Compute — landed cost per job

GPU list prices are everywhere. What nobody publishes is the landed cost of an actual job — the same number across currencies, billing grains (per-second vs per-hour vs minimum charge) and payment friction. That is what this endpoint returns.

Jobs

Two mental models are precomputed: kind="workload" (what one fine-tuning run costs, durations from a first-party benchmark report) and kind="hour" (one card for one hour, for straight per-card comparison).

curl https://tkx.org/api/v1/compute/jobs.json

// example unavailable at build time

{
  "slug": "qlora-13b",
  "gpu": "4090",
  "gpus": 1,
  "hours": 3.5,
  "source": "echo《云端 GPU 租卡行情报告 2026-07》实测基准",
  "groups": [
    {
      "group": "direct",
      "rows": [
        {
          "cloud": "autodl",
          "name": "AutoDL",
          "url": "https://www.autodl.com/",
          "unitPrice": 2.15,
          "cur": "¥",
          "billing": "perhour",
          "minBillSec": null,
          "interruptible": false,
          "substitute": null,
          "tier": "high",
          "stale": false,
          "cnyCents": 860,
          "cnyCentsLo": 792,
          "billedSec": 14400,
          "basis": [
            {
              "k": "list",
              "v": "¥2.15 / GPU-hr · autodl · high · 2026-07-04"
            },
            {
              "k": "grain",
              "v": "per-hour"
            },
            {
              "k": "min",
              "v": "no minimum"
            }
          ]
        }
      ]
    }
  ]
}

Friction groups

Grouping is derived from verifiable facts (what the checkout accepts, whether the console resolves) — not from a platform score. Set the buyer's origin outside China and every row collapses into direct.

Clouds

curl https://tkx.org/api/v1/compute/clouds.json

// example unavailable at build time