When Storage Talks

I am Storage and I have no shame in admitting that for last three decades I was treated as the electricity bill of the infrastructure stack, which is necessary but not strategic. Recently, this has changed sharply and today I want to walk you through what happened and why it matters.

For starters, I take a lot of different shapes across the infrastructure of every organization on the planet. I am the block storage sitting inside your SAN, the file storage humming underneath your NAS, and the object storage holding every training dataset and backup archive across every public cloud region you use. I am the NVMe flash feeding your AI pipeline at ridiculous speeds during training runs, and the nearline HDD holding the petabytes that AI training corpuses now demand. I am the tape cartridge sitting in an air-gapped vault waiting for the day nothing else can save you from ransomware. And I am now showing up in a small handful of research labs as strands of synthetic DNA. I hold what you give me for as long as you tell me to hold it, and I do not lose any of it in between.

For last three decades, I got cheaper every year on a schedule the entire industry planned its capacity budgets around, and the famous line that everyone repeated about me was that “storage is cheap”. Just add more storage. That line became infrastructure common sense, right up until the year it stopped being true.

My life splits cleanly now, the way Data’s and Database’s and Memory’s did before me, into Before AI and After AI eras. In BAI era, I was just the line item no one cares much and now in AAI era, I am one of the deciding factors for success along with memory.

In late 2025, the pricing curve took a U-turn. NAND contract prices moved up 55 to 60 percent quarter over quarter according to TrendForce. Enterprise NVMe now runs roughly 16 times the cost per terabyte of nearline HDD. HDD is on allocation for the first time in the history, with hyperscalers hoarding Seagate’s new 44-terabyte HAMR drives to feed AI training corpuses that need every byte they can find.

Here is the specific number that changed everything for finance teams, because it is the number that made the theoretical repatriation argument practical.

37signals moved 18 petabytes of production storage off Amazon S3 onto their owned storage. The one-time migration cost was 1.5 million dollars in capital expenditure, the ongoing cost is under 200,000 dollars per year, and the bill they were previously paying to Amazon was 1.5 million dollars every single year. The projected five-year saving on that one decision is somewhere north of ten million dollars.

The egress economics behind that math changed in early 2024. The European Union’s Data Act came into force in January of that year, and all three major US hyperscalers followed with their own policies within weeks. Google Cloud eliminated egress fees for customers fully migrating off its platform. AWS matched in March 2024 with a 60 to 90 day exit window. Microsoft Azure followed shortly after with its own caveated version. All three waivers require you to fully exit the platform, which means partial repatriation still incurs the traditional egress costs, which currently run somewhere between 90 and 120 thousand dollars per petabyte pulled out of S3. Enough CFOs and CIOs finally saw the math close on paper.

Storage that physically sits inside a rack you own is capacity you can plan around across a five-year horizon with predictable pricing. Storage that sits in a public cloud region you do not own is capacity you are renting from someone whose also owns the power to price it. The smart move now looks like a rack you own outright, running like cloud, with the operational simplicity that comes from treating infrastructure as a product instead of a project.

To keep things interesting, Tape came back into the conversation in 2022 Fujifilm started shipping LTO-10 cartridges in January 2026 at 30 terabytes native, roughly 100 terabytes compressed. Quantum reports working with five of the world’s largest hyperscalers on active tape deployments, driven by two forces at once. The NAND and HDD supply constraints already discussed, and the need for genuine air-gapped backups against ransomware attacks that have gotten too sophisticated for anything connected to a network to reliably resist. Tape also has approximately 97 percent lower carbon per terabyte than HDD, which matters more today than it did five years ago.

Here is what to actually walk away with from this confession.

The AI project stalling inside your organization right now are stalling because of the multiple factors including model, or the GPU, or the data engineer who left last quarter, or Storage. Prices are up 55 percent quarter over quarter across the flash market, capacity is on allocation for the first time, and the training data you had assumed you could use next quarter may live in a jurisdiction that changed its rules six months ago.

Storage is not the boring line item anymore. It is the line that decides whether the whole AI project works, where it physically has to live, and how much of your five-year infrastructure cost ends up sitting on someone’s balance sheet instead of yours.

Next time someone on your team tells you storage is cheap, ask them when they last looked at the purchase order and how they feeding their AI projects.