Showcase: The spreadsheet I use to track total cost of ownership for AI agent platforms.

Another week, another VC-funded AI agent startup announcing their "enterprise-grade platform" with pricing that seems to be plucked from thin air. The discourse is always about tokens per second or benchmark scores, never about the actual bill you'll get when you deploy this thing beyond a demo. It's infrastructure, people. The costs are predictable if you bother to model them.

I've been tracking the TCO for these platforms because, surprise, the listed SaaS fee is often the *smallest* part of the equation. The real bleed is in the foundational model inference, vector DB operations, and the orchestration overhead they conveniently offload to your cloud bill. Most teams just see the $5/agent/month and think they've got a deal, then get a $12,000 AWS invoice and panic.

So, here's the skeleton of the spreadsheet I use. It's not pretty, but it forces you to confront the real variables. The key is modeling *concurrent usage*, not monthly averages. An agentic workflow isn't a background job; it's a chain of synchronous, latency-sensitive calls that scale with user activity.

```plaintext
| Component | Cost Driver | Formula Example (AWS) |
|-------------------------|--------------------------------------|---------------------------------------------------------------------------------------|
| Platform SaaS Fee | Per agent, per user, etc. | =($PlatformFee * NumberOfAgents) |
| Foundation Model Inference | Input/Output tokens, concurrency | =((AvgInputTokens * $0.50/1M) + (AvgOutputTokens * $1.50/1M)) * AvgSessions/Month * CallsPerSession |
| Vector DB Read/Write | Query units, storage, writes | =((QueryUnits/Month * $0.25/1K) + (StorageGB * $0.50)) |
| Orchestration Compute | Runtime (vCPU-seconds), memory | =((TotalVCPUSeconds * $0.000044) + (TotalGBSeconds * $0.000005)) * PeakConcurrentUsers |
| "Hidden" Egress | Data transfer between services | =GBTransferredBetweenAZs * $0.01 |
| Development & Debug Ops | Logging, tracing, monitoring | =IngestionGB * $0.50 + ScanGB * $0.10 |
```

The brutal exercise is filling this in. You'll need to:
* Instrument a prototype to get *real* token counts per agentic session (they're 3-5x higher than a simple chat completion).
* Estimate your peak concurrent sessions to size the orchestration layer. Ten users hitting an agent at once requires ten parallel chains.
* Model vector DB costs based on your RAG strategy. A million queries a month isn't "high volume" anymore; it's a Tuesday.

I recently modeled a proposed "budget" platform against a DIY setup using OpenRouter, LiteLLM, and some carefully tuned spot fleets for orchestration. The SaaS platform's TCO came in at ~$8,400/month at projected scale. The DIY version, with reserved instances for the stable base and spots for the burstable agent workers, modeled to ~$3,100. That's a 63% premium for the convenience wrapper. For some, that's worth it. But most teams signing that contract have no idea they're paying it.

The point isn't that you should build it yourself. The point is you need to know *what you're buying*. These platforms are reselling compute and API calls with a markup. If you can't deconstruct their pricing into the underlying cloud services, you are not doing due diligence; you're writing a blank check and hoping the growth hackers were feeling generous this quarter.