Self-hosted Wan 2.1 on RunPod/Vast.ai for bulk generation

[rancur]

Problem

fal.ai charges $0.15-$0.25 per Wan 2.1 generation. Self-hosted on spot A100s costs $0.03-0.04 per generation -- 4-6x cheaper.

Benchmarks

GPU	480p (5s clip)	720p (5s clip)	Cost/hr (spot)	Cost/gen (480p)
A100 80GB	~170s	~523s	$0.78	$0.037
H100 SXM	~85s	~284s	$1.00	$0.024

Proposal

Create a self-hosted inference pipeline:

1. RunPod Serverless for burst workloads (pay per second, auto-scale to zero)
2. Vast.ai spot A100s for batch processing (cheapest bulk rate)
3. Retry logic for spot preemption
4. Batch queue with progress tracking

For 5,000 tracks x 3 video segments each = 15,000 generations:

• On H100 spot at 480p: 15,000 / 42 per hour = 357 GPU-hours = $357 total
• vs fal.ai: 15,000 x $0.15 = $2,250 total

Implementation

• New: src/generator/self_hosted.py -- RunPod/Vast.ai client
• src/generator/video_pipeline.py -- add "self-hosted" provider
• src/generator/model_router.py -- route to self-hosted for bulk jobs
• Docker image with Wan 2.1 + ComfyUI for deployment

Risk

• Spot instances can be preempted mid-generation (lost work)
• Need checkpointing or accept ~5% retry rate
• More DevOps overhead than API calls

Labels: cost-optimization, infrastructure