Almost every managed versus self-hosted debate turns into an argument about the monthly bill, and the monthly bill is the least useful number in the comparison. Both paths got better in the last two years. Managed platforms picked up compliance certifications, data processing agreements and regional deployment options. Self-hosting picked up one-click deploy scripts, pre-configured containers and open-weight models that hold their own on plenty of task categories. That makes this a question of organizational fit rather than technical feasibility, and the factors that settle it are more specific than most teams expect.
Where The Cost Crossover Really Sits
Managed pricing is boring in the best way. A platform subscription runs roughly $14 to $55 a month for small to mid-size workloads and $200 to $500 for higher-volume enterprise usage. Model inference adds $50 to $500 a month depending on tier and request volume. A moderate deployment lands somewhere between $100 and $800 a month, all in.
Self-hosted infrastructure starts lower and carries multipliers. A VPS that can run an orchestration layer is $5 to $40 a month. Add local inference and you are at $200 to $1,000 a month on cloud GPU, or $5,000 to $30,000 upfront for owned hardware. Storage, networking, backup and monitoring add another $20 to $200.
The crossover lands around 200 agent requests a day on mid-tier models. Below that, managed is almost always cheaper once engineering time is counted honestly. Above it, self-hosted starts showing advantages that grow with scale, and at enterprise volumes on open-weight models it can run 60 to 70 percent cheaper.
The Expenses Nobody Puts In The Spreadsheet
Engineering time is where the comparison gets uncomfortable. A standard self-hosted deployment needs 2 to 4 hours a month of real attention for routine maintenance, patching and monitoring, which is $100 to $400 a month in labor at typical salaries. Deployments with custom models, fine-tuning pipelines or multi-node architectures need 8 to 20 hours a month, which is $400 to $2,000.
Incident response is the other one. When a managed platform has an outage, the provider handles it and you wait. When your own deployment goes down at 3 AM, your team handles it. On-call rotations, emergency debugging and production fire drills are real costs and they belong in the analysis.
When Compliance Removes The Choice
Compliance is the strongest forcing function in the whole decision. Organizations under GDPR, HIPAA, DORA, the EU AI Act, or sector rules like FedRAMP or ITAR often face hard requirements about where data is processed, who can access it, and what audit trails must exist. Some of those effectively mandate self-hosting, because no managed provider meets the specific residency or access control requirement. Others are fully satisfied by a managed provider holding the right certifications.
The security half of that is worth stating plainly, because assumptions cause damage in both directions. Self-hosting is not automatically more secure just because you control the box. Managed providers get real economies of scale on security: dedicated teams, continuous vulnerability scanning, SOC 2 and ISO 27001, and fleet-wide patches deployed within hours of a disclosure. What self-hosting buys you is control over where data lives and who can reach it, which is a different property than being harder to compromise.
The Hybrid Most Teams Are Already Running
The binary framing is the part that misleads. Most self-hosted deployments still call a commercial API for every inference, which makes them self-managed orchestration sitting on top of managed inference. That is a hybrid, and budgeting for it as though it were a true self-hosted build is where the numbers go wrong. A genuinely self-hosted deployment running only local models removes the external dependency entirely, but it needs substantial GPU hardware and real ML engineering depth.
Knowing which point on that spectrum you actually occupy is the first honest step in the decision.
The Takeaway
Run the volume test first, because it settles the cost question quickly: under roughly 200 agent requests a day, managed wins on total cost once engineering hours are priced in. Then check compliance, because it can override the cost answer outright. Then be honest about which hybrid you are really running. If you want the long version, with all five decision factors, both cost models and the cases where each side clearly wins, it is here: Managed vs Self-Hosted AI Agents: Decision Guide
Top comments (0)