{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-08-26-019",
    "source_story_id": "tmp-sidebar-gpu-vs-cache-cost",
    "edition_id": "mp-2026-08-26-morning-0048",
    "edition_url": "https://themachinepress.com/edition/2026-08-26",
    "position": 21,
    "story_type": "ticker",
    "section": "chips-infrastructure",
    "editorial_classification": "editorial",
    "headline": "Cache Compression Won the Cost Axis Until Weights Became the Wall",
    "slug": "cache-compression-won-the-cost-axis-until-weights-became-the-wall",
    "dek": "A profiled simulator places tensor parallelism and KV compression on one cost-latency chart.",
    "summary": "A profiled simulator places tensor parallelism and KV compression on one cost-latency chart.",
    "body_text": "Across two model sizes and three GPU types, compression was 1.20 to 2.00 times cheaper in every comparable memory-relief configuration. Above roughly 36 billion parameters on an 80 GB device, weights—not cache—became the binding constraint, making tensor parallelism an entry requirement despite higher cost.",
    "why_it_matters": "A profiled simulator places tensor parallelism and KV compression on one cost-latency chart.",
    "limitations": [
      "Above roughly 36 billion parameters on an 80 GB device, weights—not cache—became the binding constraint, making tensor parallelism an entry requirement despite higher cost."
    ],
    "importance": 6,
    "canonical_url": "https://themachinepress.com/story/mp-2026-08-26-019/cache-compression-won-the-cost-axis-until-weights-became-the-wall",
    "json_url": "https://themachinepress.com/story/mp-2026-08-26-019.json",
    "first_published_at": "2026-08-26T09:00:00.000-04:00",
    "modified_at": "2026-08-26T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-08-26-019-001",
        "text": "A profiled simulator places tensor parallelism and KV compression on one cost-latency chart.",
        "source_ids": [
          "source-2026-08-26-021"
        ],
        "qualification": "Above roughly 36 billion parameters on an 80 GB device, weights—not cache—became the binding constraint, making tensor parallelism an entry requirement despite higher cost."
      }
    ],
    "source_ids": [
      "source-2026-08-26-021"
    ],
    "tags": [
      "LLM serving",
      "GPU cost",
      "KV compression"
    ],
    "image_url": null,
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-08-26-021",
      "title": "arXiv preprint 2608.23962",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2608.23962",
      "canonical_url": "https://arxiv.org/abs/2608.23962",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-08-24T21:42:08.000-04:00",
      "accessed_at": "2026-08-26T08:17:00.000-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-08-26-019-001"
      ]
    }
  ],
  "corrections": [],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "Cache Compression Won the Cost Axis Until Weights Became the Wall",
    "publisher": "The Machine Press",
    "published_at": "2026-08-26T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-08-26-019/cache-compression-won-the-cost-axis-until-weights-became-the-wall"
  }
}