The question posed by the Reddit thread cuts to the heart of a tension we see across modern data work: the desire for sovereignty versus the reality of operational burden. Our opinion is plain: self-hosted ML can deliver meaningful control, but only if your team is prepared to absorb the complexity it introduces. For most organizations, the trade-off is not a binary choice between freedom and constraint, it is a practical calculation about where your scarce engineering hours are best spent.
Control is real when it matters. If your use case demands data never leaves your network, requires custom model tuning that a hosted API forbids, or must comply with regulatory frameworks that treat third-party inference as a data transfer, then self-hosting is the only honest answer. In those scenarios, the alternative is not a simpler path; it is a dead end. The control you gain is control over latency, over data residency, over the model's behavior when you need to adjust it mid-flight. That is not theoretical. It is the difference between a tool that fits your constraints and one that forces you to reshape your workflow around its limitations.
But control is not free. The same thread that celebrates on-prem sovereignty also documents the hidden costs: infrastructure provisioning, model serving optimization, monitoring for drift, managing GPU utilization, handling security patches, and debugging failures that a managed service would absorb silently. Every hour your team spends keeping the self-hosted stack running is an hour not spent on the actual ML work that drives your product. For teams without dedicated ML infrastructure engineers, that shift can turn a perceived advantage into a persistent drag on velocity. The complexity does not disappear; it migrates from your vendor's operations to your own.
What this means for you is a straightforward audit. Ask: does your data or compliance posture require self-hosting, or does it just feel safer? If the answer is the former, invest in the team and tooling to make the complexity manageable, treat infrastructure as a first-class discipline, not an afterthought. If the answer is the latter, consider that a well-designed hosted solution, with clear data handling policies and auditability, may give you the control you actually need without the operational tax. The goal is not to avoid complexity for its own sake. It is to ensure the complexity you choose to own is complexity that serves your outcomes.