INFRASTRUCTURE
Cloud, on-prem, hybrid: the boring decision framework.
There's no universally right answer to where workloads should live. There is a small set of questions that gets you to the right answer for any specific workload.
The cloud-vs-on-prem debate gets framed as a religious one. It isn't. It's a per-workload decision driven by four variables: data gravity, latency requirements, cost shape, and operational maturity. Get those right and the answer falls out.
Data gravity
Wherever the data lives, the compute wants to be nearby. If you have terabytes of imaging, video, or sensor data being generated on-site, pulling it to the cloud for processing and pushing results back is slow and expensive. Process locally, sync summaries up.
If your data is born in SaaS systems already — CRM, billing, support — it's already in the cloud. Building on-prem analytics infrastructure to consume it is moving the data the wrong direction.
Latency
Sub-50ms interactive workloads (point-of-sale, manufacturing control, real-time vision) need to run physically close to the user or the machine. Round-tripping to a region 800 miles away will not feel right no matter how fast the link is.
Batch and asynchronous workloads have no latency budget worth defending. Run them wherever they're cheapest.
Cost shape
Cloud is operating expense that scales with use. On-prem is capital expense plus a fixed operating overhead. The crossover point is roughly: if your workload runs at >70% utilization 24/7 for three years, on-prem usually wins on TCO. If it spikes and idles, cloud usually wins.
The trap is the workload that started spiky and quietly became steady-state. Re-evaluate annually.
Operational maturity
On-prem assumes you can run hardware: capacity planning, firmware, backups, replacement parts, after-hours response. If you don't have that muscle and don't want to build it, the cloud's higher per-unit cost is buying you operations, not just compute. That's a fair trade.
Hybrid is a strategy, not a leftover
Most real architectures end up hybrid: identity and collaboration in the cloud, line-of-business systems wherever the data and latency demand, with disciplined boundaries between them. Hybrid only goes wrong when it's the accidental result of two half-finished migrations rather than a deliberate split.