Retailer one petabyte on Azure’s scorching tier for 5 years and the invoice runs near $1.2 million, earlier than you’ve moved a single byte again out. Attempt to retrieve that knowledge and the cloud prices once more: AWS payments roughly $90,000 in egress charges to tug a single petabyte off its platform, and Google Cloud Storage prices much more per gigabyte on the best way out than AWS does. At ten, fifty, or 5 hundred petabytes, that arithmetic stops being a line merchandise and begins being a strategic determination.
This information covers the true choices for storing petabyte-scale knowledge exterior a public cloud supplier: what they price, what they require, and which corporations have already confirmed the mannequin works.
Why the Cloud Will get Costly at Scale, Particularly
Cloud storage pricing is constructed round a easy asymmetry: getting knowledge in is reasonable or free, and getting knowledge out is just not. AWS prices $0.09 per gigabyte for the primary 10 terabytes of egress, Azure prices round $0.087/GB, and Google Cloud Storage prices roughly $0.12/GB. Cloud suppliers typically cost 4 to 6 occasions extra to retrieve knowledge than to retailer it, which suggests the sticker value on a pricing calculator hardly ever displays what you’ll truly pay as soon as your knowledge begins transferring, whether or not that’s backups, analytics pipelines, catastrophe restoration, or a full migration.
This isn’t a fringe concern. In accordance with Flexera’s State of the Cloud Report, the share of enterprise workloads that organizations have pulled again out of the general public cloud rose from 21% within the 2025 report back to roughly 23% within the 2026 version. That’s a gradual, incremental shift somewhat than a stampede, and cloud adoption total remains to be rising sooner than repatriation is shrinking it. Nevertheless it confirms the sample: corporations aren’t abandoning the cloud, they’re rebalancing it, conserving unpredictable workloads there whereas transferring regular, predictable storage onto infrastructure they management instantly.
Choice One: Run Your Personal Object Storage
Essentially the most direct substitute for S3 or Azure Blob Storage is software program that speaks the identical API however runs on {hardware} you management. Cloudian HyperStore, Scality RING, MinIO, and Caringo Swarm all ship S3-compatible object storage on odd x86 servers, which suggests purposes constructed across the S3 API don’t should be rewritten. Scality RING scales to a whole bunch of petabytes and billions of objects throughout commodity {hardware}, and MinIO ships as a binary below 100MB that runs wherever from an edge server to a full knowledge heart. TrueNAS takes a unique route, layering file, block, and object storage on ZFS for multi-petabyte deployments that want robust knowledge integrity ensures.
The attraction right here is simple: you pay as soon as for the {hardware} and the software program license (or nothing, in MinIO’s open-source case), and there’s no per-gigabyte toll on the best way out. Sturdiness comes from erasure coding somewhat than the triple replication cloud suppliers use internally, which retains the bodily footprint and the facility invoice decrease than you’d anticipate for petabyte-scale clusters.
Choice Two: Colocation As a substitute of Building
Proudly owning the {hardware} doesn’t imply constructing an information heart. Colocation enables you to set up your personal servers and storage arrays inside a third-party facility that handles energy, cooling, bodily safety, and community connectivity. Your knowledge stays on tools solely your group controls, which issues for regulated industries the place auditors wish to see precisely who can bodily contact a drive.
For steady-state, predictable workloads, on-premises and colocation ship comparable complete prices, and colocation tends to win on flexibility because you’re not committing capital to a constructing. Cloud, by comparability, runs roughly 60% extra over 5 years for equal steady-state capability as soon as storage, compute, and knowledge motion are all counted. Colocation is the sensible center floor for corporations that wish to escape cloud storage economics with out taking up services administration.
Choice Three: Tape for the Knowledge You Not often Contact
Most petabyte-scale knowledge isn’t accessed every day. It’s compliance archives, media libraries, historic logs, and backups that have to exist someplace sturdy and low cost, not someplace quick. That is the place LTO tape nonetheless wins by a large margin. LTO-9 cartridges maintain 18TB natively and value round $90 every, which works out to roughly $5 per terabyte in uncooked media price. Storing 100TB on Amazon S3 Customary for 5 years runs effectively over $130,000; the identical 100TB on LTO-9 tape prices a number of hundred {dollars} in cartridges when you personal the drive.
Tape isn’t a substitute for energetic storage, retrieval occasions are measured in minutes, not milliseconds, and it requires drives, libraries, and somebody who is aware of easy methods to handle a tape rotation. However for the chilly portion of a petabyte-scale archive, which is often nearly all of it, tape stays the most cost effective storage medium that exists.
The Corporations That Already Did This
Dropbox proved the mannequin at real scale. Between 2015 and 2016, the corporate moved 500 petabytes of buyer knowledge off AWS and onto its personal infrastructure, a undertaking it known as Magic Pocket, custom-built storage packing containers and all. In accordance with figures in Dropbox’s 2018 IPO submitting, the migration saved the corporate $74.6 million over two years: a $92.5 million drop in third-party knowledge heart spending, offset by a $53 million enhance in its personal infrastructure prices, with an extra $35.1 million saved in 2017. Dropbox nonetheless retains a small fraction of workloads on AWS the place it is smart, which is the sample price noting: this was a rebalancing, not an ideological exit.
37signals provides a smaller, newer case research. After DHH introduced the corporate’s cloud exit in October 2022, 37signals spent roughly $700,000 on Dell {hardware}, recouped that funding throughout the 12 months, and introduced its annual cloud invoice down from a $3.2 million run price to $1.3 million by 2024. The corporate now initiatives complete five-year financial savings north of $10 million, achieved with out including infrastructure workers.
Neither firm handled this as a one-time migration. Each constructed the operational capability to run storage as a everlasting perform, which is the true dedication this determination requires.
Constructing the Enterprise Case
Earlier than committing capital, run the numbers the best way these corporations did. Audit your entry patterns first: separate scorching knowledge (accessed always), heat knowledge (accessed often), and chilly knowledge (accessed hardly ever), because the proper storage tier is totally different for every. Calculate a real five-year complete price of possession to your present cloud footprint, together with egress and retrieval prices, not simply the month-to-month storage line. Select an S3-compatible software program layer should you go the on-premises or colocation route, so current purposes don’t want architectural surgical procedure. And should you function in a regulated trade, consider knowledge sovereignty necessities early: the query isn’t solely the place the info bodily sits, however which jurisdiction’s legal guidelines govern it and who could be legally compelled handy it over. Enforcement is actual cash at this level: impartial trackers put cumulative GDPR fines at greater than €7.1 billion because the regulation took impact in Might 2018, with Eire’s Knowledge Safety Fee alone accounting for roughly €4.04 billion of that complete.
None of this argues that the cloud is completed. Most corporations nonetheless belong there for compute elasticity and workloads that spike unpredictably. However at petabyte scale, storage stops behaving like a rented service and begins behaving like infrastructure, and infrastructure is nearly all the time cheaper to personal outright than to lease indefinitely.
That very same intuition, questioning what you’ve been instructed to simply accept by default, is beginning to present up past infrastructure too. As AI assistants like ChatGPT and Perplexity more and more reply questions that used to ship visitors to a search outcomes web page, businesses resembling Liplyn now assist manufacturers earn a spot inside these AI-generated solutions somewhat than leaving that visibility to probability.
Dropbox and 37signals didn’t depart the cloud as a result of it was trendy. They left as a result of somebody lastly ran the five-year math.
Datafloq is a part of Liplyn Info Group. For extra on AI visibility and digital advertising technique, go to liplyn.com.
Retailer one petabyte on Azure’s scorching tier for 5 years and the invoice runs near $1.2 million, earlier than you’ve moved a single byte again out. Attempt to retrieve that knowledge and the cloud prices once more: AWS payments roughly $90,000 in egress charges to tug a single petabyte off its platform, and Google Cloud Storage prices much more per gigabyte on the best way out than AWS does. At ten, fifty, or 5 hundred petabytes, that arithmetic stops being a line merchandise and begins being a strategic determination.
This information covers the true choices for storing petabyte-scale knowledge exterior a public cloud supplier: what they price, what they require, and which corporations have already confirmed the mannequin works.
Why the Cloud Will get Costly at Scale, Particularly
Cloud storage pricing is constructed round a easy asymmetry: getting knowledge in is reasonable or free, and getting knowledge out is just not. AWS prices $0.09 per gigabyte for the primary 10 terabytes of egress, Azure prices round $0.087/GB, and Google Cloud Storage prices roughly $0.12/GB. Cloud suppliers typically cost 4 to 6 occasions extra to retrieve knowledge than to retailer it, which suggests the sticker value on a pricing calculator hardly ever displays what you’ll truly pay as soon as your knowledge begins transferring, whether or not that’s backups, analytics pipelines, catastrophe restoration, or a full migration.
This isn’t a fringe concern. In accordance with Flexera’s State of the Cloud Report, the share of enterprise workloads that organizations have pulled again out of the general public cloud rose from 21% within the 2025 report back to roughly 23% within the 2026 version. That’s a gradual, incremental shift somewhat than a stampede, and cloud adoption total remains to be rising sooner than repatriation is shrinking it. Nevertheless it confirms the sample: corporations aren’t abandoning the cloud, they’re rebalancing it, conserving unpredictable workloads there whereas transferring regular, predictable storage onto infrastructure they management instantly.
Choice One: Run Your Personal Object Storage
Essentially the most direct substitute for S3 or Azure Blob Storage is software program that speaks the identical API however runs on {hardware} you management. Cloudian HyperStore, Scality RING, MinIO, and Caringo Swarm all ship S3-compatible object storage on odd x86 servers, which suggests purposes constructed across the S3 API don’t should be rewritten. Scality RING scales to a whole bunch of petabytes and billions of objects throughout commodity {hardware}, and MinIO ships as a binary below 100MB that runs wherever from an edge server to a full knowledge heart. TrueNAS takes a unique route, layering file, block, and object storage on ZFS for multi-petabyte deployments that want robust knowledge integrity ensures.
The attraction right here is simple: you pay as soon as for the {hardware} and the software program license (or nothing, in MinIO’s open-source case), and there’s no per-gigabyte toll on the best way out. Sturdiness comes from erasure coding somewhat than the triple replication cloud suppliers use internally, which retains the bodily footprint and the facility invoice decrease than you’d anticipate for petabyte-scale clusters.
Choice Two: Colocation As a substitute of Building
Proudly owning the {hardware} doesn’t imply constructing an information heart. Colocation enables you to set up your personal servers and storage arrays inside a third-party facility that handles energy, cooling, bodily safety, and community connectivity. Your knowledge stays on tools solely your group controls, which issues for regulated industries the place auditors wish to see precisely who can bodily contact a drive.
For steady-state, predictable workloads, on-premises and colocation ship comparable complete prices, and colocation tends to win on flexibility because you’re not committing capital to a constructing. Cloud, by comparability, runs roughly 60% extra over 5 years for equal steady-state capability as soon as storage, compute, and knowledge motion are all counted. Colocation is the sensible center floor for corporations that wish to escape cloud storage economics with out taking up services administration.
Choice Three: Tape for the Knowledge You Not often Contact
Most petabyte-scale knowledge isn’t accessed every day. It’s compliance archives, media libraries, historic logs, and backups that have to exist someplace sturdy and low cost, not someplace quick. That is the place LTO tape nonetheless wins by a large margin. LTO-9 cartridges maintain 18TB natively and value round $90 every, which works out to roughly $5 per terabyte in uncooked media price. Storing 100TB on Amazon S3 Customary for 5 years runs effectively over $130,000; the identical 100TB on LTO-9 tape prices a number of hundred {dollars} in cartridges when you personal the drive.
Tape isn’t a substitute for energetic storage, retrieval occasions are measured in minutes, not milliseconds, and it requires drives, libraries, and somebody who is aware of easy methods to handle a tape rotation. However for the chilly portion of a petabyte-scale archive, which is often nearly all of it, tape stays the most cost effective storage medium that exists.
The Corporations That Already Did This
Dropbox proved the mannequin at real scale. Between 2015 and 2016, the corporate moved 500 petabytes of buyer knowledge off AWS and onto its personal infrastructure, a undertaking it known as Magic Pocket, custom-built storage packing containers and all. In accordance with figures in Dropbox’s 2018 IPO submitting, the migration saved the corporate $74.6 million over two years: a $92.5 million drop in third-party knowledge heart spending, offset by a $53 million enhance in its personal infrastructure prices, with an extra $35.1 million saved in 2017. Dropbox nonetheless retains a small fraction of workloads on AWS the place it is smart, which is the sample price noting: this was a rebalancing, not an ideological exit.
37signals provides a smaller, newer case research. After DHH introduced the corporate’s cloud exit in October 2022, 37signals spent roughly $700,000 on Dell {hardware}, recouped that funding throughout the 12 months, and introduced its annual cloud invoice down from a $3.2 million run price to $1.3 million by 2024. The corporate now initiatives complete five-year financial savings north of $10 million, achieved with out including infrastructure workers.
Neither firm handled this as a one-time migration. Each constructed the operational capability to run storage as a everlasting perform, which is the true dedication this determination requires.
Constructing the Enterprise Case
Earlier than committing capital, run the numbers the best way these corporations did. Audit your entry patterns first: separate scorching knowledge (accessed always), heat knowledge (accessed often), and chilly knowledge (accessed hardly ever), because the proper storage tier is totally different for every. Calculate a real five-year complete price of possession to your present cloud footprint, together with egress and retrieval prices, not simply the month-to-month storage line. Select an S3-compatible software program layer should you go the on-premises or colocation route, so current purposes don’t want architectural surgical procedure. And should you function in a regulated trade, consider knowledge sovereignty necessities early: the query isn’t solely the place the info bodily sits, however which jurisdiction’s legal guidelines govern it and who could be legally compelled handy it over. Enforcement is actual cash at this level: impartial trackers put cumulative GDPR fines at greater than €7.1 billion because the regulation took impact in Might 2018, with Eire’s Knowledge Safety Fee alone accounting for roughly €4.04 billion of that complete.
None of this argues that the cloud is completed. Most corporations nonetheless belong there for compute elasticity and workloads that spike unpredictably. However at petabyte scale, storage stops behaving like a rented service and begins behaving like infrastructure, and infrastructure is nearly all the time cheaper to personal outright than to lease indefinitely.
That very same intuition, questioning what you’ve been instructed to simply accept by default, is beginning to present up past infrastructure too. As AI assistants like ChatGPT and Perplexity more and more reply questions that used to ship visitors to a search outcomes web page, businesses resembling Liplyn now assist manufacturers earn a spot inside these AI-generated solutions somewhat than leaving that visibility to probability.
Dropbox and 37signals didn’t depart the cloud as a result of it was trendy. They left as a result of somebody lastly ran the five-year math.
Datafloq is a part of Liplyn Info Group. For extra on AI visibility and digital advertising technique, go to liplyn.com.
