Most infrastructure teams understand single-turn LLM inference cost.
You send a prompt, the model responds, and you pay for tokens processed. The relationship between input size, model size, and cost is straightforward.
This mental model breaks down completely when systems evolve into **multi-step AI workflows**.
This post is part of a series on the economics of multi-step AI workflows. We examine why inference costs scale with depth, why verification is disabled in production, and why existing optimizations fail to eliminate redundant execution across workflow steps.
What Are Multi-Step AI Workflows?
Multi-step workflows—also called agentic workflows—are AI systems that reason in stages rather than single interactions.
A typical multi-step workflow might:
- Read and analyze a large document
- Form a plan based on that analysis
- Execute multiple subtasks sequentially
- Verify results at each stage
- Refine or retry based on outcomes
Each step builds on prior steps. The workflow maintains shared context across all stages.
This is fundamentally different from single-turn chat, where each interaction is independent.
Why LLM Inference Cost Scales Differently
In single-turn inference, cost scales with:
- Prompt length
- Model size
- Output length
In multi-step workflows, cost scales with:
- Workflow depth (number of steps)
- Shared context size
- Number of times that context is reprocessed
The critical difference: shared context gets reprocessed at every step.
If a workflow has ten steps and each step references the same 50,000-token context, that context is processed ten separate times—even though the information itself never changes.
Why Inference Optimization Alone Doesn't Fix It
Inference optimizations—faster runtimes, better batching, model quantization—improve individual inference calls.
They don't address redundant execution across sequential workflow steps.
These optimizations assume each call is independent. In multi-step workflows, steps are dependent and sequential. They can't be batched. They share context.
Faster inference makes each step faster, but it doesn't eliminate the redundant work of reprocessing the same context repeatedly.
Why Teams Move Toward Self-Hosted LLM Inference
As workflows deepen, API economics break down.
Cost becomes unpredictable because it scales with workflow depth, not just request volume. Execution is opaque—teams can't see what's being reprocessed. Optimization is constrained—redundant execution can't be eliminated from the outside.
Teams move to self-hosted LLM inference not just for control or data privacy, but because the cost structure changes fundamentally.
Self-hosting provides:
- Predictable cost structure (fixed infrastructure vs variable per-call)
- Execution visibility (teams can see exactly where compute is going)
- Optimization control (redundant work can be eliminated, not just made faster)
The Structural Problem
The root issue is that multi-step workflows treat every step as stateless, even when the state hasn't changed.
Modern AI systems repeatedly re-encode the same information because there is no standard way to continue execution from prior internal state.
Until that changes:
- Deep workflows remain fragile
- Verification remains optional
- Costs scale faster than value
Understanding the Full Picture
To fully understand why LLM inference cost explodes in multi-step workflows, consider:
- Why inference costs scale with workflow depth — The fundamental cost structure shift
- The hidden cost of reprocessing context — How shared context multiplies cost
- Why verifier loops get turned off — How cost structure forces quality tradeoffs
- Inference optimization vs execution efficiency — Why existing optimizations plateau
- When API economics break — Why teams move to self-hosted inference
These pieces together explain why multi-step AI workflows create fundamentally different cost structures—and why infrastructure teams need different optimization strategies.
CLC Labs is focused on execution-layer infrastructure that addresses the cost structure of multi-step AI workflows.