Hot or Not

Years ago we ran a FAN. A File Area Network.

The concept came from a company called Acopia, later acquired by F5. The hardware was called ARX. We had the big ARX 6000 boxes, director-class switch chassis with custom boards inside, and smaller ARX 2000 systems at remote sites. The whole thing sat inline in front of NetApp controllers, doing file-level tiering in real time. 15K SAS for the hot tier. SATA for the cold. The ARX decided where every file lived based on how active it was, and moved it transparently. Applications saw nothing. The cost savings were enough to justify the complexity. It caused havoc for the snapshot growth when files were moved between tiers though.

It was also fragile. At scale, something was always broken, about to break or we were still cleaning up the problem before. The kind of architecture where you knew the vendor support team by name.

FabricPool didn't exist then, it would have solved all those issues. What did exist was knowing that not all data has the same temperature, and that storing cold data on expensive, fast media was a costly game. We built a fragile solution around that because the alternative was paying for 15K SAS underneath everything, which was its own kind of expensive.

That was then. Now the cost difference is even more dramatic.

Between Q2 2025 and Q1 2026, pricing for 30TB enterprise-grade SSDs rose 257%, from around £2,400 to nearly £8,600 (https://www.blocksandfiles.com/disk/2026/01/20/vdura-says-enterprise-ssds-now-16-times-more-expensive-than-disk-drives/4090513). Over the same period, HDD pricing rose around 35%. The cost multiple between enterprise SSD and nearline disk capacity went from 6x to over 16x. VDURA modelled the real-world impact on a 25PB deployment running at 1,000 GB/s: an all-flash architecture that cost the equivalent of £6.7m annually at Q2 2025 prices costs roughly £19.4m at Q1 2026 prices, a 189% rise driven almost entirely by flash media pricing.

That is not a small problem to overcome. The cause, AI workloads consuming the silicon wafer capacity that would otherwise produce the NAND in your next storage refresh, is not resolving before 2028 at the earliest. The SK Hynix chairman said as much at GTC recently.

So here is the question that most enterprise storage planning needs to get back to.

How hot is your data?

Because most of it is stone cold. Write once, read never. Industry estimates consistently put cold or inactive data,content that hasn't been accessed in months, backup volumes, compliance archives, file share residue, AI training datasets that have already been ingested, at somewhere between 60 and 80 percent of the typical enterprise estate. According to the Komprise 2026 State of Unstructured Data Management Report (https://www.intelligentcio.com/apac/2026/01/21/the-storage-and-memory-squeeze-of-2026-why-it-leaders-must-act-now-on-unstructured-data/), 74% of organisations now hold more than 5PB of unstructured data, with 40% holding more than 10PB. Most of that data is not being actively read. Most of that data is probably sitting on expensive flash because the costs kind of worked in the last refresh cycle and it was easier that asking the complicated questions.

The fix is not complicated. The discipline to do it is. The principle of storage tiering has existed for decades. Match the performance tier to the actual access pattern of the data. Hot data, active workloads, AI inference pipelines, live transactional systems, that belongs on fast, expensive flash. Warm data, less frequently accessed file shares, secondary databases, backup targets, that belongs on a cheaper medium where the performance trade-off is acceptable. Cold data, compliance archives, historical records, datasets that serve a regulatory obligation but are never operationally accessed that should never have been on flash at all.

The reason this discipline has been easy to skip is that flash was cheap enough to absorb the waste. It no longer is.

ONTAP's Inactive Data Reporting does the diagnostic work (https://docs.netapp.com/us-en/ontap/fabricpool/determine-data-inactive-reporting-task). It scans your volumes and tells you, with real precision, how much of what you currently hold on SSD-backed storage has not been touched recently. In many environments, that number surprises people. Significantly. It is not uncommon to find that a third or more of the capacity on an expensive all-flash aggregate is data that could move to a cheaper tier without any application-visible performance impact at all.

FabricPool acts on that information automatically. It identifies cold blocks, moves them to a lower-cost object tier, on-premises via StorageGRID on to ONTAP S3 on a hybrid FAS, or cloud-based, and retrieves them transparently when they are needed again. The application sees nothing. The SAN or NAS interface is unchanged. But the flash you will be paying the 2026 premium for is being used for data that actually needs it, not data that happened to land there and never moved.

The tiering decision and the classification decision are the same decision made at different layers. Data Classification (https://www.netapp.com/data-services/classification/) tells you what your data is and how sensitive it is. Inactive Data Reporting tells you how active it actually is. Between the two, you get a complete picture of what belongs where: what needs to be on fast flash, what should move to disk or object, and what does not need to be in your primary estate at all. That picture is not theoretical. It produces a direct line to a specific number, the capacity you are currently paying flash prices for that you do not need to.

At today's prices, getting that number right is a huge win, not a nice to have. The organisation that right-tiers its estate in 2026 is buying less new flash than the one that can't. On any meaningful programme, that difference runs to hundreds of thousands of pounds.

When did you last check if your data was hot or not?