"Let's put everything on AWS" sounded in 2018 like the obvious move. In 2026 there are companies paying US$ 12,000/month for infrastructure that would cost US$ 3,000/year amortized on their own rack plus US$ 200/month in power and maintenance. And the other way: there are companies with physical servers crashing once a month because there's no redundancy and no on-call staff, when a well-planned cloud migration would solve it for a fraction of total cost.
There's no single answer. There are criteria — and this article is about which.
What each thing means
On-premise
Physical servers in your office or rented data center (colocation). You buy, install, maintain and replace the hardware. You pay power, AC, connectivity and the tech who looks after it.
Cloud (IaaS / PaaS / SaaS)
You rent virtualized infrastructure from third-party data centers (AWS, Azure, Google Cloud, DigitalOcean, Hetzner, Vultr). You pay monthly usage; the provider handles the hardware. Spans from "a VM with CPU/RAM/disk" (IaaS) to "a complete managed database service" (PaaS) to "a fully-managed software" like Microsoft 365, Salesforce (SaaS).
Hybrid
Mix of both — sensitive data or heavy workloads local, frontends and spikes in cloud. Connected via VPN or dedicated lines.
CAPEX vs OPEX: the most important thing for an SMB
On-premise is CAPEX (capital expense): pay big once, amortize over 3–5 years. Cloud is OPEX (operating expense): pay month to month as long as you use it.
This isn't just an accounting nicety. It has real implications:
- On-premise means small usage increments don't raise cost (hardware paid). But it traps you: need double tomorrow, you don't have it.
- Cloud scales linearly with usage. Good when you grow; bad when that 2-hour-per-month spike gets billed against all 730 hours.
- The average client does NOT need elastic scalability. An SMB with 30 employees using the same CRM every day doesn't benefit from "elasticity" — they pay the fixed cost under OPEX plus the risk of vendor price increases.
Side-by-side comparison
| Criterion | On-premise | Cloud |
|---|---|---|
| Cost model | CAPEX (high upfront, low after) | OPEX (recurring monthly) |
| Provisioning time | Days to weeks | Minutes |
| Scalability | Limited by purchased hardware | Practically unlimited |
| Physical maintenance | Yours | Provider's |
| Data control | Total | Shared (contract-dependent) |
| Latency for local users | Very low | Variable (better with nearby regions) |
| Availability | Depends on your team and power | Typically 99.9–99.99% SLA |
| Disaster recovery | Requires extra infrastructure | Simpler (regions, snapshots) |
| Personal data compliance (Colombia) | Simpler when data stays local | Requires analysis (Law 1581 cross-border) |
| Lock-in | Minimal (commodity hardware) | High (proprietary services/formats) |
| 5-year cost (stable load) | Usually lower | Usually higher |
| 5-year cost (variable load) | Usually higher | Usually lower |
| Required staff | More (sysadmin) | Less (devops / cloud engineer) |
Cases where on-premise wins
1. Heavy, constant workloads
A design firm rendering 24/7, a 50 TB file server accessed all day, a high-throughput transactional DB — in these cases, the 3–5 year math usually says buy hardware cheaper than rent cloud.
2. Highly sensitive or regulated data
Patient information (health), regulated financial data, strategic IP. When law requires data on Colombian soil or the reputation risk of a leak is huge, keeping it local with strict controls is simpler than negotiating clauses with a hyperscaler.
3. Latency-critical applications
Industrial apps (SCADA, plant control), collaborative CAD/CAM, real-time video editing — latency to a regional cloud (even "São Paulo") can be unacceptable. On-premise solves latency by proximity.
4. Predictable workloads without spikes
If your compute usage is flat all year, cloud elasticity adds no value — just cost.
5. Unreliable connectivity
Production plants in unstable-internet zones, warehouses in remote regions. If depending on cloud means operations halt when internet drops, on-premise is the only serious option.
Cases where cloud wins
1. Startups and new projects without traction yet
No sense investing US$ 10,000 in hardware for an idea you don't know works. Cloud lets you start at US$ 50/month and grow when the business grows.
2. Seasonal or unpredictable spikes
E-commerce with Black Friday peaks, event sites, campaign tools, monthly/yearly close processes. Paying peak capacity year-round in hardware is waste.
3. Teams distributed across multiple locations
If your team is in Cartagena, Medellín and Bogotá hitting the same systems, regional cloud gives reasonable latency to all without complex site-to-site VPN setups.
4. Services where 99.9%+ availability is critical from day 1
Replicating 99.99% redundancy on-premise (duplicate servers, redundant power, two ISPs) costs a fortune. AWS RDS Multi-AZ gives it to you built-in.
5. SaaS you don't want to administer
Microsoft 365 email, Google Workspace, Salesforce, HubSpot, GitHub. Building and maintaining on-premise equivalents is absurd for an SMB. Not even a discussion — use the SaaS.
The big cloud misconception
Many believe "cloud is cheaper". It's cheaper for variable or small use cases. It's more expensive for large stable workloads. AWS, Azure and Google make money precisely because per-unit compute price is meaningfully higher than amortized cost on owned hardware — their margin comes from there, plus the managed services.
There's a clear pattern today: companies that went 100% cloud 5 years ago are "repatriating" heavy workloads back to owned infrastructure or colocation. Dropbox did it, 37signals went public with numbers (US$ 7M/year saved). It doesn't mean cloud is bad — it means "all to cloud" without analysis is naive.
Hybrid architecture: what works best in SMBs
The pattern I most recommend to Colombian SMBs today:
- Email and collaboration: SaaS (Microsoft 365 / Google Workspace).
- Web and public services: Cloud IaaS or PaaS (Hostinger, DigitalOcean, Vercel for frontends).
- Offsite backup: Cloud storage (Backblaze B2, Wasabi).
- ERP / management system: Depends — if proprietary and critical, on-premise local with cloud backup. If third-party SaaS (SAP B1 Cloud, Siigo), cloud directly.
- Daily working files: On-premise NAS (Synology) synced to cloud.
- Heavy databases: On-premise with replication, cloud only if latency tolerates it.
How to decide well
- Measure real current load. CPU, RAM, disk, monthly traffic. Don't estimate, measure.
- Project 3 years. Estimated growth, not aspirational.
- Calculate TCO (Total Cost of Ownership) for both scenarios including: hardware/cloud, power, connectivity, software, licenses, technical staff, backups.
- Consider legal risks (Law 1581, customer contracts requiring data residency).
- Consider cost-of-exit: how much would it cost to leave the cloud in 2 years if you decide to move. That's real lock-in.
For most Colombian SMBs with 10–80 employees, the hybrid model (email and backup in cloud + management system and files on-premise) usually gives the best total cost and best control. But each case gets evaluated with numbers, not dogma.
Got a decision to make?
If you're evaluating cloud migration, workload repatriation or deciding between buying servers or renting, tell me what you have today and where you want to be and we'll build the numbers to decide on facts.