Netflix, Capital One, Spotify, Moderna and Monzo run their core systems on public cloud. Dropbox and 37signals moved large workloads back off it. The companies that use cloud computing well didn’t simply rent someone else’s servers: each had a constraint that forced the decision, and each has published enough about it to check.
- The hardest call: whose word do you take?
- 10 companies that use cloud computing, at a glance
- 1. Netflix — rebuilt itself on AWS after a database took the business down
- 2. Capital One — a large US bank that closed every data centre it owned
- 3. Spotify — left its own hardware because it wanted Google’s data tools
- 4. Dropbox — moved most of its data off AWS and built its own storage
- 5. 37signals — left the cloud loudly, and kept publishing the numbers
- 6. Nasdaq — put a market on AWS without moving the machines
- 7. Moderna — a pharmaceutical company that never built a data centre
- 8. Monzo — a licensed bank with no data centre, running on Kubernetes
- 9. The Guardian — got out of the database business
- 10. Zoom — owned data centres, with public cloud as the shock absorber
- The patterns worth copying
- Where this goes wrong
- Frequently asked questions about companies that use cloud computing
That last part is the filter. Every organisation below has written about its own infrastructure, or filed it with a regulator, or stood on a conference stage and said it. Nothing here comes from a vendor’s marketing page unless the section says so.
Originally published March 2019 by Kayla Matthews. Reviewed and updated August 11, 2026 by the SmartDataCollective editorial team.
The hardest call: whose word do you take?
The awkward judgement in a piece like this is what to do with outcomes that only the cloud provider has announced. A press release saying a retailer “modernised” and “accelerated” tells you a deal closed. It tells you nothing about what broke, what it cost, or whether anyone would do it again.
So the order of preference here is: the organisation’s own engineering blog or post-mortem first, then regulatory filings and annual reports, then recorded conference talks, then provider case studies clearly labelled as such. Where an outcome was only ever described qualitatively, it’s described qualitatively here. You won’t find a percentage in this article that the named company didn’t publish itself.
A few names from the earlier version of this page have gone. General Electric has since separated into three independent companies, so its 2017 cloud strategy describes an entity that no longer exists. HotelTonight was acquired by Airbnb. Fitbit is now part of Google, which makes its “partnership with Google Cloud” a different kind of story. Kroger’s Azure work was interesting in 2019 and has been quiet since.
SmartDataCollective has covered data and analytics since 2008. This article rests on published documentation, filings and recorded talks. We have no access to these organisations’ infrastructure, their bills or their internal post-mortems, and nothing here is based on hands-on testing.
10 companies that use cloud computing, at a glance
| Organisation | Sector | What they moved | Platform | What changed |
|---|---|---|---|---|
| Netflix | Streaming | Everything: billing, encoding, playback control | AWS, plus its own Open Connect CDN hardware | Rebuilt as distributed services after a 2008 database failure |
| Capital One | Retail banking | All applications; last data centre closed in 2020 | AWS | Exited its own data centres. Also breached in 2019 through its own misconfiguration |
| Spotify | Audio streaming | Backend services and the entire data platform | Google Cloud | Retired self-run Hadoop for managed data services |
| Dropbox | File storage | Most user file data, off S3 and onto its own hardware | AWS, then in-house “Magic Pocket” | Cost of revenue fell, per its own S-1 |
| 37signals | SaaS | Applications, then object storage, back to owned servers | Left AWS and Google Cloud for co-located hardware | Publishes its own savings figures. Unaudited |
| Nasdaq | Financial markets | One options market’s matching engine, first | AWS Outposts, sited in Nasdaq’s own facility | Cloud operating model without moving the hardware off-site |
| Moderna | Biotech | Nothing. It never built a data centre | AWS | Research and manufacturing systems designed cloud-first |
| Monzo | Banking (UK) | Built on cloud from day one | AWS with Kubernetes | A regulated bank running over a thousand services, no data centre |
| The Guardian | News media | Publishing platform and content API; self-managed database to managed | AWS | Fewer systems the team has to keep alive overnight |
| Zoom | Communications | Burst capacity layered on top of its own data centres | Own facilities, AWS, Oracle Cloud | Absorbed the 2020 surge without owning peak-sized capacity |
1. Netflix — rebuilt itself on AWS after a database took the business down
The situation
In August 2008, database corruption in Netflix’s own data centre stopped it shipping DVDs for three days. Netflix has been open about this being the trigger. The lesson its engineers drew wasn’t “our database vendor let us down”. It was that vertically scaled hardware in a single facility puts the whole company on one failure domain.
What they did
An eight-year migration to AWS, finished in January 2016 when the billing systems finally moved. Netflix’s Yury Izrailevsky wrote at the time that the company deliberately refused a lift-and-shift, choosing instead to rewrite services as loosely coupled components running across multiple availability zones. The Simian Army, Chaos Monkey and the rest of the open-source tooling came out of that period, as did the practice of killing production instances on purpose to prove failover worked.
What it changed
Netflix has written that its cost per streaming start ended up a fraction of what it had been in its own data centres, and that service availability improved. It did not publish the underlying numbers, and you should treat the cost claim as the company’s own characterisation. One detail gets skipped in most retellings: Netflix does not serve video from AWS. Open Connect appliances, its own hardware, sit inside ISP networks. Control plane in the cloud, bytes on owned kit.
What transfers
The multi-region active-active architecture doesn’t transfer to a 40-person company, and copying it will cost you more in engineering time than any outage it prevents. What does transfer is the split. Netflix put the elastic, bursty part in the cloud and kept the predictable, bandwidth-heavy part on hardware it controls. That decision is available at any size.
2. Capital One — a large US bank that closed every data centre it owned
The situation
Eight data centres, a release cadence measured in months, and a competitive problem: the fintechs Capital One was losing customers to shipped weekly. Banks are among the most cautious organisations in the world about infrastructure, and for good reason. Many companies delay cloud migration over security concerns, and regulated ones delay longest.
What they did
A staged migration to AWS running from around 2014 to 2020, ending with the closure of its final data centre. Along the way the bank built and open-sourced Cloud Custodian, a policy-as-code engine for enforcing rules on cloud resources, now a CNCF project. That artefact matters more than the migration timeline: it tells you the bank’s real problem was governance at scale, not compute.
What it changed
Capital One has described faster provisioning and shorter release cycles, and has said publicly it was the first large US bank to go all-in. It has not published a consolidated cost figure, which is unsurprising for a public company that would then have to keep publishing it.
More Read
The other half of the story belongs here too. In July 2019, an attacker exploited a misconfigured web application firewall to reach data in the bank’s S3 storage, affecting roughly 100 million US applicants and customers by Capital One’s own disclosure. The OCC issued an $80 million civil penalty. The cloud provider was not breached. The customer’s configuration was. That’s what shared responsibility means in practice, and it’s why resilience strategies for cloud security issues belong in the migration plan rather than after it.
What transfers
The policy-as-code habit. If your cloud estate is more than a few dozen resources, humans reviewing configurations will miss things. The full data-centre exit does not transfer to anyone without Capital One’s engineering budget, and it took them roughly six years with that budget.
3. Spotify — left its own hardware because it wanted Google’s data tools
The situation
Spotify ran its own data centres and its own Hadoop cluster. Its infrastructure team spent its time racking servers and babysitting a batch platform, which is not what a music company’s engineers are for. Nicholas Harteau, then VP of engineering, wrote in 2016 that operating your own infrastructure had stopped being a differentiator.
What they did
A full move to Google Cloud, announced in early 2016 and described by the company as taking around two years. By Spotify’s own accounts it involved well over a thousand services and tens of thousands of daily data jobs. The data platform migration was the harder half: self-managed Hadoop gave way to BigQuery, Dataflow and Pub/Sub, with the Scio library open-sourced to make Scala pipelines workable on Dataflow.
What it changed
Spotify has consistently framed the payoff in terms of what its engineers stopped doing, not in dollars. It has not published migration savings. What it has published is Backstage, the developer portal built to keep a large service estate navigable, which is now used well beyond Spotify.
What transfers
The reasoning does. Spotify didn’t move for cost; it moved because a managed analytics stack let a small data platform team support a large engineering organisation. If your Hadoop or Kafka cluster is being maintained by two people who also have other jobs, that’s the same argument at a different scale. Comparable stories exist further down the size curve, including Evernote’s move to Google Cloud, done by a far smaller team over a far shorter window.
4. Dropbox — moved most of its data off AWS and built its own storage
The situation
Dropbox stored user file data in Amazon S3 while running its metadata and application layer itself. As the company grew, storage stopped being an undifferentiated commodity and became the thing the business actually was. At that volume, a per-gigabyte price you don’t control is a structural problem.
What they did
Magic Pocket: a custom multi-exabyte storage system running on Dropbox’s own hardware in leased data centres, built and filled over roughly two and a half years to 2016. Dropbox’s engineering team documented the design publicly, including the decision to write its own storage layer rather than adapt an existing one. The company kept AWS in the mix for regions where it had no infrastructure of its own.
What it changed
Dropbox’s 2018 S-1 attributed a cumulative saving of roughly $75 million across 2016 and 2017 to infrastructure optimisation, and its gross margin improved over that period. That’s a filing, not a blog post, which makes it the best-evidenced number in this article. It also came with the obligation to run hardware forever.
What transfers
Almost none of it, and that’s the point. Dropbox repatriated because storage was its product and its volume justified an entire in-house storage engineering team. If you’re comparing cloud storage solutions for business at anything short of exabyte scale, the S3 bill is not your problem and building Magic Pocket would be an act of self-harm. The transferable idea is narrower: work out which line on your cloud bill is a fixed, predictable, permanent load, because that’s the only line worth owning.
5. 37signals — left the cloud loudly, and kept publishing the numbers
The situation
Basecamp and HEY have steady, predictable traffic. They don’t have a Black Friday. David Heinemeier Hansson’s argument, made in a series of posts starting in October 2022, was that cloud pricing is rational for spiky or uncertain workloads and irrational for a stable one, and that his company had been paying an elasticity premium it never used.
What they did
Bought servers, put them in two co-location facilities, and moved the applications off AWS and Google Cloud during 2023. Then went further: in 2024 and 2025 the company moved its object storage off S3 onto its own arrays, a much larger job than the compute migration, and wrote about the mechanics of both. It also open-sourced Kamal, the deployment tool built to make owned servers feel like a platform.
What it changed
37signals has published its hardware costs, its old cloud bills and a running savings total, and says its operations headcount did not grow. These are self-reported and unaudited figures from a company with a strong public position on the subject, and they should be read that way. The technical claims are checkable in the code they’ve released. The financial ones aren’t.
What transfers
The diagnostic question, which is short: is your load predictable? If demand doesn’t move much month to month, you’re paying for optionality you never exercise. What doesn’t transfer is the assumption that you have people who can run hardware. 37signals had a capable ops team before it left, and hiring one specifically to save on cloud is how the arithmetic quietly reverses.
6. Nasdaq — put a market on AWS without moving the machines
The situation
Matching engines are latency-sensitive in a way almost nothing else is. Nasdaq wanted the operating model of cloud, including elastic capacity and managed services, without adding network hops between the exchange and the firms co-located next to it. Those two goals normally fight each other.
What they did
In 2021 Nasdaq announced a multi-year partnership with AWS, and in late 2022 migrated one of its US options markets, MRX, onto AWS Outposts installed inside Nasdaq’s own data centre in Carteret, New Jersey. The hardware is AWS-managed and AWS-shaped; it just doesn’t sit in an AWS region. Adena Friedman has described the programme publicly as phased, market by market, rather than a single cutover.
What it changed
Nasdaq has described the migration as successful and has continued the programme. It has not published latency comparisons, which is exactly what you’d expect from an exchange operator whose members would litigate the methodology. Take the absence of numbers as normal for the sector, not as evasion.
What transfers
Edge deployments of cloud platforms are now a mainstream option rather than a curiosity, and they exist for a specific class of problem: physical location matters, but you’d still like the cloud’s API and its managed services. Factory floors, hospitals and trading venues all qualify. If your constraint is data residency rather than microseconds, a region in the right country solves it more cheaply than an Outpost.
7. Moderna — a pharmaceutical company that never built a data centre
The situation
Moderna was founded in 2010, well after AWS was a serious option, and made a deliberate decision not to own infrastructure. mRNA design is computational before it is chemical, so the research bottleneck sits in modelling and lab automation rather than in glassware.
What they did
Built its research and manufacturing systems on AWS from the start, including its Drug Design Studio, which generates candidate mRNA sequences for protein targets. Moderna’s CIO has described the approach in public talks: standardised, automated processes designed once and then run repeatedly, with the digital layer treated as part of the manufacturing process rather than as support for it.
What it changed
Moderna has said publicly that the first clinical batch of its COVID-19 vaccine candidate shipped to the NIH 42 days after the viral sequence was selected. That figure is the company’s own and is widely cited. It should not be attributed to cloud computing alone: regulatory posture, prior platform work and a great deal of money were all in play. What’s fair to say is that Moderna had no procurement cycle for compute standing between a decision and a run.
What transfers
Greenfield organisations in regulated sectors no longer have to justify cloud; they have to justify hardware. Established ones don’t get to be Moderna, because Moderna’s advantage was the absence of legacy. The narrower lesson holds across healthcare, where cloud computing has changed patient engagement for organisations that never had the option of building their own platform.
8. Monzo — a licensed bank with no data centre, running on Kubernetes
The situation
Monzo applied for a UK banking licence with the intention of running entirely on public cloud, at a point when the standard core banking system was a vendor package on hardware in a bank-owned facility. UK regulators had published guidance on outsourcing to the cloud, so the path existed. Nobody had walked it at this scale.
What they did
Built a microservice platform on AWS with Kubernetes and Cassandra, and wrote about it in unusual detail. Monzo has published its service counts (well over a thousand, growing every year), its network isolation approach for controlling which services can call which, and full post-mortems for its outages, including the ones that were its own fault.
What it changed
It demonstrated that a deposit-taking, regulated bank can run without owning a server, which changed the conversation for every incumbent bank’s architecture team. Monzo hasn’t published infrastructure cost comparisons, and it has no legacy estate to compare against.
What transfers
The engineering documentation is genuinely useful reading if you’re building a service platform, and it’s free. What doesn’t transfer is the count. A thousand microservices for a company of Monzo’s size only works because they invested heavily in the tooling that makes creating one cheap. Attempting that ratio without the platform team produces a distributed monolith and a pager rota nobody wants. The wider point about the benefits and risks of cloud computing applies with force in regulated sectors: the regulator will hold you responsible for your provider’s failure as if it were yours.
9. The Guardian — got out of the database business
The situation
The Guardian ran its own MongoDB cluster on EC2 for the content management system behind its publishing platform, managed with Ops Manager. Its developers wrote candidly about what that involved: OS patching, cluster upgrades and enough operational overhead that the team was administering a database instead of shipping the tools journalists needed.
What they did
Migrated the CMS datastore to PostgreSQL on AWS RDS, documented in a 2018 engineering post that walked through the whole thing, including a period of writing to both databases in parallel and reconciling them before the cutover. The Guardian’s digital team has published extensively about running on AWS more broadly.
What it changed
Less maintenance, fewer components the team owned outright, and a database whose failover behaviour was the provider’s problem. The Guardian described the outcome in operational terms and didn’t publish a cost figure. Given the size of the estate, the saving was probably in engineer-hours rather than in the bill, which is the more honest way to describe most managed-service migrations anyway.
What transfers
This one transfers almost completely, and it’s the most useful case study on this page for a mid-sized team. Self-managed databases on cloud VMs are the single most common form of unfinished migration: you’ve taken on cloud pricing and kept all the operational work. The dual-write-and-reconcile pattern for a zero-downtime cutover is also worth stealing, because it’s the only version of this that lets you back out.
10. Zoom — owned data centres, with public cloud as the shock absorber
The situation
Zoom ran its own co-located data centres for the bulk of its meeting traffic, on the same logic as 37signals: known load, owned hardware. Then in March and April 2020 daily meeting participants went up by an order of magnitude in weeks. No procurement process moves at that speed.
What they did
Added capacity across public cloud providers while keeping its own facilities running. Oracle announced in April 2020 that Zoom had selected Oracle Cloud Infrastructure, stating that Zoom was moving around seven petabytes of data a day through OCI servers. That figure comes from Oracle, not Zoom, and it’s the sort of number a provider has every incentive to publish. Zoom also expanded on AWS during the same period.
What it changed
Zoom absorbed a demand curve that would have broken a hardware-only company. The architecture that resulted is the interesting bit: a stable baseline on owned kit, with public cloud carrying the part of the curve nobody could have forecast.
What transfers
The hybrid split, if you can engineer for it. Running the same workload in two places demands portability you have to design for deliberately, usually containers with no provider-specific services in the hot path. Most organisations discover during an incident that they can’t actually shift load, because six managed services quietly became load-bearing. Test the failover before you need it, or you own an architecture diagram rather than a capability.
The patterns worth copying
Read the ten together and a few things recur.
A constraint forced the move, not a strategy deck. Netflix had a three-day outage. Dropbox had a storage bill that scaled with its product. Zoom had a pandemic. None of these started with “cloud-first” as a principle. If you can’t name the specific thing that hurts today, you’re building a business case for going to the cloud out of vocabulary rather than pain.
The successful migrations rewrote things; the disappointing ones copied them. Netflix said outright that it refused a lift-and-shift. Spotify replaced its data platform rather than hosting the old one. Moving a virtual machine to a more expensive virtual machine changes your invoice and nothing else.
The direction of travel isn’t uniform. Two of these companies moved substantial workloads back onto their own hardware and both were right to. Predictable load with high volume is the profile where ownership wins.
Managed services are the actual product. Spotify wanted BigQuery. The Guardian wanted RDS. Almost nobody in this list moved because EC2 was cheaper than a server. They moved to stop employing people to do undifferentiated work, which also means they signed up for a skills shift: cloud platform engineering is a different discipline from data-centre operations, and the demand for cloud training tracks that shift closely.
The ones with the best outcomes published their failures. Monzo’s post-mortems, Netflix’s chaos engineering write-ups, the Guardian’s account of what its Mongo cluster cost in attention. Publishing forces precision.
Where this goes wrong
A page that shows ten wins and no reversals is an advertisement. Here’s the other side, and it’s better documented than the industry likes to admit.
The bill goes up and nobody can explain why. Data egress, idle non-production environments running at the weekend, over-provisioned instances chosen by a nervous engineer in month one and never revisited, per-request charges on a chatty service. Andreessen Horowitz’s 2021 essay “The Cost of Cloud, a Trillion Dollar Paradox” made the case that at sufficient scale, cloud spend materially depresses gross margins for public software companies. The venture firm has an interest in that argument; the arithmetic in it still stands up.
Lift-and-shift with no follow-through. The workload runs, the data centre lease ends, and you’ve swapped capital expenditure for a variable operating cost while keeping every architectural constraint you had before. Nobody publishes this. It’s the most common outcome.
Misconfiguration, which is now the dominant breach pattern. Capital One’s 2019 incident wasn’t an AWS failure. It was a firewall rule. The essential cloud security practices are unglamorous and they are almost always the gap: least-privilege roles, no public storage buckets, key rotation, logging that someone reads.
Committed spend you can’t grow into. Multi-year reserved capacity discounts look excellent in the negotiation and awkward in month fourteen when the roadmap changes. The same pattern shows up in SaaS procurement, where buyers routinely overpay for AI-driven services bundled into contracts they signed for something else.
Exit costs, which are improving. Egress fees were a genuine barrier to leaving. Following the EU Data Act, Google, AWS and Microsoft all removed egress charges for customers migrating away from their platforms during 2024. That helps with the bytes. It does nothing about the managed services you rewrote your application around, which is where lock-in actually lives.
Frequently asked questions about companies that use cloud computing
Which companies are known for using cloud computing?
Netflix, Capital One, Spotify, Airbnb, Monzo, Moderna, Nasdaq, Zoom and the Guardian all run production systems on public cloud and have described those systems publicly. Netflix and Capital One are the most-cited because both exited their own data centres entirely and both published enough detail to be useful. Dropbox and 37signals are worth knowing for the opposite reason: both moved significant workloads back to hardware they own.
What are the top 10 companies in cloud computing?
That question splits two ways, which is why search results for it are such a mess. If you mean providers, the usual list is Amazon Web Services, Microsoft Azure, Google Cloud, Alibaba Cloud, Oracle, IBM, Tencent Cloud, Huawei Cloud, Salesforce and Snowflake, with the ordering depending entirely on whether the tracker counts SaaS revenue. If you mean companies that use cloud computing at scale, the ten case studies above are a better answer, because they show what was moved rather than who sold it.
Does Netflix use cloud computing?
Yes. Netflix completed its migration to AWS in January 2016, after roughly eight years, and runs its control plane, encoding pipeline and billing there. It does not stream video from AWS: video is served from Open Connect, Netflix’s own CDN hardware installed inside internet service provider networks. The distinction matters, because Netflix is frequently cited as an all-in cloud company when it’s a well-designed hybrid.
Who is the biggest cloud computing company?
Amazon Web Services has consistently held the largest share of cloud infrastructure services in the estimates published by Synergy Research and Canalys, with Microsoft second and Google third. Microsoft is larger by total cloud revenue if you include Microsoft 365 and other software, which is why you’ll see two different answers depending on the source. The published percentages differ between trackers because they define the market differently, so treat any single figure with caution.
What is an example of a cloud computing case study?
Dropbox is the strongest one available, because the evidence is in an SEC filing rather than a blog post. Dropbox moved the majority of user file data off Amazon S3 onto its own storage system between roughly 2013 and 2016, and its 2018 S-1 attributed around $75 million in cumulative savings across 2016 and 2017 to that infrastructure work. A good cloud computing case study names the constraint, states what was actually migrated, and cites a source the company is legally accountable for.
How many companies use cloud computing?
Published adoption figures range from around 90% to nearly 100%, and the spread tells you the question isn’t well formed. If “uses cloud computing” includes any organisation with a Microsoft 365 subscription, the answer approaches everybody. If it means running production infrastructure on IaaS, the number is far lower and varies enormously by sector and country. We don’t hold survey data of our own and won’t quote someone else’s percentage as though it were settled. Check the methodology behind any adoption statistic before you put it in a board pack, specifically what the researchers counted as cloud.
