1. What is the problem?

Today most AI pipelines on the network run using a one-container-per-pipeline setup.

In practice this means:

• Each pipeline type runs in its own container environment
• Orchestrators must maintain multiple containers to support multiple pipelines
• Most orchestrators end up specialising in one pipeline type
• GPUs cannot easily pivot between different workloads without redeployment

This creates two ecosystem issues:

Limited flexibility

Even if GPUs have available capacity, orchestrators cannot easily switch between different pipeline types.

Operational complexity

Running multiple containers increases configuration overhead, environment drift, and maintenance burden.

As a result, orchestrators often choose a single pipeline to support rather than experimenting with multiple AI services.

The constraint appears to be software orchestration, not GPU capability.

2. Why does this matter to the ecosystem?

This primarily affects the supply side of the network.

The ecosystem currently has roughly 100 AI-capable orchestrators, meaning supply growth is constrained.

When supply is capped, the key growth lever becomes:

revenue per GPU

Multi-pipeline capability could improve:

• revenue per GPU
• revenue per orchestrator
• supply flexibility across pipeline types
• time-to-serve new AI workloads

Without this flexibility, the network risks developing specialised supply that cannot adapt quickly to demand changes.

3. Proposed experiment

This proposal tests whether ComfyStream can enable multi-pipeline orchestrators by dynamically loading workflows on a single GPU.

Instead of running multiple containers, an orchestrator would run:

one ComfyStream runtime capable of loading different workflows on demand.

The experiment aims to determine whether this is technically viable and operationally useful.

4. Demonstration application: ComfyMeme

To test this capability in practice, the experiment includes building a small demonstration application called ComfyMeme.

ComfyMeme generates AI-remixed animated memes using short GIF / WebP clips from the Giphy API.

Example pipeline:

Giphy meme clip
→ frame extraction
→ Stable Diffusion + LoRA stylisation
→ animated meme output

Memes are intentionally chosen because they are:

• easy to understand
• fast to generate
• culturally shareable
• capable of producing organic traffic

The demo therefore acts as both:

• a public application
• a stress test for multi-pipeline orchestration

5. Relationship to ComfyStream Cloud

The longer-term concept discussed in the AI SPE roadmap is ComfyStream Cloud — a platform where creators could deploy Comfy workflows as hosted AI applications.

Instead of this model:

creator → workflow JSON → user runs locally

Workflows could become hosted services:

creator → workflow → hosted endpoint → users

ComfyMeme acts as a first proof-of-concept of this idea.

If hosted workflows prove viable, future work could explore creator deployment tools and monetisation mechanisms.

6. Scope limitations

This experiment intentionally avoids solving several large problems:

• automatic model distribution
• arbitrary workflow compatibility
• creator monetisation infrastructure

Instead, it assumes a curated model set preloaded on orchestrator nodes.

The goal is simply to validate dynamic workflow execution on the network.

7. Success criteria

Success should demonstrate both technical viability and economic potential.

Technical signals:

• one ComfyStream runtime successfully serving multiple workflows
• acceptable workflow switching latency
• stable execution across multiple requests

Adoption signals:

• at least three orchestrators running ComfyStream

Economic signal:

• at least one orchestrator earning revenue from two distinct pipeline types on a single GPU

8. Deliverables

The experiment would produce:

• ComfyMeme demonstration application
• ComfyStream configuration enabling multi-pipeline loading
• documentation for orchestrator setup
• public demo endpoint
• written findings on performance, latency, and operational challenges

9. Expected duration

Estimated timeline: 4–6 weeks

This includes:

• building the demo application
• orchestrator deployment testing
• community demonstration
• documentation of results

10. Summary

This proposal tests whether dynamic workflow loading via ComfyStream can enable orchestrators to serve multiple AI pipelines from a single GPU.

The experiment combines:

• infrastructure validation
• a public demonstration application
• measurable economic outcomes

The key outcome is simple:

prove that a single orchestrator GPU can earn revenue from multiple pipelines.

If successful, this would strengthen orchestrator incentives and lay the groundwork for future hosted workflow platforms such as ComfyStream Cloud.