Unpopular opinion: the training and fine-tuning costs make Claw a bad fit for SMBs.

I've been seeing a lot of excitement around Claw, especially for automating customer support and internal documentation. It's a powerful model, no doubt. But after running some numbers for a few small-to-midsize business (SMB) clients wanting to implement it, I keep hitting the same wall: the ongoing training and fine-tuning costs are a silent budget killer for operations at this scale.

Let's break down what a typical SMB implementation might look like. You're not just paying for API calls. To get good, domain-specific results, you need to fine-tune the base model with your own data. This process has three major cost components:

1. **Data preparation & cleaning:** Engineering hours to structure support tickets, knowledge base articles, etc.
2. **Fine-tuning compute costs:** The actual GPU time to run the training jobs on your data. A single run on a substantial dataset can easily run into hundreds of dollars.
3. **Iteration and evaluation:** You rarely get it right the first time. You'll need multiple cycles, each costing more in compute and engineering time.

Here's a simplified, hypothetical cost structure for a 6-month project I modeled for a client with ~50k historical documents:

```yaml
# Simplified Cost Breakdown (Hypothetical AWS SageMaker Example)
Fine-Tuning Phase (Months 1-2):
- Data Engineering (40 hrs @ $75/hr): $3,000
- 5x Training Jobs (ml.g5.12xlarge @ $9.83/hr, avg 8 hrs/job): ~$400
- Storage & Logs: ~$100

Monthly Inference & Maintenance (Months 3-6):
- API Endpoint (ml.g5.2xlarge @ $1.68/hr, 24/7): ~$1,200/month
- Monthly Data Refresh & Light Re-tuning (10 hrs eng + compute): ~$1,000/month
```

**Total estimated 6-month cost: ~$9,500.** And that's before accounting for the initial model selection, prompt engineering, and integration work. For an SMB, that's a significant ongoing investment just to *maintain* the system's accuracy as their business data changes.

For many SMB use cases, a well-architected combination of simpler, purpose-built tools often delivers 80% of the value for 20% of this cost. Think about using a rules-based classifier to route tickets, paired with a semantic search (like OpenAI embeddings or even open-source models) over your knowledge base. The TCO is far more predictable and manageable.

I'm curious if others have done similar deep-dives. Have you found a cost-effective way to operationalize these large models at SMB scale, or is the ROI just not there yet compared to more traditional ML/AI approaches?

-- Amy