Store one petabyte on Azure’s hot tier for five years and the bill runs close to $1.2 million, before you’ve moved a single byte back out. Try to retrieve that data and the cloud charges again: AWS bills roughly $90,000 in egress fees to pull a single petabyte off its platform, and Google Cloud Storage charges even more per gigabyte on the way out than AWS does. At ten, fifty, or five hundred petabytes, that arithmetic stops being a line item and starts being a strategic decision.
This guide covers the real options for storing petabyte-scale data outside a public cloud provider: what they cost, what they require, and which companies have already proven the model works.
Why the Cloud Gets Expensive at Scale, Specifically
Cloud storage pricing is built around a simple asymmetry: getting data in is cheap or free, and getting data out is not. AWS charges $0.09 per gigabyte for the first 10 terabytes of egress, Azure charges around $0.087/GB, and Google Cloud Storage charges roughly $0.12/GB. Cloud providers generally charge four to six times more to retrieve data than to store it, which means the sticker price on a pricing calculator rarely reflects what you’ll actually pay once your data starts moving, whether that’s backups, analytics pipelines, disaster recovery, or a full migration.
This isn’t a fringe concern. According to Flexera’s State of the Cloud Report, the share of enterprise workloads that organizations have pulled back out of the public cloud rose from 21% in the 2025 report to roughly 23% in the 2026 edition. That’s a steady, incremental shift rather than a stampede, and cloud adoption overall is still growing faster than repatriation is shrinking it. But it confirms the pattern: companies aren’t abandoning the cloud, they’re rebalancing it, keeping unpredictable workloads there while moving steady, predictable storage onto infrastructure they control directly.
Option One: Run Your Own Object Storage
The most direct replacement for S3 or Azure Blob Storage is software that speaks the same API but runs on hardware you control. Cloudian HyperStore, Scality RING, MinIO, and Caringo Swarm all deliver S3-compatible object storage on ordinary x86 servers, which means applications built around the S3 API don’t need to be rewritten. Scality RING scales to hundreds of petabytes and billions of objects across commodity hardware, and MinIO ships as a binary under 100MB that runs anywhere from an edge server to a full data center. TrueNAS takes a different route, layering file, block, and object storage on ZFS for multi-petabyte deployments that need strong data integrity guarantees.
The appeal here is straightforward: you pay once for the hardware and the software license (or nothing, in MinIO’s open-source case), and there’s no per-gigabyte toll on the way out. Durability comes from erasure coding rather than the triple replication cloud providers use internally, which keeps the physical footprint and the power bill lower than you’d expect for petabyte-scale clusters.
Option Two: Colocation Instead of Construction
Owning the hardware doesn’t mean building a data center. Colocation lets you install your own servers and storage arrays inside a third-party facility that handles power, cooling, physical security, and network connectivity. Your data stays on equipment only your organization controls, which matters for regulated industries where auditors want to see exactly who can physically touch a drive.
For steady-state, predictable workloads, on-premises and colocation deliver comparable total costs, and colocation tends to win on flexibility since you’re not committing capital to a building. Cloud, by comparison, runs roughly 60% more over five years for equivalent steady-state capacity once storage, compute, and data movement are all counted. Colocation is the practical middle ground for companies that want to escape cloud storage economics without taking on facilities management.
Option Three: Tape for the Data You Rarely Touch
Most petabyte-scale data isn’t accessed daily. It’s compliance archives, media libraries, historical logs, and backups that need to exist somewhere durable and cheap, not somewhere fast. This is where LTO tape still wins by a wide margin. LTO-9 cartridges hold 18TB natively and cost around $90 each, which works out to roughly $5 per terabyte in raw media cost. Storing 100TB on Amazon S3 Standard for five years runs well over $130,000; the same 100TB on LTO-9 tape costs a few hundred dollars in cartridges once you own the drive.
Tape isn’t a replacement for active storage, retrieval times are measured in minutes, not milliseconds, and it requires drives, libraries, and someone who knows how to manage a tape rotation. But for the cold portion of a petabyte-scale archive, which is usually the majority of it, tape remains the cheapest storage medium that exists.
The Companies That Already Did This
Dropbox proved the model at genuine scale. Between 2015 and 2016, the company moved 500 petabytes of customer data off AWS and onto its own infrastructure, a project it called Magic Pocket, custom-built storage boxes and all. According to figures in Dropbox’s 2018 IPO filing, the migration saved the company $74.6 million over two years: a $92.5 million drop in third-party data center spending, offset by a $53 million increase in its own infrastructure costs, with an additional $35.1 million saved in 2017. Dropbox still keeps a small fraction of workloads on AWS where it makes sense, which is the pattern worth noting: this was a rebalancing, not an ideological exit.
37signals offers a smaller, more recent case study. After DHH announced the company’s cloud exit in October 2022, 37signals spent roughly $700,000 on Dell hardware, recouped that investment within the year, and brought its annual cloud bill down from a $3.2 million run rate to $1.3 million by 2024. The company now projects total five-year savings north of $10 million, achieved without adding infrastructure staff.
Neither company treated this as a one-time migration. Both built the operational capacity to run storage as a permanent function, which is the real commitment this decision requires.
Building the Business Case
Before committing capital, run the numbers the way these companies did. Audit your access patterns first: separate hot data (accessed constantly), warm data (accessed occasionally), and cold data (accessed rarely), since the right storage tier is different for each. Calculate a genuine five-year total cost of ownership for your current cloud footprint, including egress and retrieval charges, not just the monthly storage line. Choose an S3-compatible software layer if you go the on-premises or colocation route, so existing applications don’t need architectural surgery. And if you operate in a regulated industry, factor in data sovereignty requirements early: the question isn’t only where the data physically sits, but which jurisdiction’s laws govern it and who can be legally compelled to hand it over. Enforcement is real money at this point: independent trackers put cumulative GDPR fines at more than €7.1 billion since the regulation took effect in May 2018, with Ireland’s Data Protection Commission alone accounting for roughly €4.04 billion of that total.
None of this argues that the cloud is finished. Most companies still belong there for compute elasticity and workloads that spike unpredictably. But at petabyte scale, storage stops behaving like a rented service and starts behaving like infrastructure, and infrastructure is almost always cheaper to own outright than to lease indefinitely.
That same instinct, questioning what you’ve been told to accept by default, is starting to show up beyond infrastructure too. As AI assistants like ChatGPT and Perplexity increasingly answer questions that used to send traffic to a search results page, agencies such as Liplyn now help brands earn a place inside those AI-generated answers rather than leaving that visibility to chance.
Dropbox and 37signals didn’t leave the cloud because it was fashionable. They left because someone finally ran the five-year math.
Datafloq is part of Liplyn Information Group. For more on AI visibility and digital marketing strategy, visit liplyn.com.
