{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-09-11-008",
    "source_story_id": "tmp-story-ai-training-power-elasticity",
    "edition_id": "mp-2026-09-11-morning-0064",
    "edition_url": "https://themachinepress.com/edition/2026-09-11",
    "position": 8,
    "story_type": "dispatch",
    "section": "chips-infrastructure",
    "editorial_classification": "editorial",
    "headline": "Thirty Percent Less Power Did Not Cost Every Training Job Equally",
    "slug": "thirty-percent-less-power-did-not-cost-every-training-job-equally",
    "dek": "Across 189 H100 and H200 runs, a job-aware allocator recovered 63% of the throughput gap to an oracle under a power cap.",
    "summary": "Across 189 H100 and H200 runs, a job-aware allocator recovered 63% of the throughput gap to an oracle under a power cap.",
    "body_text": "The study defines a Power Flexibility Index to measure how much LLM-training throughput changes when GPU power is reduced. Its evidence covers 131 H200 runs, 24 H200 validations and 34 matched H100 runs across dense and mixture-of-experts models, pretraining and fine-tuning, and deployments up to 32 GPUs. Under a 30% power reduction, allocating power by the learned index recovered about 1,500 tokens per second per job—63% of the gap between equal allocation and perfect foresight. The result shows measurable job-level flexibility, not a general claim about data-center electricity or grid impacts.",
    "why_it_matters": "Across 189 H100 and H200 runs, a job-aware allocator recovered 63% of the throughput gap to an oracle under a power cap.",
    "limitations": [
      "The result shows measurable job-level flexibility, not a general claim about data-center electricity or grid impacts."
    ],
    "importance": 8,
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-11-008/thirty-percent-less-power-did-not-cost-every-training-job-equally",
    "json_url": "https://themachinepress.com/story/mp-2026-09-11-008.json",
    "first_published_at": "2026-09-11T09:00:00.000-04:00",
    "modified_at": "2026-09-11T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-09-11-008-001",
        "text": "Across 189 H100 and H200 runs, a job-aware allocator recovered 63% of the throughput gap to an oracle under a power cap.",
        "source_ids": [
          "source-2026-09-11-008"
        ],
        "qualification": "The result shows measurable job-level flexibility, not a general claim about data-center electricity or grid impacts."
      }
    ],
    "source_ids": [
      "source-2026-09-11-008"
    ],
    "tags": [
      "AI infrastructure",
      "GPU power",
      "training"
    ],
    "image_url": "https://themachinepress.com/issues/2026-09-11/ai-power-server-file.webp",
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-09-11-008",
      "title": "arXiv preprint 2609.11542",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2609.11542",
      "canonical_url": "https://arxiv.org/abs/2609.11542",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-09-10T09:40:13.000-04:00",
      "accessed_at": "2026-09-11T08:25:00.000-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-09-11-008-001"
      ]
    }
  ],
  "corrections": [],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "Thirty Percent Less Power Did Not Cost Every Training Job Equally",
    "publisher": "The Machine Press",
    "published_at": "2026-09-11T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-11-008/thirty-percent-less-power-did-not-cost-every-training-job-equally"
  }
}