Cloud Repatriation
Who Left the Cloud, What They Saved, and When It Pays
A sourced tracker of 122 moves by 112 companies: who left public cloud, who went back and who stayed. Then the math: egress, exit waivers, and when colocation or bare metal beats optimized cloud.
Model My Cloud CostsUpdated November 12, 2026. Full tracker available as a CSV download.
Cloud repatriation is moving workloads off public cloud (AWS, Microsoft Azure or Google Cloud) onto infrastructure you control: your own servers in a colocation facility, rented bare metal, or a hybrid where the steady core moves and the variable edge stays. It is not a rejection of cloud. It is a decision about which workloads belong where, made on total cost over three to five years rather than on this month's bill.
The public record is narrower than the surveys suggest. Our research team tracks 124 moves by 112 companies and public bodies, each with a public source for what it did. 51 moved at least some workloads off public cloud, but only 10 of them left completely and finished; most moved part of their workloads, went to another cloud, or are still moving. 29 moves went onto cloud or back to it, 12 of them AI capacity deals, and 4 companies renewed or expanded large cloud commitments. That is the work our data center consulting practice does: model the move against optimized cloud, then place it in wholesale colocation or bare metal with the operators that fit.
Send your annual spend, the main workloads and what is driving the question. You'll hear back within 24 hours, including whether the analysis is worth running for you.
Is Cloud Repatriation Real? What Surveys Say and What Companies Did
Yes, but on a much smaller scale than the surveys suggest. Surveys of IT leaders, most of them sponsored by vendors that sell private cloud or hardware, put repatriation at anywhere from a fifth of workloads to more than two-thirds of companies. IDC's own survey work found that only 8 to 9 percent of companies plan a full exit. The named, sourced moves in our tracker look like that smaller number: specific workloads, moved for cost, egress or control, with cloud kept for the rest.
Sources: Virtualization Review on Cloudera; Broadcom; The New Stack on Flexera; ITPro on IDC; TechTarget on IDC.
The companies with the most to gain from a repatriation wave describe it carefully too. Asked about it on Equinix's April 2026 earnings call, CEO Adaire Fox-Martin said some customers find "certain parts of their dataset need to move into a private environment or be repatriated from cloud," but "I would not say that this is a broad-based conversation across our customer base." (transcript)
And the big cloud contracts keep getting renewed. Pinterest disclosed a $4.0 billion AWS commitment running through May 2031, Airbnb at least $1.7 billion in hosting commitments through 2031, and Reddit $880 million with AWS and $580 million with Google Cloud. Each of these commitments was renewed or expanded, not cut. So the picture is this: repatriation is real for steady, data-heavy workloads, and it sits alongside cloud rather than replacing it.
Cloud Exit Tracker: Which Companies Left Public Cloud, Where They Went and What They Saved
Dataset, updated November 2026. 124 entries, 112 companies and public bodies. Free to use under CC BY 4.0. DOI: 10.5281/zenodo.24223306.
Most companies that leave public cloud leave part of it. Of 51 named companies and public bodies with a public source for moving workloads off a public cloud, 10 left completely and finished. 18 moved part of their workloads and kept cloud for the rest, 6 have a move planned or under way, 14 went to another, smaller cloud, and 3 later moved back or shut the setup down. 31 of the 51 put a number on cost, and 10 of those numbers come from vendor case studies.
Every row below is a named company with a public source. Of the 124 rows, 17 rest on a filing, earnings call or government record, 55 on the company's own blog, talk, paper or post, 29 on a vendor's case study or a cloud provider's own release, and 23 on press reports. The source type is a column in the data file. Rows marked "Partial source" (21 of them) are ones where the source supports only part of the row. Each exit carries a status, and a short note says what the source shows when the label needs it. The table opens on every exit, 52 entries. Switch the first filter to see the moves onto cloud, the companies that stayed, the AI capacity decisions and a watch list of 3 entries we no longer count as exits, or filter by exit status, destination, stated cost or AI only.
| Company | From cloud | To where | Year | Workload | Cost stated | Source |
|---|---|---|---|---|---|---|
| 37signals (Basecamp, HEY)Left the cloudFull exitCompletion reported by Structure Research (Jun 2025) and The Register (May 2025) | AWS (S3) | Own storage (Pure Storage) in colocation | 2025 | S3 object storage (almost 6 PB to move) | S3 spend was just under $1.5M a year; The Register (May 2025) reports about $1.3M a year in savings; moved the data in a free egress window | 37signals blog (DHH) |
| InstagramLeft the cloudFull exit | AWS | Facebook's own data centers | 2013 to 2014 | Entire production stack and photo store | No figure in the talk; a Facebook engineer said the move was reducing costs vs Amazon (via The Register) | SREcon15 talk (USENIX) |
| WhatsAppLeft the cloudFull exit | IBM SoftLayer (per CNBC) | Facebook's own data centers | 2017 to 2019 | Backend and user data (1.5B users) | No figure; three years of work to migrate data into Facebook's data centers | Code BEAM SF 2019 talk |
| Blue (blue.cc)Left the cloudFull exit | Render and AWS | Bare metal (Hetzner) | 2026 | Frontend, MySQL database and backend services | No figure; says the monthly bill fell from thousands of dollars to a fraction of that | Blue blog (Mar 2026) |
| DatapultLeft the cloudFull exit | AWS | Hetzner and OVHcloud (server type not stated) | Not stated (post Jun 2025) | Full stack, keeping ISO 27001 | Cloud bill cut by 90% from $24,000 a year on AWS | Datapult founder's Substack (Jun 2025) |
| Zig Software FoundationLeft the cloudFull exit | AWS (S3 and CloudFront) | Single server (Hetzner) | 2024 | ziglang.org website and release downloads | No figure; AWS monthly costs were rising with traffic | Zig blog (Sep 2024) |
| LinkPoolLeft the cloudFull exitVendor source | Hyperscale cloud (not named) | Colocation (Pulsant) | 2023 | Blockchain node infrastructure | Up to 85% cost savings on memory, CPU and disk in the first 12 months (vendor figure) | Pulsant case study (Aug 2023) |
| Prerender.io (saas.group)Left the cloudFull exitPartial source | AWS | In-house setup (not named) | 2022 (date per Indie Hackers; not in the cited post) | Rendering and cache infrastructure | Annual server costs cut from about $1M to about $200K, about $800K a year saved | saas.group blog (Mar 2026) |
| freedesktop.org (X.Org Foundation)Left the cloudFull exitMoved again in March 2025 when donated hosting ended | Google Cloud | Bare metal (Packet / Equinix Metal) | 2020 to 2021 | Self-hosted GitLab for open-source projects | Google bill peaked above $6,000 a month in Jan 2020 and was near $3,000 before the move; projected to save about $30 a day; the new hosting was donated, and when that ended it moved again in March 2025 | The Register (Apr 2021) |
| SpyFuLeft the cloudFull exitVendor source | AWS | Bare metal (phoenixNAP) plus own storage | Not stated | Data processing, Elasticsearch and storage | Monthly cloud cost cut 50% vs its previous AWS setup (vendor case study) | phoenixNAP case study |
| DropboxLeft the cloudPartial moveMore than 90% of user data on own infrastructure; AWS holds the rest (10-K FY2025) | AWS | Own servers in colocation | 2015 to 2016 | User file storage (Magic Pocket); more than 90% of user data now on its own infrastructure, the rest on AWS | Infrastructure costs fell $39.5M in 2016 and $35.1M more in 2017, per its S-1, which credits infrastructure optimization | Dropbox S-1 (2018) |
| 37signals (Basecamp, HEY)Left the cloudPartial moveSavings overlap with the S3 row | AWS | Own servers in colocation | 2022 to 2023 | Compute for seven cloud apps, including HEY (S3 storage left later, in a separate row) | Cloud bill cut from $3.2M a year to $1.3M in 2024; projects well over $10M saved over five years, a total that includes the S3 exit | 37signals blog (DHH) |
| X (Twitter)Left the cloudPartial moveCut cloud use and moved more on premises; original post read through reproductions | Not named | Own data centers | 2023 | Media storage and other workloads (cloud use cut, not ended) | Monthly cloud costs down 60%; cloud data storage size down 60%; cloud data processing costs down 75% | X Engineering post on X (Oct 2023) |
| MozLeft the cloudPartial moveCTO kept up to 2,000 servers on AWS | AWS | Own data centers (private cloud) | 2012 to 2014 | Crawl, index and other 24/7 processing (kept up to 2,000 servers on AWS) | Spent $6.2M on AWS vs $2.8M on own data centers in 2013; CTO expected AWS to fall to $500K to $1M a year | GeekWire (Feb 2014) |
| Saxo BankLeft the cloudPartial move | Not named | Own data centers (self-hosted Kubernetes) | 2024 to 2025 | Core parts of its Kubernetes platform | Costs down 80%, cluster creation fifteen times faster, 30% better CIS benchmark (KubeCon EU 2025 abstract) | KubeCon EU 2025 talk |
| RailwayLeft the cloudPartial moveGoogle Cloud account suspension in May 2026 took down its API, control plane and databases | Google Cloud | Own servers in colocation | 2024 to 2025 | Customer workloads on Railway Metal; API, control plane, databases and some customer compute were still on Google Cloud in May 2026 | No savings figure; cited the effect of Google Cloud pricing, including egress fees, on what it could offer customers | Railway blog (Jan 2025) |
| BlueskyLeft the cloudPartial move | AWS | Own servers in colocation; some bare metal (Vultr) | 2023 to 2024 | Core data services; personal data servers (small services stay with a cloud provider) | No figure; says bandwidth and storage in a data center are a lot cheaper than in the cloud, citing egress fees | Fork Around And Find Out podcast (Feb 2025) |
| GrabLeft the cloudPartial move | Not named | Own Mac minis in colocation (Malaysia) | 2025 | macOS build and test machines for iOS apps | Projected TCO savings of over $2.4M over three years | Grab Engineering blog (Nov 2025) |
| OneUptimeLeft the cloudPartial move | AWS | Own servers in colocation | 2023 | SaaS stack on Kubernetes (an AWS failover cluster, archives and CDN stay) | Saves over $1.2M a year, over 76% less than AWS (2025 update); was saving about $230K a year in 2023 | OneUptime blog (Oct 2025) |
| ProsopoLeft the cloudPartial move | AWS (MongoDB Atlas) | Bare metal (Hetzner) | 2025 | MongoDB database (from MongoDB Atlas) | Database cost cut from over $3,000 a month to a $160 a month server, about 95% lower; excludes backup storage | Prosopo blog (Nov 2025) |
| Fleet Device ManagementLeft the cloudPartial move | AWS (CloudFront) | Bare metal (Hetzner) | 2021 | Agent update server (downloads) | Peak monthly CloudFront bill of $2,457 (the month before was $370) vs $41.63 on Hetzner; the source calls it a 59x saving | Fleet engineering page |
| Namespace LabsLeft the cloudPartial move | AWS | Own servers in colocation (US and Europe) | 2024 | CI and build compute platform | No figure; says the unit economics on AWS didn't quite work out | Namespace blog (Feb 2026) |
| Wikimedia FoundationLeft the cloudPartial move | AWS | Bare metal (provider not named) | 2024 | Web performance testing tools and results | No figure; moved for steadier test results and more tests for a lower total price | Wikimedia Phabricator blog (Aug 2024) |
| RedNote (Xiaohongshu)Left the cloudPartial move | Not named | Own data centers | Not stated | Core online services (public cloud kept as backstop) | No figure; began building its own cloud once resources reached large scale, citing cost control | InfoQ China, QCon Beijing talk |
| OptoroLeft the cloudPartial move | AWS | Own data center (on premises) | Before 2018 | Application platform (still bursts Kubernetes into the cloud) | Projected 60% savings when it ran the numbers; later said it saved more than that | The New Stack (May 2020) |
| SeagateLeft the cloudPartial move | AWS | Private cloud on own storage | Not stated (reported 2020) | Hadoop data lake only (hybrid of AWS, Azure and own data centers) | Spend cut by as much as 25%, per CIO.com's summary of an interview with its CIO | CIO.com (Mar 2020) |
| Derby City CouncilLeft the cloudPartial moveVendor source | Not named | Own private cloud (Nutanix) | By 2019 | 400+ data center workloads (Office 365 stays) | Data center space costs from £200K a year to under £37.5K after moving to a cheaper provider; about £1.7M IT savings projected over five years | Nutanix case study |
| Hashgraph (Hedera)Left the cloudPartial moveVendor sourceTest environments only | Google Cloud (GKE) | Bare metal (Latitude.sh) | Not stated | CI/CD testing and performance environments only | Compute cost cut from $60K to $75K a month on GKE to $10K a month, 83% to 87% lower (vendor case study) | Latitude.sh case study |
| Magic EdenLeft the cloudPartial moveVendor sourceValidators only | AWS | Bare metal (Latitude.sh) | Not stated | Solana validator infrastructure only (still tied to its Kafka clusters on AWS) | Nearly 90% lower infrastructure cost (vendor case study) | Latitude.sh case study |
| ZoomLeft the cloudPlannedStated strategy; 10-K still lists AWS and Oracle Cloud | AWS and Oracle | Leased colocation data centers | Stated in 2023 | Meeting capacity; its 10-K (Feb 2026) still lists AWS and Oracle Cloud for certain critical functions | No figure; 2023 ESG report says it will migrate more into leased colocation data centers | Runtime, citing Zoom 2023 ESG report |
| Health Data Hub (France)Left the cloudPlannedNo data moved as of April 2026 | Azure (named in secondary reports) | Regional cloud (Scaleway) | Planned for late 2026 to early 2027 | National health data platform | No figure; sovereignty-driven move; an April 2026 report says no data had moved yet | Health Data Hub release (Apr 2026) |
| GEICOLeft the cloudUnder wayTarget, program began 2023; no share repatriated reported | Not named | Own private cloud in colocation (OCP) | 2023 to 2029 (target; the paper also says 2028) | Cloud workloads; target is at least half repatriated; no share moved so far is reported | Cloud spend had passed $300M a year; more than 50% lower cost per compute core and more than 60% per GB of storage, against its prior vendor hardware and cloud services | GEICO OCP whitepaper |
| AkamaiLeft the cloudUnder way10-Q: continuing to migrate; 70% is a projection | Not named | Own cloud (Akamai Connected Cloud) | Under way in 2024 | Internal business apps on third-party clouds (migration continuing, per its 10-Q) | Third-party cloud costs fell (10-Q); company blog: 40% less in year one, 70% projected in year two | Akamai 10-Q, Q1 2024 |
| RobloxLeft the cloudUnder wayAIStill migrating in 2026 | Not named | Own data centers (GPUs) | 2025 to 2026 | AI inference (migration continuing in 2026) | 10x efficiency gain for certain AI workloads after Q4 2025 infrastructure investments; no dollar figure | Roblox 8-K, Feb 2026 |
| CloudastructureLeft the cloudUnder wayStill hosts on Google and Amazon | AWS and Google Cloud | Own servers in colocation | Not stated | Cloud video surveillance services (moving more services; still hosts on Google and Amazon) | No figure; moving services to its own computers in colocation for the same result at lower cost | Cloudastructure 10-K FY2025 |
| LendsqrLeft the cloudCloud to cloud | AWS | Hetzner Cloud | Not stated (post May 2026) | Core app stack (a few AWS services kept) | Monthly bill cut from about $13,000 to $1,113 in the first full month, about 91% lower | Lendsqr blog (May 2026) |
| BHASHINI (India, MeitY)Left the cloudCloud to cloudAIVendor source | Hyperscale cloud (not named) | GPU cloud (Yotta) | 2025 to 2026 | National language-AI stack (200+ TiB) | 20 to 30 percent cost savings (Yotta's figure); done with zero data loss, per Yotta | Yotta press release (Feb 2026) |
| Digital SocietyLeft the cloudCloud to cloud | AWS and DigitalOcean | Regional cloud (Hetzner Cloud) | 2025 | Hosted services (AWS kept for DNS and email sending) | Monthly bill of $559.36 (peak months, added across providers) vs $132.96, a 76% saving, while tripling capacity | Digital Society blog (Oct 2025) |
| everysizeLeft the cloudCloud to cloudVendor source | AWS | Hetzner Cloud (managed) | Not stated (case study May 2025) | App servers, MySQL and Elasticsearch (Google Cloud kept for some workloads) | Annual server costs cut from €50,000 to just under €10,000, an 80% saving (vendor case study) | We Manage case study (May 2025) |
| HopsworksLeft the cloudCloud to cloud | AWS | Regional cloud (OVHcloud) | 2024 | Serverless SaaS and its ~8,000 users | AWS hosting was about $8K a month; OVHcloud's own case study says overall cloud costs fell 62%; AWS egress was a key risk | Hopsworks blog (Mar 2025) |
| Ola (ANI Technologies)Left the cloudCloud to cloudDestination is a group company's cloud | Azure | Krutrim cloud (Ola group company) | 2024 | Workloads that ran on Azure (Ola also used AWS) | No figure; Azure spend went to zero within a week, after a dispute with Microsoft-owned LinkedIn | Business Standard, quoting CEO on X |
| TarteelLeft the cloudCloud to cloudAIVendor source | AWS | GPU cloud (CoreWeave) | 2023 | Speech recognition and text-to-speech inference | About 56% lower cost and 22% better latency (vendor case study) | CoreWeave case study (Mar 2023) |
| searchmysite.netLeft the cloudCloud to cloud | AWS | Regional cloud (Hetzner Cloud) | 2022 | Search engine server | About £35 a month on AWS vs about £7 a month on Hetzner, around 80% cheaper | searchmysite.net blog (Dec 2022) |
| ParlerLeft the cloudCloud to cloudForced off AWS; came back on SkySilk, a cloud host | AWS | SkySilk (cloud host), from Feb 2021 | 2021 | Apps and website (all hosting) | Forced exit, not a cost decision: AWS suspended its services in January 2021 | W.D. Wash. court order (Jan 2021) |
| SystranLeft the cloudCloud to cloudVendor source | AWS (S3) | Regional cloud (OVHcloud) | Not stated (blog Feb 2021) | Translation models and data (object storage) | S3 bill was about $4K a month; three-month cost about $13K on AWS vs $11K on OVHcloud; paid back in two months | OVHcloud blog (Feb 2021) |
| Arcee AILeft the cloudCloud to cloudAIVendor source | AWS (EKS) | GPU cloud (Together AI) | Not stated | Model inference | No dollar figure; up to 95% faster time to first token; CEO cited better price-performance | Together AI case study |
| Mantis Ad NetworkLeft the cloudCloud to cloudVendor source | Not named | OVHcloud bare metal, later OVHcloud public cloud | Not stated | Compute and broader infrastructure | No figure; founder cited legacy costs it couldn't get out of; bundled bandwidth, no transfer fees | OVHcloud case study |
| ShipitLeft the cloudCloud to cloudVendor source | AWS | Regional cloud (UpCloud) | Not stated | Shipping SaaS platform | Ongoing cloud costs roughly half what its previous provider would have cost; about 130ms faster responses (vendor case study) | UpCloud case study |
| Vitec MV (Vitec Software Group)Left the cloudCloud to cloudVendor source | AWS and Microsoft | Regional cloud (UpCloud) | Not stated | Production across three environments | No cost figure; moved for European hosting and jurisdiction; 10-15 ms latency gain (vendor case study) | UpCloud case study |
| ZyngaLeft the cloudReversed or endedAWS still hosted a third of traffic; back on AWS by end of 2015 | AWS | Own infrastructure in leased data centers | 2011 | Most social game traffic (AWS still hosted about a third) | No figure | Zynga 10-K FY2011 |
| HathoraLeft the cloudReversed or endedAbout 20% stayed on cloud; game hosting wound down after the Fireworks AI purchase | AWS (EKS) | Bare metal (providers not named) | Not stated (interview Jul 2025) | Game server compute (about 20% stayed on AWS and Google Cloud); Hathora was bought by Fireworks AI and is winding down game hosting | CTO: moving 80% of workloads to bare metal cuts compute costs nearly in half and bandwidth over 90% | The CTO Club interview (Jul 2025) |
| GrouponLeft the cloudReversed or endedBack to multi-cloud by early 2023 (10-K FY2024) | AWS | Colocation (RagingWire, Sacramento) | 2011 to 2016 | Cloud-hosted infrastructure (moved back to multi-cloud by early 2023, in a separate row) | No figure; its data center chief said cost was the leading driver of the shift to colo | Data Center Frontier (Aug 2016) |
| xAI (SpaceX)Moved to the cloudAIVendor source | Own data centers (Colossus) and Oracle | Oracle Cloud | 2025 | Training and inference for next-generation Grok models | No figure; Oracle cites the cost efficiency of its AI infrastructure | Oracle press release (Jun 2025) |
| GrouponMoved to the cloudGroupon reversed its 2011 to 2016 move to colocation | Own data centers | Multi-cloud (providers not named) | 2023 | Public websites, internal applications and services, back-end BI systems | No figure; completed the migration from its own data centers in early 2023 | Groupon 10-K FY2024 |
| X (Twitter)Moved to the cloud | Own data centers | Google Cloud | 2018 | Cold storage and some Hadoop clusters (plan; 300 PB is its total Hadoop storage) | No figure; planned to move cold storage and some Hadoop clusters to Google Cloud | Google Cloud Next '19 talk |
| ZyngaMoved to the cloud | Own infrastructure | AWS | 2015 | Game traffic and computing systems | No figure; by end of 2015 AWS hosted almost all of its game traffic and computing systems | Zynga 10-K FY2015 |
| GitHubMoved to the cloud | Own data centers | Azure | 2025 to 2027 | All GitHub infrastructure | No figure; internal CTO memo cites capacity limits in its Virginia data center | The New Stack (Oct 2025) |
| AnthropicMoved to the cloudAI | None | Google Cloud (TPUs) | 2025 to 2026 | Training and inference on up to one million TPUs | No cost figure; up to one million TPUs and well over a gigawatt of capacity online in 2026 | Anthropic news (Oct 2025) |
| Monex SecuritiesMoved to the cloudPartial source | Own servers (on premises) | Oracle Cloud | 2026 | Core database system | Projected: about 40% lower cost vs expanding on-premises infrastructure | @IT (May 2026) |
| Thinking Machines LabMoved to the cloudAI | None | Google Cloud | 2026 | Training on GB300 NVL72 systems | No figure; cited speed and reliability; deal value not disclosed | ChannelLife (Apr 2026) |
| AnthropicMoved to the cloudAIVendor source | None | Azure | 2025 | Compute capacity on Azure | Committed to buy $30B of Azure compute, plus contract up to one gigawatt more capacity | Microsoft blog (Nov 2025) |
| MetaMoved to the cloudAIPartial source | Own data centers | Google Cloud | 2025 | AI infrastructure capacity (six-year deal) | Reported as more than $10B over six years; no company statement found | Data Centre Review (Aug 2025) |
| OpenAIMoved to the cloudAI | Azure | AWS | 2025 | Training and inference on NVIDIA GPUs | $38B multi-year agreement; Amazon said on Feb 27 2026 it expanded the deal by $100B over eight years | OpenAI news (Nov 2025) |
| OpenAIMoved to the cloudAI | None | Oracle Cloud (Stargate) | 2025 | Stargate data center capacity | No figure; 4.5 GW of added Stargate capacity, over 5 GW under development | OpenAI news (Jul 2025) |
| Safe Superintelligence (SSI)Moved to the cloudAIPartial source | None | Google Cloud (TPUs) | 2025 | Research and training on TPUs | No figure; Google Cloud says SSI uses TPUs for its research and development | TechCrunch (Apr 2025) |
| Stack Overflow (Stack Exchange)Moved to the cloud | Own servers in a New Jersey data center | Azure and Google Cloud | 2023 to 2025 | Stack Overflow for Teams and the public sites | No figure; the trigger was its New Jersey data center vendor closing the site | Stack Overflow blog (Dec 2025) |
| AppleMoved to the cloudAI | None | Google Cloud (TPUs) | 2024 | Pre-training of Apple Foundation Models | No figure; trained on Google Cloud TPU clusters (8192 TPUv4 chips for the server model) | Apple paper on arXiv (Jul 2024) |
| PicnicMoved to the cloud | Own servers in its warehouse | AWS | 2022 to 2024 | Warehouse control system | No figure; started in the cloud, ran warehouse servers on site, then moved them to AWS after high-impact incidents, more upkeep and features built twice | devopsdays Amsterdam 2024 talk |
| Riot GamesMoved to the cloudVendor source | Own data centers | AWS | 2017 to 2024 | League of Legends and VALORANT game servers | No figure; cited flexibility and reliability; 14 data centers closed | AWS blog on re:Invent 2023 talk |
| Character.AIMoved to the cloudAIVendor source | None | Google Cloud (TPUs and H100 VMs) | 2023 | LLM training and inference | No figure; uses Google Cloud TPUs and A3 H100 VMs to train and serve its models | Google Cloud press release (May 2023) |
| MidjourneyMoved to the cloudAIVendor source | None | Google Cloud | 2023 | Model training (TPUs) and inference (GPUs) | No figure; trains on Google TPUs and runs inference on Google GPU clusters | Google Cloud press release (Mar 2023) |
| SquarespaceMoved to the cloudPartial source | Own data centers | Google Cloud | 2022 to 2023 | Data group's platform (one team) | No figure; goal was 100% cloud native by end of 2023; Google says the move of critical workloads is still under way | LeadDev talk (Dec 2022) |
| UberMoved to the cloudVendor source | Own data centers | Oracle Cloud and Google Cloud | 2023 | Some critical workloads to each provider; AWS use unchanged (press) | No figure; CEO said it needed a cloud provider that would help reduce overall infrastructure costs | Oracle press release (Feb 2023) |
| Stability AIMoved to the cloudAIVendor source | None | AWS | 2022 | Foundation model training | Amazon release says SageMaker cut training time and costs by 58% vs self-managed infrastructure | Amazon press release (Nov 2022) |
| WayfairMoved to the cloud | Own data centers | Google Cloud | 2021 to 2022 | All data center applications and shared services | No figure; moved 100% of its applications and shared services to Google Cloud | Wayfair tech blog (Oct 2022) |
| Capital OneMoved to the cloud | Own data centers | AWS | 2020 | Thousands of applications; eight data centers closed | No figure; an executive cited the flexibility to scale quickly if its projections were wrong | SiliconANGLE at AWS re:Invent (Dec 2020) |
| EtsyMoved to the cloudVendor source | Own servers | Google Cloud | 2018 to 2020 | Marketplace site and data (2,000 servers, 5.5 PB) | No figure; described as more cost effective; completed a two-year migration | Google Cloud press release (Feb 2020) |
| SpotifyMoved to the cloud | Own colocated data centers | Google Cloud | 2015 to 2018 | About 2,000 services and 100+ PB of data | No figure; closed the first of four data centers in December 2017 and the other three in 2018 | Spotify Engineering blog (Dec 2019) |
| Customer.ioMoved to the cloud | Bare metal (Hetzner, then OVH) | Google Cloud | 2017 | Messaging platform | No savings figure; bare metal peaked near $40K a month, Google Cloud now about $800K; moved for reliability | Co-founder's blog (Oct 2025) |
| SentryMoved to the cloud | Bare metal in colocation (Texas) | Google Cloud | 2017 | Error-tracking SaaS platform | After resizing on Google Cloud, cut infrastructure costs by roughly 20% | Sentry blog (Oct 2017) |
| NetflixMoved to the cloud | Own data centers | AWS | 2008 to 2016 | Streaming service infrastructure | No figure; completed its cloud migration and shut down its last data center pieces | Netflix (Feb 2016) |
| PinterestStayed or renewedVendor source | AWS | AWS (expanded commitment) | 2026 to 2031 | Cloud services, including AI visual search | Planned $4B commitment for AWS cloud services through 2031 | Amazon press release (Jun 2026) |
| LyftStayed or renewed | AWS | AWS (renewed commitment) | 2026 to 2030 | Platform hosting | Committed to spend at least $562.5M from 2026 through 2030, at least $100M in each period | Lyft 10-K FY2025 |
| RedditStayed or renewed | AWS and Google Cloud | AWS and Google Cloud (renewed) | 2026 to 2029 | Platform hosting | Committed $880M to AWS and $580M to Google through 2029 | Reddit 10-Q, Q3 2026 |
| UK Police National Database (Home Office)Stayed or renewed | On premises (run by CGI) | Stays on premises (moving in-house) | 2024 to 2025 | Planned cloud migration of the database | Cloud move paused after about £35.1M spent; £20.3M of stabilization work expected | Home Office response to MPs (Apr 2026), via The Register |
| LinkedInStayed or renewed | Own data centers | Own data centers (kept) | 2019 to 2023 | Planned move of its data center stack to Azure | No figure; shelved the Azure move and kept investing in its own data centers, using Azure alongside | CNBC via NBC10 (Dec 2023) |
| GitLabStayed or renewed | Not named | Stayed in the cloud | 2016 to 2017 | Planned move of GitLab.com to own bare metal | No figure; dropped the bare metal plan, citing simplicity and not wanting to be an infrastructure company | GitLab blog (Mar 2017) |
| AirbnbStayed or renewed | AWS | Unnamed data hosting provider (AWS named in a risk factor) | Not stated | Platform hosting | Committed at least $1.7B for vendor services through 2031; AWS named as its primary host | Airbnb 10-K FY2025 |
| Standard CharteredStayed or renewedAIPartial source | Not named | Stays on public cloud | Not stated | General workloads (no major repatriation) | No figure; no major repatriation, though some workloads may move back on premises (paraphrase) | TechTarget (Sep 2026) |
| DropboxNever moved to cloudAI | None | Own GPU servers in colocation | 2025 | Dash AI: embeddings, inference and fine-tuning | No figure; keeps control over performance, cost and energy efficiency | Dropbox tech blog (Jul 2025) |
| comma.aiNever moved to cloudAI | None | Own data center (San Diego office) | 2026 | Model training, metrics and data (600 GPUs) | About $5M spent vs an estimated $25M+ had it done the same work in the cloud | comma.ai blog (Feb 2026) |
| ZSky AINever moved to cloudAIPartial source | None | Own GPUs (RTX 5090) | 2026 | AI image generation inference | About $1,472 a month self-hosted (7 GPUs) vs an estimated ~$12,000 a month cloud equivalent (priced for 12) | ZSky AI blog (Apr 2026) |
| Mistral AINever moved to cloudAI | Not named | Own AI cloud (Mistral Compute) | 2025 | AI compute for Mistral and its customers | No figure; pitched as an alternative to US or China-based cloud and AI providers | Mistral AI news (Jun 2025) |
| VeridionNever moved to cloud | None | Bare metal plus own GPU rigs | 2025 | Crawling, big data and GPU clusters | About $60K a month actual vs about $2M a month modeled at cloud list prices | Veridion blog (Nov 2025) |
| DeepSeek (High-Flyer)Never moved to cloudAI | None | Own cluster (Fire-Flyer 2, China) | 2024 | Deep learning training (10,000 A100 GPUs) | Costs cut by half and energy use by 40% vs NVIDIA DGX-A100 systems (not a cloud comparison) | DeepSeek paper on arXiv (Aug 2024) |
| AhrefsNever moved to cloud | None | Rented colocation (Singapore) | 2020 to 2023 | Crawler and index (850 servers) | Estimates about $400M avoided over about 2.5 years vs a modeled AWS cost of about $448M | Ahrefs blog (Mar 2023) |
| RBC (Royal Bank of Canada)Never moved to cloudAI | None | Own GPU farm (private cloud) | 2020 | Machine learning platform | No figure; bought chips directly from Nvidia to cut out the middle person | Data Center Knowledge (Aug 2020) |
| Bank of AmericaNever moved to cloud | None | Own private cloud (own data centers) | 2019 | Core infrastructure (backbone) | Expenses cut about 40%, about $2B a year, vs its own old infrastructure, not vs public cloud | Bank of America Q3 2019 earnings call |
| PhonePeNever moved to cloud | Mostly own servers; some data on public cloud | Own servers in colocation (India) | 2016 | Payments platform (1.04M cores, 30.95 PB) | Server cost about Rs 0.06 per transaction; Rs 33.73B spent on technology infrastructure from 2016 to Sep 30 2025 | PhonePe draft prospectus (Jan 2026) |
| Inflection AIAdded a GPU cloudAIPartial sourceNot an exit: added a platform; no prior cloud stated | None stated | GPU cloud (Intel Tiber AI Cloud) | 2024 | Inflection 3.0 for enterprise | Up to 2x better price performance on Intel Gaudi 3, per the company | Inflection AI blog (Oct 2024) |
| AnthropicAdded a GPU cloudAI | AWS and Google Cloud (both kept) | Rented GPU cluster (SpaceX Colossus 1) | 2026 | Claude inference capacity for Pro and Max | No price or term stated; over 300 MW and 220,000+ NVIDIA GPUs within the month | Anthropic news (May 2026) |
| Character.AIAdded a GPU cloudAIVendor sourcePartial source | Other providers (not named) | Alternative cloud (DigitalOcean) | 2026 | LLM inference on AMD GPUs | About 2x queries per second per server vs a generic setup on other providers (vendor case study) | DigitalOcean blog (Jan 2026) |
| Codeium (Windsurf)Added a GPU cloudAIVendor sourcePartial source | Not named | GPU cloud (Crusoe) | 2025 | GPU infrastructure for its code assistant | 50% cost savings vs quotes from traditional cloud providers (vendor case study) | Crusoe blog (Jan 2025) |
| OpenAIAdded a GPU cloudAIVendor source | Azure | GPU cloud (CoreWeave) | 2025 | AI training and inference capacity beyond Azure | No savings figure; total CoreWeave contract value up to about $22.4B | CoreWeave press release (Sep 2025) |
| HedraAdded a GPU cloudAIVendor sourcePartial source | Not named | GPU cloud (Together AI) | Not stated | Video generation inference | 60% total cost savings (vendor case study) | Together AI case study |
| LinumAdded a GPU cloudAIVendor sourcePartial source | Not named | GPU cloud (Crusoe) | Not stated | Text-to-video model training | 25% lower training costs (vendor case study, via an aggregator) | Crusoe case study (via casestudies.com) |
| PoolsideBuilt new AI capacityAIPartial sourceNot an exit: new campus and CoreWeave cluster; no AWS move in the source | None stated | Own 2 GW campus (Texas) with a CoreWeave GPU cluster | 2025 | New model training capacity | No figure; the post announces a campus and a CoreWeave deal and does not mention moving off AWS; the FT later reported the CoreWeave partnership was unraveling | Poolside blog (Oct 2025) |
| xAI (SpaceX)Built new AI capacityAIPartial sourceNot an exit: kept Oracle GPUs, then added Oracle for next-generation Grok models (row AI-02b) | Oracle Cloud (kept) | Own data centers (Colossus, Memphis) | 2024 | Frontier model training (Grok 2 trained on Oracle GPUs) | No cost figure; ended talks on a larger Oracle deal over timing and built its own cluster; kept renting about 16,000 Oracle GPUs | DCD, quoting Musk on X |
| CohereBuilt new AI capacityAIPartial sourceNot an exit: new government-backed capacity | None stated | New AI data center in Canada (partners not named) | 2024 | New training and inference capacity | No figure; up to $240M, the first investment under Canada's $2B Sovereign AI Compute Strategy | Government of Canada release (Dec 2024) |
| AnthropicBuilt new AI capacityAI | AWS and Google Cloud (both kept) | Own data centers (built with Fluidstack) | 2025 to 2026 | New training and inference capacity (Texas, New York) | No figure; says it will prioritize cost-effective, capital-efficient approaches | Anthropic news (Nov 2025) |
| Eli LillyBuilt new AI capacityAIVendor source | None | Own NVIDIA DGX SuperPOD (Indianapolis) | 2025 to 2026 | Drug discovery and genomics AI | No figure; NVIDIA calls it the most powerful AI factory wholly owned by a pharma company | NVIDIA blog (Feb 2026) |
| PNC Financial ServicesBuilt new AI capacityAIPartial source | Token-priced AI APIs | Own AI data centers (planned) | 2026 | AI models and GPU compute | No figure; CEO says PNC is building its own AI factory to avoid third-party AI vendor costs | American Banker (Jun 2026) |
| Hyundai Motor GroupBuilt new AI capacityAIPartial source | None | NVIDIA AI factory in Korea (hosting not stated) | 2025 | Physical AI training and deployment | No figure; plans AI training, validation and deployment on 50,000 NVIDIA Blackwell GPUs | Hyundai press release (Oct 2025) |
| Lockheed MartinBuilt new AI capacityAI | None | Own air-gapped AI factory (on premises) | 2025 | Generative AI (Google Gemini) on premises | No figure; Google's AI deployed in its secure, on-premises, air-gapped environments | Lockheed Martin press release (Oct 2025) |
| Mayo ClinicBuilt new AI capacityAI | None | NVIDIA DGX SuperPOD (location not stated) | 2025 | Digital pathology model training | No cost figure; expected to cut four weeks of work to one | Mayo Clinic News Network (Jul 2025) |
| BNYBuilt new AI capacityAI | None | Own NVIDIA DGX SuperPOD (on premises) | 2024 | AI platform compute | No figure; expands its on-premises AI infrastructure | BNY press release (Mar 2024) |
| New Belgium BrewingBuilt new AI capacityAI | None | On premises (Azure still used for sales data, ERP and CRM) | Not stated | Specialized open AI models | No figure; its IT operations director says a frontier model is not fit for purpose for everything | TechTarget (Sep 2026) |
| Belastingdienst (Dutch Tax Administration)National or sovereign move | Microsoft 365 | Own data centers and open source tools | 2026 to 2028 | Email, calendars, storage and collaboration (about 5,000 of 47,500 staff had moved) | No figure; a review found the public cloud design fell short on security; email moves to own servers in 2027 | The Register (Oct 2026) |
| Deutsche TelekomNational or sovereign moveAI | None | Sovereign AI cloud (Munich) | 2025 | Sovereign AI compute for German industry | No figure; says AI inference should run within Germany for true AI sovereignty | Deutsche Telekom (Nov 2025) |
| European Union (InvestAI)National or sovereign moveAI | None | EU AI gigafactories | 2025 | Public AI gigafactories for model development | No savings figure; a new €20 billion European fund for AI gigafactories | European Commission (Feb 2025) |
| International Criminal CourtNational or sovereign movePartial source | Microsoft Office software | openDesk (hosting not stated) | 2025 | Office and collaboration software | No figure; the court confirmed the move after US sanctions; Microsoft says it never suspended service | The Register (Oct 2025) |
| WalmartWatch listPartial sourceNot shown as an exit: hybrid, still on Azure and Google Cloud | Azure and Google Cloud | Own private cloud (OpenStack) | 2020 to 2022 | Hybrid setup: own OpenStack cloud alongside Azure and Google Cloud | TechTarget reported 10% to 18% lower cloud spending; a Walmart VP said it can run at much lower cost; no workload is shown leaving a cloud | TechTarget (Jul 2022); AWS filing to UK CMA (Nov 2023) |
| Klink AIWatch listAIVendor sourcePartial sourceNot shown as an exit: bare metal to bare metal | AWS, then IBM SoftLayer | Bare metal (Hivelocity) | 2019 | AI platform compute and storage (the described move is bare metal to bare metal) | Five times cheaper than IBM SoftLayer and AWS, per its co-founder (vendor case study) | Hivelocity case study (Jul 2019) |
| JPMorgan ChaseWatch listAIPartial sourceHybrid plan, not a renewal or a decision to stay | Not named | Own data centers, colocation and cloud | 2025 | AI experiments in cloud; at-scale AI hybrid | No figure; cloud for experiments, own and colocation capacity for control and regulatory reasons | CIO Dive (Jan 2025) |
How this dataset is built
- What counts. A row is a specific move that a company, a filing, a regulator or a government body describes in public: workloads off public cloud, onto it, a decision to stay, or new AI capacity built instead of rented. Plans with a named destination count and are labeled as plans.
- Exit status. "Full exit" means the source shows all of the company's hosting, or all of the named product, off public cloud and the move finished. "Partial move" means it finished but the source shows cloud kept, or only one workload moved. "Planned" and "Under way" are not finished. "Cloud to cloud" means the destination is another cloud, such as Hetzner Cloud, OVHcloud or a GPU cloud. "Reversed or ended" means the company later moved back or shut the setup down.
- Sources. Each row's source type is recorded: a filing, earnings call or government record; the company's own blog, talk, paper or post; a vendor's case study or a cloud provider's release; or a press report. Vendor sources are tagged because the vendor sells the destination.
- Cost figures. Kept in the words of whoever stated them, with projections left as projections. We do not convert unit-cost cuts into dollar totals. "Cost figure stated" means the source gives a dollar amount, a percentage or a multiple for cost, or a contract value.
- Watch list. Rows we once counted as exits but whose sources show a hybrid setup or a move between hosts rather than a move off cloud. They stay in the data so the change is visible.
- Checks. Every row was checked against its source when it was added. In November 2026 a second, independent check reopened the sources; rows it found overstated were rechecked against their sources and relabeled, and every count on this page was recomputed from the data file.
- Destinations. Moves to smaller or European clouds are grouped separately so they are not counted as colocation. A company that moved twice appears once per move.
The dataset is updated as new moves are published. To suggest a company or a correction, contact us.
Cite this dataset
Cloud Repatriation Examples
Dropbox
$74.6MInfrastructure cost decreases, 2016 and 2017Moved the vast majority of user content off AWS into its own infrastructure in colocation facilities it leases and operates. Its IPO filing credits the program with infrastructure cost decreases of $39.5 million in 2016 and $35.1 million in 2017.
Dropbox S-1, 201837signals (Basecamp, HEY)
$3.2M to $1.3MAnnual cloud bill, run rate to 2024Cut its cloud bill from a $3.2 million a year run rate to $1.3 million for 2024, and projects total savings of "well over ten million dollars over five years" after also moving its storage off AWS S3.
37signals, Oct 2024GEICO
Over 50% lowerCost per compute coreWith cloud bills over $300 million a year, the insurer is moving to its own Open Compute hardware in colocation and reports more than 50% lower cost per compute core and more than 60% per gigabyte of storage than its prior hardware and cloud services. Its target is to move at least half its cloud workloads by 2029.
GEICO whitepaper; The StackX (Twitter)
60% lowerMonthly cloud costsSaid in October 2023 that doing "much more on-prem" and optimizing its cloud use had cut its monthly cloud costs by 60%.
X EngineeringAkamai
40% lowerPublic cloud spend, year oneA cloud provider moving its own IT off the large public clouds onto its own platform. Its 10-Q says third-party cloud costs fell, and its own blog says public cloud spend dropped 40% in year one, with 70% projected in year two.
Akamai 10-Q; Akamai blogZynga, out and back
2011 out, 2015 backGame traffic off AWS, then back onMoved most game traffic onto its own infrastructure in 2011, when AWS hosted about a third of it. By the end of 2015, AWS hosted "almost all of our game traffic and computing systems." Repatriation is not always one-way.
Zynga 10-K, 2011; 10-K, 2015Cloud Repatriation and AI in 2026
AI is the loudest part of the cloud repatriation conversation and the thinnest part of the record. We found four named AI workloads moved off a public cloud with a public source, and none is a finished move to owned hardware. Arcee AI, BHASHINI (India, MeitY) and Tarteel went to another cloud, each on the vendor's own word, and Roblox, the only public company that says so in an SEC filing, is still migrating. Roblox told shareholders in February 2026: "In 2026, we will continue to migrate AI inference workloads from third-party cloud providers to our own GPU-equipped data centers." (Roblox 8-K)
Meanwhile the largest AI buyers are signing bigger cloud contracts, not leaving. OpenAI committed an incremental $250 billion to Azure and $38 billion to AWS, a deal Amazon expanded by a further $100 billion in February 2026. Anthropic committed more than $100 billion to AWS and $30 billion to Azure, while also building its own data centers with Fluidstack in a $50 billion program. Most AI "moves" are a second provider added, not an exit.
Azure's list price for an H100 GPU-hour on demand, against a modeled owned H100 in colocation at full utilization. At 50% utilization the owned figure roughly doubles and loses to a three-year AWS reservation at $2.97. Utilization decides it.
Azure retail price API; AWS pricing; CloudRift model (Jan 2026)A GB300 NVL72 rack draws up to 142 kW, about 16 times the roughly 9 kW average rack in Uptime Institute's 2025 survey, and more than 80% of facilities have no rack above 30 kW. Finding a facility that can take it is usually the hard part.
Vertiv GB300 reference design; Uptime Institute 2025For steady training or inference, owning rack-scale hardware in a liquid-cooled facility usually wins over two years or more. For bursty work, rented GPUs still do. Our AI and GPU colocation page covers the facility side, and GPU rental covers what renting costs against owning.
When Does Cloud Repatriation to Colocation Make Financial Sense?
Cloud repatriation to colocation makes financial sense when three things are true at once: the workload runs at steady utilization, roughly 60 percent or more around the clock; the spend is large enough that the overhead of dedicated infrastructure is small next to the saving, usually $50,000 a month or more; and the comparison still wins against optimized cloud, not list price. High outbound data transfer makes the case stronger. If any of the three is missing, optimizing cloud is usually the better answer.
Roughly 60 percent or more, around the clock.
Large enough that dedicated overhead is small next to the saving.
The case has to clear what optimizing cloud alone would save.
Price optimized cloud first. Reserved capacity, committed-use discounts, cutting idle resources and a serious conversation with your account team can take 15 to 25 percent off a bill that has never been negotiated. The repatriation case has to beat that number. Our cloud vs colocation calculator gives a first read on yours.
How to Calculate Savings From Moving Off Public Cloud Into Colocation
To calculate the savings, compare total cost over three to five years, not the first year. Dedicated infrastructure loses year one to migration cost and, if you buy hardware, capital outlay. It wins years two through five if the workload is steady, and the crossover usually lands in year two. The cloud side is your bill by workload, with an optimized-cloud line to keep it fair. The dedicated side has to carry everything a cloud bill hides:
Hardware
Amortized over three to five years, with a refresh cycle built in, or leased if you prefer operating expense.
Facility
Colocation rent per kilowatt or bare metal rental, power, cross-connects and remote hands. Primary-market colocation asked about $196 per kW per month for 250 to 500 kW deals in CBRE's H2 2025 survey. Current figures are in our data center cost guide.
Interconnect
Private links back to whatever stays on cloud, plus transit for the traffic that used to ride the cloud provider's network.
Operations
Remote hands and a managed service provider rather than new hires, usually 5 to 10 percent of dedicated infrastructure spend.
Platform rebuild
Whatever replaces the managed services you use today: databases, load balancers, Kubernetes, observability. Each is a product on cloud and a project on dedicated infrastructure, and it is often the larger line.
Migration and risk
The one-time move, months of running both environments in parallel, and headroom for hardware failures and growth.
A worked example at enterprise scale
The figures below are illustrative, built for a data platform company spending $1.2 million a month on cloud. They show the shape of the analysis, not any client's numbers.
| Monthly line | Cloud today | Hybrid, year one | Hybrid, year three |
|---|---|---|---|
| Cloud compute and storage | $720,000 | $310,000 | $290,000 |
| Cloud egress | $280,000 | $60,000 | $55,000 |
| Cloud managed services | $200,000 | $120,000 | $110,000 |
| Bare metal rental | 0 | $210,000 | $40,000 |
| Colocation facility | 0 | $25,000 | $95,000 |
| Hardware amortization | 0 | $20,000 | $130,000 |
| Interconnect and transit | 0 | $35,000 | $40,000 |
| Operations (remote hands, managed service) | 0 | $30,000 | $30,000 |
| Platform rebuild (managed services replaced) | 0 | $40,000 | $45,000 |
| Migration, amortized | 0 | $60,000 | 0 |
| Monthly total | $1,200,000 | $910,000 | $835,000 |
| Annual saving | $3.5 million | $4.4 million |
Illustrative model, Metro Colo Advisory, November 2026. Free to cite with a link to this page.
Three things stand out. Egress falls furthest in percentage terms, about 79 percent, and it arrives first: from $280,000 a month to $60,000 in year one. Bare metal carries year one while the colocation build runs in parallel, which is why bare metal is high in year one and low in year three. And operations plus platform rebuild, at $70,000 to $75,000 a month, is the line in-house models leave out: without it the example would show $5.3 million a year instead of $4.4 million, overstating the saving by about a fifth. The year-three colocation footprint here is about 500 kW, a wholesale colocation lease rather than a few cabinets.
Cloud Egress Fees and the Exit Waivers From AWS, Azure and Google Cloud
Egress is what a cloud provider charges to move data out of its network, and for a business that sends a lot of data outward it is often the largest single line on the bill. AWS charges $0.09 per gigabyte for the first 10 TB a month of data transfer out to the internet, stepping down to $0.05 above 150 TB. Azure starts at about $0.087, and Google Cloud's premium tier starts higher. At AWS list prices, 500 TB a month comes to about $29,500 in egress alone, before compute or storage. On dedicated infrastructure that line becomes a network port and a transit commit: Fleet reported an AWS CloudFront bill that reached $2,457 in one month, for traffic that costs about 37 euros a month on Hetzner.
Sources: AWS pricing; Azure bandwidth pricing; Google Cloud network pricing; Fleet. The 500 TB figure is our calculation from AWS's published tiers.
The one-time cost of getting your data out when you leave is now largely waived. All three providers offer exit credits, and the European Union goes further: under the Data Act, providers may charge only cost-based switching fees until January 12, 2027, and no switching charges after that. The conditions matter for a phased move.
| Provider | What is waived | The conditions |
|---|---|---|
| AWSsince March 2024 | Data transfer out to the internet, credited when you move all your data off AWS or off a particular service | Request through AWS Support. You do not have to close your account. Customers outside the EU have 180 days to complete the move; EU and UK customers follow AWS's Data Act and UK switching terms |
| Microsoft Azuresince March 2024 | Data transfer out to the internet when leaving Azure | Open a support request with your start date; 60 days from that date to move. All subscriptions on the account must be canceled before the credit. ExpressRoute, VPN, Front Door and CDN egress are excluded. UK-billed customers get 180 days |
| Google Cloudsince January 2024 | Data transfer fees on internet transfers when migrating off a Google Cloud service | Submit an Exit Notice, start within a 30-day initiation period, then a migration period of at least 30 days. You must terminate the service you are leaving; moving only part of a service does not qualify |
| EU Data ActArticle 29 | Switching charges, including data egress for switching | Reduced to the provider's direct cost until January 12, 2027; no switching charges allowed after that date |
Sources: AWS; Microsoft; Google Cloud; Regulation (EU) 2023/2854, Article 29. The UK competition regulator found in October 2026 that "the uptake of these schemes has been limited so far" (CMA). Terms change; confirm with your provider before you plan around them.
Repatriation from AWS: what to plan besides the data
The exit credit covers the data you take with you. It does not cover the ongoing egress on workloads that stay, and it does not unwind your commitments. Plan three other lines alongside the move: Savings Plans, Reserved Instances and any enterprise commitment, which should be reduced only as workloads leave; the managed services each workload depends on, which need a replacement; and the Direct Connect or other private links that keep the hybrid side connected. Our data center connectivity guide covers the private links.
Cloud vs Colocation vs Bare Metal: The Four Paths
Every repatriation decision resolves into one of four paths, and the differences in speed and capital are larger than the differences in cost.
| Path | What it is | Time to production | Capital | Best for |
|---|---|---|---|---|
| Stay on cloud, optimized | Reserved instances, savings plans, rightsizing, egress renegotiation | Weeks | None | Bursty workloads, small footprints, teams with no infrastructure capacity |
| Bare metal | Rented dedicated servers, no hypervisor, no capital | Weeks to three months | None; operating expense at a lower rate | Steady compute that needs to move fast, or a first step before colocation |
| Colocation, owned hardware | You own the servers and lease space and power, at scale a wholesale lease of 500 kW and up | Nine to fifteen months | Hardware amortized over three to five years, or leased | Large steady footprints and AI clusters where three-year economics matter more than speed |
| HybridWhere most land | Steady workloads on dedicated infrastructure, variable ones on cloud, private interconnect between them | Follows the dedicated path chosen | Depends on the dedicated path | Most companies above $50,000 a month; where the analysis usually lands |
On the P&L, bare metal improves EBITDA without capital outlay because it stays operating expense. Colocation with owned hardware improves it more, but part of the cost moves to depreciation and there is cash up front. For adtech, SaaS and data companies, infrastructure usually sits in cost of revenue, so a saving shows up in gross margin. Many companies do both: bare metal for the first wave, colocation once the workload profile is proven.
Which Workloads Should Move, and Which Should Stay?
Strong candidates to move
- AI training and steady inference on GPU clusters
- Databases, data warehouses and core application tiers running 24/7
- Anything with high outbound transfer: bidding, ad serving, content delivery, high-volume APIs
- Storage-heavy applications where the bill is mostly capacity you never release
- Workloads with data residency or compliance reasons for physical control
Better left on cloud
- Workloads that spike ten times or more against baseline
- Anything under roughly $20,000 a month on its own
- Applications built deep on proprietary managed services
- True global distribution across many regions at once
- Anything expected to change shape in the next twelve months
Best Practices for Migrating GPU Clusters From Cloud to Colocation
Moving a GPU cluster from cloud to colocation is a facility problem first and a migration problem second. The order that avoids the expensive mistakes:
Prove the utilization
Pull at least 90 days of GPU utilization by job. Owning wins on steady load and loses on idle hardware, so the case rests on this number.
Secure power and density before hardware
Current rack-scale systems need about 130 to 140 kW per rack and liquid cooling. Confirm the facility can deliver that density and the power on your timeline before you order a single server.
Design the network and storage together
Plan the GPU fabric, storage and the private links back to cloud as one design, so training data sits next to the compute instead of across a metered link.
Move the data on the exit waiver
Use your provider's exit credit for the bulk data move, and time it inside the waiver window.
Keep cloud or a GPU cloud for bursts
Size owned capacity for the steady base and rent the peaks. Reduce cloud commitments only as workloads land.
Our AI and GPU colocation service finds facilities that can serve the density, and GPU cluster rental covers the rented side.
Cloud Exit Strategy: The Sequence That Works
A cloud exit strategy is the order in which workloads move, the parallel-running period for each, and the point at which the cloud commitment is reduced. Done in the wrong order, it costs more than staying.
Workload assessment, weeks one to three
Billing export by workload, utilization data and the people who run the systems. This produces the movable list, the stay list and the kilowatt requirement.
Path and facility selection, weeks three to six
Bare metal, colocation or both, with facilities that fit the connectivity, density and compliance profile. See site selection.
Contracting, weeks six to ten
Escalation, minimum commitment and renewal terms negotiated before signature. See the data center lease guide.
Hardware
Bare metal is usually ready within weeks once contracted. Owned hardware is a procurement cycle of four to twelve months.
Migration in waves, three to six months
Non-critical first, production once stability is confirmed, with both environments running in parallel for each wave. See data center migration.
Cloud commitment reduction
Reserved instances and committed-use contracts come down as workloads leave, not before. Cutting the commitment early is the most common expensive mistake.
How Metro Colo Advisory Works on Cloud Repatriation
We run cloud repatriation as a fixed-fee data center consulting engagement, four to six weeks, that produces the path-by-path P&L this page describes: every cost counted, against optimized cloud, with an editable model your finance team keeps. The consulting fee is the same whether the answer is move, stay or somewhere in between.
What you get
The four paths as a P&L over three to five years with cash timing, the movable and stay lists workload by workload, your real egress line on cloud and on dedicated infrastructure, and a recommendation with the migration sequence and risks.
What it costs
From $45,000Engagements start at $45,000, fixed fee. Multi-site and portfolio work is scoped separately and commonly runs into six figures.
When the answer is move
We take the requirement to colocation and bare metal operators, find the capacity and negotiate the rate and terms. That placement is paid by the operator you choose as a referral fee, not by you, and using us for placement is never a condition of the analysis.
Send your annual cloud spend, the main workloads and what is driving the question. If the numbers do not justify the engagement, we will say so.
The public file is free to use under CC BY 4.0. For a custom cut by market, operator or date, scheduled updates, email us what you are working on. If it is a site, a deal or a lease, the same research backs our data center consulting.
Frequently Asked Questions
What is cloud repatriation?
Cloud repatriation is moving applications, data or compute from a public cloud such as AWS, Azure or Google Cloud to infrastructure you control: owned servers in colocation or on-premises, rented bare metal, or a hybrid of dedicated and cloud. It is sometimes called reverse cloud migration or a cloud exit.
Is cloud repatriation a real trend in 2026?
For specific workloads, yes. Vendor surveys claim most companies are moving workloads back, but IDC found only 8 to 9 percent plan a full exit, and the named, sourced moves in our tracker are mostly steady, data-heavy workloads moved for cost or egress, with cloud kept for the rest.
Which companies have left the cloud?
Few have left completely. Of 51 named companies and public bodies in our tracker that moved workloads off public cloud, 10 finished a full exit, including 37signals, Instagram and WhatsApp (both to Facebook's data centers) and freedesktop.org. Most moved part: Dropbox keeps more than 90% of user data on its own infrastructure and the rest on AWS, and Moz and X cut cloud use without ending it. GEICO and Akamai are still moving. Zynga and Groupon left and later moved back. The tracker on this page labels each exit and links to each source.
Is colocation more cost effective than AWS?
For steady workloads running 24/7, usually yes over three years: by our estimate, 30 to 50 percent on the workloads that move once operations and platform costs are counted. For bursty or small workloads, AWS usually costs less. The comparison has to be against optimized AWS pricing, not list price.
Does cloud repatriation actually save money after connectivity costs?
It can, when egress is a large part of the bill. Private links back to cloud and internet transit are real costs, but they are usually far smaller than per-gigabyte egress at volume. In our worked example, interconnect and transit add $35,000 to $40,000 a month while egress falls by more than $200,000.
Will AWS waive egress fees if we leave?
Yes, for the data you move out. AWS credits data transfer out to the internet when you move all your data off AWS or off a particular service, by request through AWS Support. You do not have to close your account, and customers outside the EU have 180 days to finish. Ongoing egress on workloads that stay is still charged.
Is cloud storage more expensive than on-premises in 2026?
At steady, large volumes it usually is, once retrieval and egress are included. 37signals was spending just under $1.5 million a year on AWS S3 before moving that data to its own storage hardware. For small or rarely accessed data, cloud storage is usually cheaper than owning the hardware.
How long does cloud repatriation take?
Bare metal takes weeks to provision and three to six months to migrate live workloads. Colocation with owned hardware takes nine to fifteen months from decision to production. Most enterprise moves take six to eighteen months end to end, with the saving arriving as each wave lands.
Should we renegotiate with our cloud provider instead?
Do both. A documented cost for the alternative is what moves a cloud provider on price, and in our experience some companies take 15 to 25 percent off without moving anything. The repatriation case only stands if it beats that optimized number.
Is cloud repatriation a one-way decision?
No. Zynga moved most of its game traffic to its own infrastructure in 2011 and was back on AWS by the end of 2015, and Groupon, which moved to colocation by 2016, finished moving back to multi-cloud in early 2023. Hybrid setups keep cloud connectivity, and bare metal and colocation contracts run one to five years, so workloads can move back if their profile changes.
Can I use this data?
Yes. The tracker is free to use in news coverage and research with credit to Metro Colo Advisory and a link to this page. The CSV download includes the source link for every row.
How the Tracker Is Built
A company is included when it, or a filing, regulator or government body, publicly describes a specific move of workloads off public cloud, onto it, or a decision to stay. Sources are ranked: filings and government documents first, then the company's own blog, talk or executive statement, then press reporting that quotes the company. Rows that rest on a secondary report are marked "Partial source." Vendor case studies are tagged because the vendor sells the destination, so you can weigh them accordingly. Savings are kept in the words of whoever stated them; we do not convert projections into results. Moves to smaller or European clouds are shown as their own group so they are not counted as colocation. New moves are added as they are published. To suggest a company or a correction, contact us.
Deciding and modeling: cloud vs colocation calculator, colocation pricing guide, FinOps consulting.
Where workloads go: wholesale colocation, bare metal, managed private cloud, AI and GPU colocation.
More research: all Metro Colo Advisory research and datasets.