"We'll just deploy it on Kubernetes." Teams discussing AI agent infrastructure often hear this proposal. It sounds reasonable. An existing K8s setup makes one more container seem easy.
Many teams discover six months in that it is harder than expected. The cost model below covers self-hosting AI agents in the DACH market with real numbers.
The Visible Costs (What Everyone Budgets For)
These are the costs teams typically account for when planning AI agent infrastructure:
The visible bill looks manageable. The hidden costs change the calculation.
The Hidden Costs That Kill Budgets
1. Engineering Time to Build It
Deploying the container is only the start. The team also needs to build:
- •Webhook ingestion layer: Accept external events, validate signatures, handle retries
- •Queue consumers: Process messages reliably with error handling and dead-letter queues
- •Agent orchestration: Manage concurrent runs, timeouts, and state
- •Deployment pipelines: CI/CD for agent code with versioning and rollback
- •Observability stack: Logging, tracing, metrics, dashboards
Conservative estimate: 2–3 months of senior engineer time. Senior software engineers in Germany earn a median base salary in the €78k–€95k range per Levels.fyi Germany compensation data; with typical employer overhead (~30% on top of gross), fully loaded cost is €100k–€125k/year, or roughly €17k–€31k just to reach "it works."
2. Ongoing DevOps Burden
Infrastructure doesn't run itself. Someone needs to:
- →Monitor for outages and performance degradation
- →Apply security patches and updates
- →Scale infrastructure as usage grows
- →Debug production issues at 3am
- →Handle on-call rotations
This usually takes at least 20–30% of a DevOps engineer's time, which works out to €1,800–€3,100/month in ongoing labour cost at fully-loaded DACH DevOps engineer rates, often from the team's most expensive engineers.
3. Tools and Software
4. The Opportunity Cost
The hidden cost does not show up on any invoice: what could the engineers have built instead?
Every hour spent debugging Kubernetes networking is an hour not spent on product features that differentiate the business. Every sprint on "agent infrastructure improvements" is a sprint competitors can use to ship customer-facing features.
The 12-Month TCO Calculation
Adding it up for a realistic DACH-market scenario: a mid-size team running a moderate workload of 5,000 agent runs per month, average 30-second run duration, with cited DACH salary and EU cloud pricing assumptions.
How our pricing works in this model: the monthly subscription converts into an equal amount of Project credit. Usage draws down that credit at the published per-unit rates (€0.0337 per LLM-agent start, €0.00042 per runtime second, €0.00342/GB-hour of storage, and €0.0276/synced retrieval item). This example uses a one-month term and estimates 5,000 runs per month at base rates, with no term or volume discount. Volume discounts start at the 10,000-run package. At 5,000 LLM-agent runs/month averaging 30 seconds, run starts and runtime use €231.50/month: the €200 Pro credit plus €31.50 of purchased credit. This example excludes model tokens, storage, and retrieval usage. Extra credit can be added manually or through capped auto-refill. Enterprise contracts define custom commercial terms.
Why Existing Kubernetes Is Not Enough
The most common objection deserves a direct answer.
Kubernetes provides orchestration. AI agents still have unique requirements:
- •Unpredictable execution times: A simple query might take 2 seconds, a complex one might take 5 minutes
- •Token tracking: Cost control requires the exact token consumption of every run
- •Execution traces: Standard APM doesn't capture LLM reasoning steps or tool calls
- •Hot-reload development: Existing CI/CD is rarely built for 2-second iteration cycles
The result is a custom platform on top of the existing infrastructure, bringing the calculation back to the cost estimates above.
When Self-Hosting Does Make Sense
To be fair, there are legitimate reasons to self-host:
- 1.Extreme data sensitivity: Regulated industries where data cannot leave the organization's infrastructure (though even here, private cloud options exist)
- 2.Massive scale: At millions of agent invocations per day, the math might favor self-hosting
- 3.Core competency: Building expertise makes sense when AI infrastructure is the product
For teams where AI agents are a feature rather than the product, the math strongly favors managed platforms.
Validate Before Investing in Self-Hosting
Our recommendation: start with a managed platform and validate the use case before committing to infrastructure investment.
With our Basic plan, teams can:
- •Deploy agents in minutes, not months
- •Validate that the integration patterns work
- •Get real usage data to inform build-vs-buy decisions
- •Ship AI features while evaluating long-term options
If usage eventually justifies self-hosting, the required platform scope will be clear. Otherwise, as is true for most teams, the managed start avoids a very expensive learning experience.
Further detail is available in our guide to replacing self-hosted AI agents with a managed platform, our managed vs. self-hosted TCO comparison, our pricing page for the full cost comparison, and the quickstart guide.