"Move everything to the cloud" was the default advice for a decade. More recently, some well-known companies have publicly moved workloads back to their own hardware and reported large savings. For an Indian business deciding where to run its ERP, website, file server or AI workloads, both camps sound convincing. The honest answer is that it depends on the workload, and you can work it out with a fairly simple cost model.
Why do cloud vs on-premise cost comparisons usually go wrong?
Most comparisons put a monthly cloud invoice next to the price of a server and declare a winner. Both sides of that comparison leave things out.
- On-premise comparisons forget power, UPS and battery replacement, cooling, the space, internet links, hardware warranty, spare parts, and above all the people who patch, monitor and fix the server at 2 a.m.
- Cloud comparisons forget data transfer (egress) charges, storage that grows quietly, snapshots never deleted, over-sized instances nobody revisits, premium support plans, and the engineering time needed to manage cloud accounts properly.
A total cost model you can fill in
Compare over three to five years, which is roughly a hardware life cycle. Fill in your own numbers; local power tariffs, bandwidth prices and salaries vary too much across Indian cities for any generic figure to be reliable.
| Cost line | On-premise / own server | Public cloud |
|---|---|---|
| Compute | Server purchase spread over its life, plus warranty | Instance hours, or reserved / committed-use pricing |
| Storage | Disks, RAID, spares, backup hardware | Block and object storage per GB per month, snapshots |
| Power and cooling | Server and AC consumption at your tariff, UPS and batteries | Included |
| Space | Server room or colocation rack rental | Included |
| Network | Business internet lines, ideally two from different ISPs, static IPs | Data transfer out charges, load balancers, NAT gateways |
| People | Admin time for hardware, OS, patching, monitoring | Admin time for cloud accounts, security, cost control |
| Software licences | OS, database, virtualisation | Often bundled into instance price, sometimes higher |
| Disaster recovery | Second site or off-site backups | Second region or backups, plus their transfer costs |
| Downtime risk | Single site, hardware failure, power cuts | Provider outages, misconfiguration |
Add up each column over the period. Then do one more thing most people skip: model growth. If your storage doubles every year, the cloud line grows every month, while on-premise grows in steps when you buy new disks.
What does a real cloud exit look like?
One widely discussed recent example is 37signals, the company behind Basecamp and HEY. Its co-founder David Heinemeier Hansson has published the figures: a cloud bill at a run rate of about $3.2 million a year, about $700,000 spent on Dell servers (which he says was recouped during 2023 as cloud contracts expired), and a 2024 cloud bill of about $1.3 million, all of it storage that was also planned to move. He reports savings of almost $2 million a year and projects more than $10 million over five years, with the team running the infrastructure unchanged in size.
He is also candid about the limits: the cloud still makes sense for very early companies that are not sure they will survive the year or do not need a whole machine, and for workloads with large, unpredictable swings in load. 37signals fitted the new hardware into data-centre racks it already used and kept its existing operations team; a company starting from zero would have to pay for those.
The lesson for an Indian business is not "leave the cloud". It is: once your cloud bill is substantial and your workload is steady, run the numbers seriously.
When does the public cloud win?
- Unpredictable or spiky load: festive-season e-commerce, exam result days, marketing campaigns.
- Early-stage products where you do not yet know the demand, and speed matters more than unit cost.
- Global users who need low latency in several regions.
- Managed services you would struggle to run yourself: managed databases with automated failover, object storage with high durability.
- Small teams with no infrastructure skills, where the alternative is an unmaintained server in a cupboard.
When does on-premise or a private server still win?
- Steady, predictable workloads that run all day, every day: ERP, internal applications, file servers, build servers.
- Large data volumes that move a lot: CCTV footage, design files, backups, media. Storage and egress charges dominate cloud bills here.
- GPU-heavy AI workloads running continuously, where renting GPUs by the hour adds up quickly.
- Data that must stay under your direct control for client, contractual or regulatory reasons. Note that CERT-In's 2022 directions already require logs of ICT systems to be maintained within Indian jurisdiction for a rolling 180 days, whatever hosting you use.
- Factories and sites with poor connectivity, where local systems must keep running when the internet does not.
A worked way to read your own numbers
Take one workload, say your ERP. Pull the last twelve months of its cloud invoices, split into compute, storage, data transfer and support. Then get two quotes: a rented private server of similar capacity, and the hardware, warranty and colocation or power cost of running your own. Add realistic admin hours to each option at your actual hourly cost. If the private options are cheaper by a wide margin and the workload has been steady for a year, it is worth a pilot. If the gap is small, stay where you are and optimise instead; migrations carry their own risk and cost.
The middle options most businesses overlook
- Dedicated or virtual private servers from a hosting provider: predictable monthly cost, no hardware to own, often cheaper than equivalent hyperscaler instances for steady workloads, though you should compare quotes.
- Colocation: your own servers in a professional data centre with proper power, cooling and connectivity.
- Hybrid: steady workloads on your own or rented servers, spiky or global workloads in the public cloud, backups in a second location.
Do not ignore the cost of leaving
Whatever you choose, keep the option to move. In March 2024 AWS announced credits for data transfer out to the internet for customers moving their data off AWS, following a similar move by Google, with reporting linking both to the European Data Act. That helps with transfer fees, but the bigger exit cost is usually engineering: applications built on provider-specific services. Containers, open-source databases and infrastructure as code keep that cost low.
A quick decision checklist
- Is the workload steady or spiky?
- How much data does it store and send out each month?
- Do you have, or can you hire or contract, people to run servers properly?
- Are there data residency or client requirements?
- What happens to the business if this system is down for four hours?
- What would it cost to move this workload again in three years?
How Cehpoint can help
We build the cost model with your real bills and usage, then design and run whichever mix makes sense: cloud, private servers or hybrid. See our cloud and DevOps service. If the trigger is a rising bill, start with a cloud and AI cost optimization review, and if you would rather own the server outright, look at our lifetime VPS option.