Orchestrator Selection and Pricing

How gateways select orchestrators for transcoding and AI inference jobs: stake-weighted selection, price filtering, and capability matching.

---

When a gateway receives a job (a transcoding segment or an AI inference request), it selects an orchestrator from the active set to process the work. Selection combines three criteria: capability matching (does the orchestrator support the requested pipeline and model?), price filtering (is the orchestrator’s price within the gateway’s maximum?), and stake-weighted random selection among qualifying candidates.

Capability Matching

For AI inference jobs, the gateway matches against orchestrators registered in the AI Service Registry on Arbitrum One. The registry records which pipelines and model IDs each orchestrator advertises as warm. A gateway requesting text-to-image with model ByteDance/SDXL-Lightning only considers orchestrators advertising that combination.

Capability data is read at session start; the gateway caches the active set and refreshes periodically. Cold capabilities (advertised but not currently loaded) are visible in the registry but may result in longer job latency as the orchestrator loads the model.

Price Filtering

Orchestrators advertise a price per unit. The gateway filters out any orchestrator whose price exceeds -maxPricePerUnit:

livepeer -broadcaster \
  -maxPricePerUnit 1000000000000000 \   # max 0.001 ETH/unit
  ...

For transcoding, the unit is wei per pixel. For AI batch pipelines, the unit is pipeline-specific (per output, per second, or per request depending on orchestrator configuration). Orchestrators advertising a price of 0 accept any work regardless of the gateway’s maximum.

Setting -maxPricePerUnit 0 disables price filtering: the gateway accepts orchestrators at any price.

Stake-Weighted Selection

Among orchestrators that pass capability and price filters, the gateway applies stake-weighted random selection. An orchestrator with twice the staked LPT of another has approximately twice the probability of being selected for a given job, holding other factors equal.

The discovery ping time also influences selection: orchestrators that consistently fail to respond or return errors are removed from the active selection pool until they recover.

This mechanism is stable for the duration of a stream: once a session is assigned to an orchestrator, the gateway continues routing that session’s segments to the same orchestrator unless it fails.

Restricting to Specific Orchestrators

From go-livepeer PR #3634, gateway clients can include or exclude specific orchestrators by serviceAddr in their requests:

# Include only specific orchestrators
livepeer -broadcaster \
  -orchAddr "https://orch-1.example.com,https://orch-2.example.com" \
  ...

This is useful for operators running their own orchestrators, for testing, or for agreements between gateway and orchestrator operators outside the open market.

Viewing Active Orchestrators

The Livepeer Explorer at explorer.livepeer.org shows the active orchestrator set with stake, price per pixel, and performance scores. The AI capability dashboard at tools.livepeer.cloud/ai/network-capabilities shows AI-specific warm capabilities.

Setting Prices as an Orchestrator

Orchestrators set pricing per pipeline type. For transcoding:

livepeer -orchestrator \
  -pricePerUnit 1000  # wei per pixel
  ...

For AI batch pipelines, pricing is set in aiModels.json per pipeline entry:

{
  "pipeline": "text-to-image",
  "model_id": "ByteDance/SDXL-Lightning",
  "price_per_unit": 5000000000000000,
  "warm": true
}

For live-video-to-video, price_per_unit is in wei per second.

Competitive pricing relative to the gateway’s -maxPricePerUnit determines whether an orchestrator enters the selection pool at all. Orchestrators priced above the gateway’s maximum receive no work from that gateway regardless of stake.

Market Dynamics

The active set posts prices on-chain via the AI Service Registry and via the off-chain GetOrchestratorInfo RPC. Gateways negotiate directly during session establishment.

The Livepeer Explorer shows historical price data per orchestrator. The tools.livepeer.cloud/ai/network-capabilities dashboard shows current warm AI capabilities but does not expose per-orchestrator pricing.

Related Pages

Probabilistic Micropayments

How tickets are priced and the per-unit model used in selection.

ETH Escrow and Deposits

Funding the gateway deposit that backs PM tickets sent to orchestrators.

BYOC Overview

Registering custom capabilities that influence orchestrator selection.

Alt-Gateway Overview

Off-chain gateway clients and their orchestrator discovery patterns.