Why on-prem storage is having a moment again: the cloud-cost backlash

On-prem storage is cool again, and the reason is not nostalgia. It is the bill. After fifteen years of "just put it in the cloud," creative teams in 2026 are doing the arithmetic and discovering that for a media library that mostly sits still and occasionally streams, renting space on someone else's S3 costs more than owning a box. The backlash has a name now, cloud repatriation, and it has numbers behind it. This is a field guide to why the pendulum is swinging back toward the rack in the closet, and where the swing is real versus where the cloud still wins.

The backlash is real, and it has a balance sheet #

The intellectual seed was planted in 2021, when Andreessen Horowitz published "The Cost of Cloud, a Trillion Dollar Paradox." Sarah Wang and Martin Casado argued that for the 50 largest public software companies, cloud spend ran roughly half of cost of revenue, and that repatriating workloads could recover billions in profit. It was a heretical thing to say in 2021. In 2026 it reads like a forecast that came true.

The current data is blunt. The FinOps Foundation's State of FinOps 2026 report, drawn from more than 1,200 organizations representing over $83 billion in annual cloud spend, found that 72% of companies exceeded their cloud budgets in the last fiscal year, and only 2% of CIOs reported spending less than they projected (FinOps Foundation, checked Jun 2026). Industry estimates put wasted cloud spend at roughly 29% in 2026, ticking up after years of slow improvement. When the cheap, elastic thing turns out to be neither cheap nor predictable, smart teams stop renting the parts that do not need to be rented.

Egress is the tax nobody priced in #

The cloud's dirty secret for media is egress: the fee to pull your own data back out. Storing it is cheap. Touching it is not. As of June 2026, Amazon S3 charges $0.09 per GB for the first 10 TB of internet egress each month, after a 100 GB free tier (aws.amazon.com/s3/pricing, checked Jun 2026). For video, where a single project can mean pulling tens of terabytes across a finishing pass, that is a meter running on every read.

Think of it like a self-storage unit where they charge you nothing to drive boxes in, a small monthly rent to keep them, and then a fee every single time you open the door to take something out. For documents you rarely reopen, fine. For a 4K timeline you scrub a thousand times a day, the door fee is the whole expense. This is exactly why we wrote a separate piece on what leaving the cloud actually costs, and why teams comparing backends should read our breakdown of B2 vs Wasabi vs plain S3 as a media backend. The newer object stores blunt the egress problem, but they do not make a remote object store behave like a local SSD.

Per-seat fatigue: paying for chairs, not footage #

The second driver is per-seat fatigue. The SaaS suites price by the head, and a media team's footage does not scale with its headcount. As of June 2026, LucidLink's Business tier runs $32 per member per month ($27 with annual billing) with 400 GB of pooled storage per member, and overage beyond the pool at $0.08 per GB (lucidlink.com and the AWS Marketplace listing, checked Jun 2026). Frame.io's plans run $15 to $25 per seat per month with bundled terabytes (frame.io/pricing, checked Jun 2026). Both are good products. But a five-editor shop with 30 TB of media is buying five seats and then paying a storage tax on the 28 TB the seats do not cover.

We did the full arithmetic in the LucidLink 2026 cost teardown: a 5-seat, 10 TB team lands near $9,300 in year one, and the seats are under a fifth of that. The storage line, not the seat line, is the bill. When the storage you own dwarfs the seats you employ, per-seat pricing stops being a convenience and starts being a leak.

The 2026 cost climate has a twist: drives are not cheap either #

Here is the honest complication, and I am not going to skip it just because it weakens the on-prem pitch. 2026 is a brutal year to buy hard drives. The same AI buildout that fills data centers has bought out the spinning-disk supply. Western Digital CEO Irving Tan said the company's nearline capacity is "fully allocated through calendar year 2026," with firm purchase orders from its top seven customers for the whole year; Seagate has said the same (per coverage of both companies' earnings, checked Jun 2026). Consumer drives took the hit: a 4 TB WD Blue that ran roughly $67 to $85 in mid-2025 was around $99 by early 2026, part of a price surge near 50% over five months.

So owning storage in 2026 means buying into a spiky market. That is real, and it changes the math at the margins. But it cuts both ways: the same demand wave is what makes cloud capacity expensive and is part of why the hyperscalers have no incentive to drop egress prices. A drive you buy once, even at an inflated 2026 price, still stops costing you money the day it is paid off. A cloud bill never does. If you are timing a build, our guides on keeping storage costs down as drive prices rise and HDD vs SSD vs hybrid for media libraries walk through how to buy smart in this climate rather than panic-buy.

The proof point: 37signals walked the walk #

The most-cited repatriation case is not a storage vendor's marketing. It is 37signals, the company behind Basecamp, which left AWS publicly and showed its receipts. They cut their cloud spend from about $3.2 million a year to well under $1 million, a saving close to $2 million in a year, after spending about $700,000 on Dell hardware (The Register and DataCenterDynamics, reporting through 2024-2025, checked Jun 2026). For the storage half, they moved to a dual-datacenter Pure Storage setup with a combined 18 PB of capacity, which they said would cost under $200,000 a year to operate against the millions S3 had been charging.

One company is an anecdote, not a law. But the broader signal is consistent: in 2026 surveys, a strong majority of enterprise IT leaders said they planned to move at least one workload off public cloud to on-prem or private infrastructure, and repatriation case studies routinely report 30% to 60% cost reductions on steady-state workloads (industry repatriation coverage, checked Jun 2026). The common thread is the word "steady-state." The cloud is brilliant for spiky, unpredictable load. A media library that grows slowly and gets read constantly is the opposite of spiky, and it is exactly the workload that comes home.

Where on-prem still loses, honestly #

Repatriation is not a religion, and I would be lying if I pretended a NAS is the answer for everyone. On-prem loses when you are genuinely distributed across continents with no central office, when you have no one willing to own a box and a backup routine, or when your workload truly is spiky and short-lived. It loses when an outage means a deadline missed and there is no IT person to drive in at 2 a.m. The cloud sells you out of those problems, and for some teams that insurance is worth every dollar of egress. If your team lives on a VPN and the bottleneck is the link, not the storage model, our piece on why a NAS feels slow over a VPN matters more than any cost table.

The honest framing is not "cloud bad, on-prem good." It is that the default flipped. For a decade the safe assumption was cloud-first. In 2026, for a creative team with a stable, heavy media library, the safe assumption is to price the owned box first and reach for the cloud only where it earns its premium. The tools have caught up, too: a self-hosted mount can now give editors a real Finder volume with block-level streaming and a local SSD cache, so "on-prem" no longer means "wait for the whole file to copy." That is the lane JuiceMount was built for, and it is also where it does not fit: if you have no NAS and no appetite to run one, a managed cloud service is the better answer, full stop.