runpod-spot-pricing

name	runpod-spot-pricing
description	Daily snapshot of RunPod community-cloud (spot) vs secure-cloud (on-demand) GPU pricing. Tracks A100/H100/4090/3090 and flags spot dislocations worth queueing deferred work against.
schedule	0 13 * * *
commits	true
permissions	["contents:write"]
tags	["compute","gpu","depin","pricing"]

Today is ${today}. Read memory/MEMORY.md and memory/topics/compute-pulse.md (if it exists) before starting. If soul/SOUL.md + soul/STYLE.md exist and are populated, match that voice; otherwise be terse and neutral.

Why this skill exists

RunPod runs two pools side-by-side: community cloud (interruptible bid-priced spot, host-supplied) and secure cloud (RunPod-operated on-demand). The gap between them is the price of preemption risk. When the gap widens past ~50% on a given GPU, fault-tolerant batch work (embeddings backfills, fine-tunes with checkpointing, eval sweeps) should be queued onto spot. When it compresses, the spot pool is saturated and on-demand wins.

This skill is the data feed that makes that call legible. It complements compute-pulse (which covers the narrative layer) and compute-futures-eda (which covers the on-chain GPU markets) by giving a clean centralized-spot benchmark.

Inputs

The pre-fetch script (scripts/prefetch-runpod.sh) calls the RunPod GraphQL API outside the sandbox and writes .runpod-cache/gpu-types.json. The query asks for lowestPrice separately per pool (via GraphQL aliases secure: and community:) so the comparison below is like-for-like — not the cheapest across both pools, which would mix bid prices from one pool against on-demand prices from the other. Schema (relevant fields):

{
  "data": {
    "gpuTypes": [
      {
        "id": "NVIDIA A100 80GB PCIe",
        "displayName": "A100 80GB",
        "memoryInGb": 80,
        "secureCloud": true,
        "communityCloud": true,
        "secure":    { "minimumBidPrice": null, "uninterruptablePrice": 1.64 },
        "community": { "minimumBidPrice": 0.79, "uninterruptablePrice": null }
      }
    ]
  }
}

If .runpod-cache/gpu-types.json is missing or empty, the prefetch failed — note it in the output, do not fabricate numbers, and stop. Do NOT call the RunPod API directly from inside the sandbox.

What to do

1. Load .runpod-cache/gpu-types.json. Filter to the GPUs we care about: A100 (40GB + 80GB), H100 (PCIe + SXM + NVL), H200, L40S, 4090, 3090.

2. For each, compute:

   • spot = community.minimumBidPrice (skip if null or communityCloud: false)
   • ondemand = secure.uninterruptablePrice (skip if null or secureCloud: false)
   • discount_pct = (ondemand - spot) / ondemand * 100

   Read spot from the community alias and on-demand from the secure alias — never cross pools. If only one pool publishes a price for a given GPU, skip it rather than comparing across pools.

3. Append a row per GPU to memory/topics/runpod-pricing.md under a dated section. If the file doesn't exist, create it with this header:

   # RunPod pricing — daily snapshot
   
   Spot = community-cloud `minimumBidPrice`. On-demand = secure-cloud `uninterruptablePrice`. Both $/hr per GPU.
   
   ## YYYY-MM-DD
   | GPU | Spot $/hr | On-demand $/hr | Discount | Note |
   

4. Compare today's snapshot against the prior day's section in the same file. For each GPU, compute spot delta (% change). If spot moved more than ±15% day-over-day, OR if today's discount_pct is above 50% (genuine dislocation), add a one-line callout in a ### Signals subsection under today's date.

5. If there are ≥1 signal lines, end with ./notify "runpod-spot: <one-sentence summary of the strongest signal>". Otherwise do not notify — this is mostly a quiet data feed.

Output

A ## Summary block listing:

• file modified: memory/topics/runpod-pricing.md
• GPU count snapshotted
• signals fired (if any) with the specific discount/delta numbers
• prefetch status (success / failed / no API key)

Sandbox note

Auth-required API → pre-fetch pattern. The workflow runs scripts/prefetch-runpod.sh (with RUNPOD_API_KEY from secrets) before Claude starts, and the cached JSON lands at .runpod-cache/gpu-types.json. If RUNPOD_API_KEY is unset the prefetch exits 0 silently — the skill should detect the missing cache file and note "RUNPOD_API_KEY not configured" in its summary, not retry from inside the sandbox.

Setup checklist (one-time, not in scope for this skill run)

• Add RUNPOD_API_KEY to GitHub Actions secrets (read-only key is fine — this skill only queries).
• Wire runpod-spot-pricing into the prefetch dispatch in the workflow YAML so scripts/prefetch-runpod.sh runs for this skill.
• Optional: add a compute chain in aeon.yml that runs this skill before compute-pulse so the weekly narrative read can consume: the pricing table.