word across HPCwire, BigDATAwire, StorageNewsletter, SDxCentral, Yahoo Finance and several smaller sites. None of them measured anything. The one piece of real analysis, by Chris Mellor at Blocks and Files, reports the claims without testing them.

There is a place to test them. MLPerf Storage is a shared industry benchmark, run the way a standardized exam is run, so different vendors' systems can be compared on the same tasks. Version 3.0 came out in September 2026 and added the exact workloads VDURA is selling against.

VDURA did not submit to it. Neither did DDN, Dell, VAST Data or WEKA. Nineteen organizations did, including Microsoft Azure, NVIDIA and Everpure, publishing 143 audited results, so measured numbers exist for the same workloads V12 is sold against.

Skipping a benchmark is not proof of a bad product. Submitting costs money and vendors opt out for ordinary reasons. But when rivals have sat the shared exam and a vendor has not, buyers are being asked to take its performance claims on trust.

What to watch

Three things.

• Whether anyone submits to the next round of MLPerf Storage, which would turn this argument from marketing into measurement.
• Whether flash prices ease later in the year. If they do, the case for mixing in hard drives weakens and the all-flash vendors get their story back.
• Whether neocloud operators buy multi-tenant storage as a product or build it themselves. That assumption is what VDURA, a company that was called Panasas until May 2024, is betting its repositioning on.