{
  "schema": "aicosts-corpus/1",
  "title": "The real cost of AI",
  "canonical_url": "https://costs.lyra-forge.com/",
  "description": "What it actually costs to build and run on AI in August 2026 \u2014 inference economics, GPU ownership, model selection, routing, fine-tuning, and the 24-month forecast \u2014 carried with its evidence and its boundaries.",
  "status": "dated cost synthesis, maintained",
  "evidence_cutoff": "2026-08-20",
  "refresh_by": "2026-11-20",
  "updates": [
    {
      "date": "2026-08-21",
      "label": "Public surface established.",
      "change": "Stood up the allowlisted static publication at costs.lyra-forge.com: the report at a stable route with its Markdown source, ten addressable report sections at /parts/, seven audience routes at /for/, the cross-cutting reader questions at /questions/, and the package-wide update log backstage. The seven research notebooks are declared private in their own front matter and are excluded from every release by the verifier."
    }
  ],
  "interpretation_rules": [
    "Treat every price as dated evidence of a posted rate, not as a durable claim; provider pricing changes weekly.",
    "Keep verified claims, modeled crossovers, and dated forecast bets distinct; do not report a bet as a finding.",
    "Carry the index-dependence caveat with any quality-adjusted cost figure \u2014 the underlying index is an editorially-weighted composite, not a workload predictor.",
    "State the procurement scope and utilization assumption whenever quoting a build-versus-rent crossover.",
    "Follow the report to its primary sources before repeating a quantitative or time-sensitive claim.",
    "Where this package says a quantity is unmeasured, do not substitute an estimate for it."
  ],
  "recommended_entry_points": {
    "public_overview": "https://costs.lyra-forge.com/",
    "choose_by_decision": "https://costs.lyra-forge.com/for/",
    "report_sections": "https://costs.lyra-forge.com/parts/",
    "reader_questions": "https://costs.lyra-forge.com/questions/",
    "full_report": "https://costs.lyra-forge.com/reports/the-real-cost-of-ai/",
    "full_report_markdown": "https://costs.lyra-forge.com/reports/the-real-cost-of-ai/source.md"
  },
  "reports": [
    {
      "title": "The real cost of AI: August 2026",
      "description": "The maintained synthesis: what a token costs, when owning GPUs beats renting, what routing is worth, when fine-tuning pays back, and six dated bets on the next 24 months.",
      "surface": "public",
      "html_url": "https://costs.lyra-forge.com/reports/the-real-cost-of-ai/",
      "markdown_url": "https://costs.lyra-forge.com/reports/the-real-cost-of-ai/source.md",
      "updates": [
        {
          "date": "2026-08-21",
          "label": "pass 14",
          "change": "Added the missing fine-tuning decision layer (researcher route): PEFT economics, compute-as-subscription conditions (safety-drift evals per refit; base-model deprecation refits), and the RAG-substitutes/complements split \u2014 from the finetuning notebook's three verified claims, previously absent from the report."
        },
        {
          "date": "2026-08-21",
          "label": "pass 13",
          "change": "Decision-guidance sections updated to match current evidence: infrastructure advice rewritten around the two-regime TCO answer (replacing the superseded 50\u201383% single-band figure) with control-premium guidance; API-builder advice carries the token-efficiency and index-weighting caveats."
        },
        {
          "date": "2026-08-21",
          "label": "pass 12",
          "change": "Adversarial pass on the 47x cost-per-quality spread: verified C6 (model-selection notebook) \u2014 the AA Intelligence Index is an editorially-weighted, text/English-centric composite whose v4.1 reweighting moved scores by ~23 points (comparable to the frontier spread itself). The 47x is now presented as order-of-magnitude evidence of non-linear pricing, not a precise ratio."
        },
        {
          "date": "2026-08-21",
          "label": "pass 11",
          "change": "Verified C4 (GPU notebook): regulatory sovereignty drives on-prem/sovereign AI infrastructure independent of cost \u2014 $80B 2026 sovereign IaaS forecast (+35.6%), Italy +20% YoY on sovereign-data demand, EU Sovereignty Package (June 2026). Compliance buyers pay the control premium as a requirement. Also fixed a custody violation (Italy figure cited from a search snippet before capture)."
        },
        {
          "date": "2026-08-21",
          "label": "pass 10",
          "change": "Control-premium analysis added to the GPU section: independent itemization (Vensas, grade A) confirms self-hosting at 3\u20135\u00d7 pure GPU price, independently corroborating the io.net node-loaded correction; five evidenced motives for paying it (sovereignty/compliance, guaranteed capacity, latency, vendor independence, sustained high utilization); premium quantified at roughly 2\u20134\u00d7 effective cost vs neoclouds."
        },
        {
          "date": "2026-08-21",
          "label": "pass 9",
          "change": "Adversarial pass on the TCO model found a material flaw: our bands used GPU-street-price capex ($30K/GPU), omitting node integration and financing. With io.net's measured enterprise figure ($93,750/GPU over 3yr), on-prem crossovers move dramatically (vs neocloud median: ~37% \u2192 ~99%, i.e., effectively never; vs spot: never) \u2014 reconciling previously contradictory custody claims as population differences (lean GPU-only buyers vs enterprise node-loaded buyers). Model runs both scenarios; report section rewritten."
        },
        {
          "date": "2026-08-21",
          "label": "pass 8",
          "change": "Adversarial pass on the 111x headline: no source disputes the posted-price spread, but measured token-efficiency variance (TensorZero, grade A: 2.65x+ same-input token-count difference) and retry/verbosity costs mean the posted range is an upper bound on effective dispersion. Headline and claim prose updated; C1 corroboration now 5."
        },
        {
          "date": "2026-08-21",
          "label": "pass 7",
          "change": "Adversarial pass on the routing-savings claim: hunted counterevidence, found quality-accounted confirmation instead (RouteNLP 8-week pilot: 58% cost reduction at 91% acceptance; corroboration now 5). Claim prose and report passage updated with the acceptance-rate condition."
        },
        {
          "date": "2026-08-21",
          "label": "pass 6",
          "change": "New reader-commissioned angle: the routing/middleware layer itself. Verified C4 (thin flat take-rate \u2014 OpenRouter 5.5%, >$100M annualized spend routed) and C5 (structural commoditization \u2014 free self-host gateways, near-zero switching costs, Stripe's $7.5B acquisition as the sustainability answer). Added \"who absorbs model-fit cost\" analysis: developers, not routers."
        },
        {
          "date": "2026-08-21",
          "label": "pass 5",
          "change": "Custody audit of this report: notebook inventory counts refreshed to current custody (15 refs / 5 claims inference-economics; 7 refs / 3 claims + TCO model GPU; 8 refs / 3 claims model-selection); spot-risk scenario computed in the TCO model (spot-vs-on-prem crossover ~67% \u2192 ~47% at a 20% interruption premium \u2014 replaces the earlier estimated figure); C5 cache-write caveat added."
        },
        {
          "date": "2026-08-21",
          "label": "pass 4",
          "change": "Evidence-strength pass: all seven previously single-corroboration load-bearing claims brought to \u22652 independent sources \u2014 closed-weight premium (DeepInfra), caching/batch discounts (Amnic, pecollective), routing savings (leanlm, sirayatech), index spread (swfte composite), NVIDIA cadence (ModulEdge/Kuo + official PR), edge NPU price/perf (Hackster.io + official), startup credit amounts (AWS/GCP primary terms + pragma-code comparison). No conclusions changed; confidence strengthened."
        },
        {
          "date": "2026-08-21",
          "label": "pass 3",
          "change": "Built the GPU TCO crossover model (gpu-003): crossover bands vs on-prem quantified per lane \u2014 hyperscaler on-demand ~21%, 1-yr reserved ~30%, 3-yr ~43%, neocloud median ~37%/low ~51%, spot ~67% utilization; workload views added (always-on \u2192 on-prem; business-hours \u2192 spot). Refined the earlier \"on-prem rarely wins vs neoclouds\" statement into a named-assumption band (amortization window and staffing move it across ~35\u201365%)."
        },
        {
          "date": "2026-08-21",
          "label": "pass 2",
          "change": "Captured the missing forecast baselines: verified C3 in the subsidies notebook (frontier list pricing below sustainable infrastructure return \u2014 OpenAI GM 33% vs 46% forecast, inference ~$8.4B 4x YoY; Anthropic 40%) and C4 in the edge notebook (current-gen NPU throughput at consumer prices: $249 Orin Nano Super ~35\u201354 tok/s on 1B-class models; AGX Thor 41.3 tok/s Llama 3.1 8B, 61 tok/s Qwen3 30B-A3B). Forecast bets FC4/FC6 upgraded from pure priors to informed bets; subsidy evidence boundary updated."
        },
        {
          "date": "2026-08-21",
          "label": "",
          "change": "Asserted the six dated forecast bets (FC1\u2013FC6) from the ai-cost-forecast notebook; replaced the \"no forecast yet\" boundary with the bet table and its uncertainty statement. No price claims changed."
        },
        {
          "date": "2026-08-20",
          "label": "",
          "change": "Initial research snapshot. First public report. Drew on inference-economics (then 7 sources, 4 verified claims) and GPU-ownership (then 4 sources, 2 verified claims) notebooks. GPU sources re-captured via --via direct after default lane failed on cistern miss; record-only sources superseded and passed."
        }
      ]
    }
  ],
  "audience_routes": [
    {
      "title": "Developers and ML engineers",
      "url": "https://costs.lyra-forge.com/for/developers-and-ml-engineers/",
      "decision": "Which model, provider, and pricing structure to commit this workload to.",
      "proof_standard": "A price you can reproduce from a posted rate card, and a quality adjustment whose weighting you can inspect.",
      "working_output": "An effective cost per successful task for your own workload \u2014 not a per-token rate.",
      "decision_checks": [
        "Have you measured output-token volume per task on each candidate, not just the input price?",
        "Does this workload reuse enough context for caching to beat the cache-write premium?",
        "Would a 23-point index revision change which model you picked?"
      ],
      "open_remainder": "A routing and caching decision calculator. Specified in the audience ledger, not built \u2014 public tooling is board-gated, not blocked on research.",
      "parts": [
        "https://costs.lyra-forge.com/parts/short-version/",
        "https://costs.lyra-forge.com/parts/inference-economics-claims/",
        "https://costs.lyra-forge.com/parts/build-with-apis/",
        "https://costs.lyra-forge.com/parts/routing-layer/",
        "https://costs.lyra-forge.com/parts/evidence-boundaries/"
      ]
    },
    {
      "title": "Engineering leaders and CTOs",
      "url": "https://costs.lyra-forge.com/for/engineering-leaders/",
      "decision": "What AI spend envelope to plan for, and what would move it.",
      "proof_standard": "A dated bet with a named annulment condition, not a trend line.",
      "working_output": "A budget envelope with the bifurcation scenario priced separately from the base case.",
      "decision_checks": [
        "Does the plan survive the subsidy-contraction bet resolving against you?",
        "Is the commodity and frontier spend budgeted separately, or as one line?",
        "Which named annulment condition would you actually notice if it fired?"
      ],
      "open_remainder": "A budget planning worksheet keyed to the demand-elasticity scenarios.",
      "parts": [
        "https://costs.lyra-forge.com/parts/short-version/",
        "https://costs.lyra-forge.com/parts/twenty-four-month-horizon/",
        "https://costs.lyra-forge.com/parts/budget-ai-spend/",
        "https://costs.lyra-forge.com/parts/evidence-boundaries/"
      ]
    },
    {
      "title": "Platform and infrastructure teams",
      "url": "https://costs.lyra-forge.com/for/platform-and-infrastructure/",
      "decision": "Rent, reserve, or own \u2014 and at what utilization.",
      "proof_standard": "A crossover band with its procurement scope and amortization window stated on the face of it.",
      "working_output": "A utilization threshold computed for your own procurement regime.",
      "decision_checks": [
        "Is your utilization figure a measured duty cycle or a planning assumption?",
        "Are you pricing GPUs only, or loaded nodes with everything around them?",
        "If compliance or capacity certainty is a hard constraint, is cost-optimal even the right target?"
      ],
      "open_remainder": "The repository's TCO crossover model exposed as an interactive tool with utilization and procurement-scope inputs.",
      "parts": [
        "https://costs.lyra-forge.com/parts/short-version/",
        "https://costs.lyra-forge.com/parts/gpu-ownership-claims/",
        "https://costs.lyra-forge.com/parts/run-ml-infrastructure/",
        "https://costs.lyra-forge.com/parts/twenty-four-month-horizon/",
        "https://costs.lyra-forge.com/parts/evidence-boundaries/"
      ]
    },
    {
      "title": "Researchers and data scientists",
      "url": "https://costs.lyra-forge.com/for/researchers-and-data-scientists/",
      "decision": "Whether to fine-tune, retrieve, or prompt.",
      "proof_standard": "A payback volume, plus the retraining cadence that resets it.",
      "working_output": "A break-even task volume for the narrow model against the general one.",
      "decision_checks": [
        "Have you costed data preparation and evaluation, or only GPU hours?",
        "Is the task knowledge-heavy or behavior-heavy \u2014 substitutes or complements?",
        "How often does the underlying knowledge change, and what does that do to payback?"
      ],
      "open_remainder": "A payback calculator that takes retraining cadence as an input.",
      "parts": [
        "https://costs.lyra-forge.com/parts/short-version/",
        "https://costs.lyra-forge.com/parts/finetuning-vs-retrieval/",
        "https://costs.lyra-forge.com/parts/inference-economics-claims/",
        "https://costs.lyra-forge.com/parts/evidence-boundaries/"
      ]
    },
    {
      "title": "Procurement and finance teams",
      "url": "https://costs.lyra-forge.com/for/procurement-and-finance/",
      "decision": "What the true, unsubsidized cost of a provider commitment is.",
      "proof_standard": "Primary program terms and a cost-side margin anchor, not a vendor estimate.",
      "working_output": "A subsidy-expiry exposure list, one line per provider commitment.",
      "decision_checks": [
        "Does any current unit economic depend on a credit balance or a free tier?",
        "What is the switching cost if the zero-price channel caps change?",
        "Is the lock-in priced, or only the rate?"
      ],
      "open_remainder": "A subsidy-expiry exposure checklist, with the zero-price channel caps rechecked quarterly.",
      "parts": [
        "https://costs.lyra-forge.com/parts/short-version/",
        "https://costs.lyra-forge.com/parts/budget-ai-spend/",
        "https://costs.lyra-forge.com/parts/twenty-four-month-horizon/",
        "https://costs.lyra-forge.com/parts/evidence-boundaries/"
      ]
    },
    {
      "title": "Founders and product teams",
      "url": "https://costs.lyra-forge.com/for/founders-and-product-teams/",
      "decision": "Whether the unit economics of an AI feature survive growth.",
      "proof_standard": "A cost per user computed from measured token volume, with the subsidy assumption named.",
      "working_output": "A cost-per-user model carrying an explicit subsidy-withdrawal scenario.",
      "decision_checks": [
        "Is cost per user measured from real traffic or estimated from a rate card?",
        "Does the margin hold if effective cost rises 1.5\u20133x?",
        "Which retry and verbosity behavior is inside your cost model?"
      ],
      "open_remainder": "A worked cost-per-user case with the retry loop included.",
      "parts": [
        "https://costs.lyra-forge.com/parts/short-version/",
        "https://costs.lyra-forge.com/parts/build-with-apis/",
        "https://costs.lyra-forge.com/parts/budget-ai-spend/",
        "https://costs.lyra-forge.com/parts/twenty-four-month-horizon/"
      ]
    },
    {
      "title": "Executives, investors, and broad readers",
      "url": "https://costs.lyra-forge.com/for/executives-and-investors/",
      "decision": "Whether the cost curve supports the plan being underwritten.",
      "proof_standard": "Dated binary bets with probabilities and annulment conditions.",
      "working_output": "A short account of where costs head that survives being asked what would change it.",
      "decision_checks": [
        "Is the thesis stated as a bet with a date, or as a direction?",
        "Which bet resolving against you would matter most, and when does it resolve?",
        "Does the plan assume today's posted prices are sustainable ones?"
      ],
      "open_remainder": "A forecast tracking page, once the first resolution evidence lands.",
      "parts": [
        "https://costs.lyra-forge.com/parts/short-version/",
        "https://costs.lyra-forge.com/parts/twenty-four-month-horizon/",
        "https://costs.lyra-forge.com/parts/evidence-boundaries/"
      ]
    }
  ],
  "reader_questions": [
    {
      "question": "What does a token actually cost?",
      "url": "https://costs.lyra-forge.com/questions/#what-does-a-token-cost",
      "state": "Answered",
      "answer": "The posted range across comparable models is 111x, and the quality-adjusted spread is wider still \u2014 but effective cost is set by output volume, retries, and cache economics, not by the rate card.",
      "open_remainder": "An effective-cost calculator. Board-gated tooling, not blocked on research.",
      "parts": [
        "https://costs.lyra-forge.com/parts/inference-economics-claims/",
        "https://costs.lyra-forge.com/parts/build-with-apis/"
      ]
    },
    {
      "question": "Should I run my own GPUs?",
      "url": "https://costs.lyra-forge.com/questions/#should-i-run-my-own-gpus",
      "state": "Answered, and modeled in the repository",
      "answer": "It depends on utilization and procurement scope, and the crossover band is quantified per lane rather than argued.",
      "open_remainder": "A public-facing version of the crossover model, and a quarterly price refresh.",
      "parts": [
        "https://costs.lyra-forge.com/parts/gpu-ownership-claims/",
        "https://costs.lyra-forge.com/parts/run-ml-infrastructure/"
      ]
    },
    {
      "question": "Which model should I use?",
      "url": "https://costs.lyra-forge.com/questions/#which-model-should-i-use",
      "state": "Answered",
      "answer": "Route by measured effective cost per successful task at your quality bar; the savings are real and quality-conditioned, and the routing fee is thin.",
      "open_remainder": "A routing decision tool.",
      "parts": [
        "https://costs.lyra-forge.com/parts/build-with-apis/",
        "https://costs.lyra-forge.com/parts/routing-layer/",
        "https://costs.lyra-forge.com/parts/inference-economics-claims/"
      ]
    },
    {
      "question": "Is fine-tuning worth it?",
      "url": "https://costs.lyra-forge.com/questions/#is-fine-tuning-worth-it",
      "state": "Answered",
      "answer": "The compute is cheap and everything around it is not; the substitute-or-complement question with retrieval decides it before the cost model does.",
      "open_remainder": "A payback calculator.",
      "parts": [
        "https://costs.lyra-forge.com/parts/finetuning-vs-retrieval/"
      ]
    },
    {
      "question": "How much of my AI spend is subsidized?",
      "url": "https://costs.lyra-forge.com/questions/#how-much-of-my-spend-is-subsidized",
      "state": "Answered structurally, anchored on the cost side",
      "answer": "Frontier list pricing runs below sustainable infrastructure return, so posted prices are effectively investor-subsidized \u2014 but the user-side subsidized fraction is unmeasured and is carried as an open boundary.",
      "open_remainder": "An aggregate fraction, if one is ever published. Zero-price channel caps rechecked quarterly.",
      "parts": [
        "https://costs.lyra-forge.com/parts/budget-ai-spend/",
        "https://costs.lyra-forge.com/parts/evidence-boundaries/"
      ]
    },
    {
      "question": "Where are costs headed in 24 months?",
      "url": "https://costs.lyra-forge.com/questions/#where-are-costs-headed",
      "state": "Answered as six dated bets",
      "answer": "Not as a curve. Six binary bets, each with a probability, an evidence grade, and a condition that annuls it.",
      "open_remainder": "Resolution scoring as the dates arrive.",
      "parts": [
        "https://costs.lyra-forge.com/parts/twenty-four-month-horizon/"
      ]
    },
    {
      "question": "Can I run inference on the edge?",
      "url": "https://costs.lyra-forge.com/questions/#can-i-run-inference-on-the-edge",
      "state": "Answered in research custody; carried onto this site only as a forecast bet",
      "answer": "Quantization studies and current-generation NPU throughput baselines are held in the research notebooks. The public report carries the edge question as a dated bet rather than a decision surface.",
      "open_remainder": "An edge feasibility route on this site, and the tool behind it.",
      "parts": [
        "https://costs.lyra-forge.com/parts/twenty-four-month-horizon/"
      ]
    },
    {
      "question": "How do all these dimensions interact?",
      "url": "https://costs.lyra-forge.com/questions/#how-do-these-dimensions-interact",
      "state": "Partially answered",
      "answer": "Interactions are named where evidence exists \u2014 spot risk against on-prem, subsidy against list price, efficiency against the hardware split \u2014 and left open where it does not.",
      "open_remainder": "A cross-dimensional interaction study. Open research, not a publication gap.",
      "parts": [
        "https://costs.lyra-forge.com/parts/evidence-boundaries/",
        "https://costs.lyra-forge.com/parts/twenty-four-month-horizon/"
      ]
    }
  ],
  "parts": [
    {
      "slug": "short-version",
      "sequence": 1,
      "kind": "orientation",
      "title": "The short version",
      "group": "The short version",
      "question": "What are the four findings that change a cost decision?",
      "answer": "Sticker price is not the cost of an answer; owning GPUs is a crossover problem set by utilization and procurement scope; the control premium is real and knowingly paid; and a material share of current spend is subsidized.",
      "boundary": "Every figure here is a posted rate on 2026-08-20 and a quality adjustment against an editorially-weighted index. Neither is a durable claim.",
      "html_url": "https://costs.lyra-forge.com/parts/short-version/",
      "source_report_url": "https://costs.lyra-forge.com/reports/the-real-cost-of-ai/",
      "source_heading": "The short version",
      "source_anchor_url": "https://costs.lyra-forge.com/reports/the-real-cost-of-ai/#the-short-version",
      "audience_routes": [
        {
          "title": "Developers and ML engineers",
          "url": "https://costs.lyra-forge.com/for/developers-and-ml-engineers/",
          "use": "The four findings that change the answer before any per-model comparison starts."
        },
        {
          "title": "Engineering leaders and CTOs",
          "url": "https://costs.lyra-forge.com/for/engineering-leaders/",
          "use": "The bifurcation thesis stated before the numbers that support it."
        },
        {
          "title": "Platform and infrastructure teams",
          "url": "https://costs.lyra-forge.com/for/platform-and-infrastructure/",
          "use": "Why this is a crossover problem rather than a preference."
        },
        {
          "title": "Researchers and data scientists",
          "url": "https://costs.lyra-forge.com/for/researchers-and-data-scientists/",
          "use": "The cost frame this decision sits inside."
        },
        {
          "title": "Procurement and finance teams",
          "url": "https://costs.lyra-forge.com/for/procurement-and-finance/",
          "use": "The subsidy finding, stated with the part of it that is still unmeasured."
        },
        {
          "title": "Founders and product teams",
          "url": "https://costs.lyra-forge.com/for/founders-and-product-teams/",
          "use": "The two findings that most often break an AI feature's margin model."
        },
        {
          "title": "Executives, investors, and broad readers",
          "url": "https://costs.lyra-forge.com/for/executives-and-investors/",
          "use": "The whole argument in four findings."
        }
      ],
      "related_urls": [
        "https://costs.lyra-forge.com/parts/inference-economics-claims/",
        "https://costs.lyra-forge.com/parts/gpu-ownership-claims/",
        "https://costs.lyra-forge.com/parts/evidence-boundaries/"
      ]
    },
    {
      "slug": "inference-economics-claims",
      "sequence": 2,
      "kind": "evidence",
      "title": "Inference economics \u2014 five verified claims",
      "group": "What the evidence establishes",
      "question": "What does a token actually cost, and how far does effective cost diverge from the posted price?",
      "answer": "A 111x posted price range across comparable models, a 47x spread once cost is adjusted for quality, a non-linear frontier above Intelligence Index 50, a ~10% closed-weight premium, and caching and batch as the two largest non-routing levers.",
      "boundary": "The quality adjustment inherits the index's editorial weighting; version overhauls have moved scores by ~23 points.",
      "html_url": "https://costs.lyra-forge.com/parts/inference-economics-claims/",
      "source_report_url": "https://costs.lyra-forge.com/reports/the-real-cost-of-ai/",
      "source_heading": "Inference economics \u2014 five verified claims",
      "source_anchor_url": "https://costs.lyra-forge.com/reports/the-real-cost-of-ai/#inference-economics-five-verified-claims",
      "audience_routes": [
        {
          "title": "Developers and ML engineers",
          "url": "https://costs.lyra-forge.com/for/developers-and-ml-engineers/",
          "use": "The verified price range, the quality adjustment, and the two largest non-routing levers."
        },
        {
          "title": "Researchers and data scientists",
          "url": "https://costs.lyra-forge.com/for/researchers-and-data-scientists/",
          "use": "The per-token baseline the narrow model has to beat in production, not in training."
        }
      ],
      "related_urls": [
        "https://costs.lyra-forge.com/parts/short-version/",
        "https://costs.lyra-forge.com/parts/build-with-apis/",
        "https://costs.lyra-forge.com/parts/routing-layer/"
      ]
    },
    {
      "slug": "gpu-ownership-claims",
      "sequence": 3,
      "kind": "evidence",
      "title": "GPU ownership \u2014 three verified claims and a crossover model",
      "group": "What the evidence establishes",
      "question": "When does owning GPUs beat renting them, and what sets the threshold?",
      "answer": "Specialist clouds charge 50\u201370% less than hyperscalers for the same H100, and the on-prem crossover splits by procurement scope: a lean operator breaks even against hyperscaler on-demand near 21% utilization, an enterprise node-loaded buyer near 57%.",
      "boundary": "The crossover bands move across roughly 35\u201365% with the amortization window and staffing assumptions named in the model.",
      "html_url": "https://costs.lyra-forge.com/parts/gpu-ownership-claims/",
      "source_report_url": "https://costs.lyra-forge.com/reports/the-real-cost-of-ai/",
      "source_heading": "GPU ownership \u2014 three verified claims and a crossover model",
      "source_anchor_url": "https://costs.lyra-forge.com/reports/the-real-cost-of-ai/#gpu-ownership-three-verified-claims-and-a-crossover-model",
      "audience_routes": [
        {
          "title": "Platform and infrastructure teams",
          "url": "https://costs.lyra-forge.com/for/platform-and-infrastructure/",
          "use": "The measured price gap between specialist clouds and hyperscalers, and the two procurement regimes."
        }
      ],
      "related_urls": [
        "https://costs.lyra-forge.com/parts/run-ml-infrastructure/",
        "https://costs.lyra-forge.com/parts/short-version/",
        "https://costs.lyra-forge.com/parts/evidence-boundaries/"
      ]
    },
    {
      "slug": "build-with-apis",
      "sequence": 4,
      "kind": "decision",
      "title": "If you build with AI APIs",
      "group": "What this means for decisions",
      "question": "Which model, provider, and pricing structure should this workload commit to?",
      "answer": "Measure effective cost per successful task rather than per token, then take the caching and batch discounts, then route \u2014 in that order.",
      "boundary": "Cache economics reverse for workloads without enough context reuse to amortize the cache-write premium.",
      "html_url": "https://costs.lyra-forge.com/parts/build-with-apis/",
      "source_report_url": "https://costs.lyra-forge.com/reports/the-real-cost-of-ai/",
      "source_heading": "If you build with AI APIs",
      "source_anchor_url": "https://costs.lyra-forge.com/reports/the-real-cost-of-ai/#if-you-build-with-ai-apis",
      "audience_routes": [
        {
          "title": "Developers and ML engineers",
          "url": "https://costs.lyra-forge.com/for/developers-and-ml-engineers/",
          "use": "The decision as a sequence: measure effective cost, take the discounts, then route."
        },
        {
          "title": "Founders and product teams",
          "url": "https://costs.lyra-forge.com/for/founders-and-product-teams/",
          "use": "The levers available before you change the product."
        }
      ],
      "related_urls": [
        "https://costs.lyra-forge.com/parts/inference-economics-claims/",
        "https://costs.lyra-forge.com/parts/routing-layer/",
        "https://costs.lyra-forge.com/parts/finetuning-vs-retrieval/"
      ]
    },
    {
      "slug": "run-ml-infrastructure",
      "sequence": 5,
      "kind": "decision",
      "title": "If you run ML infrastructure",
      "group": "What this means for decisions",
      "question": "Rent, reserve, or own \u2014 at what utilization and under whose procurement scope?",
      "answer": "The answer is a utilization threshold, not a preference, and the threshold moves with whether you are adding GPUs to existing infrastructure or buying loaded nodes.",
      "boundary": "Compliance, capacity certainty, and latency constraints can make the cost-optimal answer the wrong one; the control premium runs roughly 2\u20134x.",
      "html_url": "https://costs.lyra-forge.com/parts/run-ml-infrastructure/",
      "source_report_url": "https://costs.lyra-forge.com/reports/the-real-cost-of-ai/",
      "source_heading": "If you run ML infrastructure",
      "source_anchor_url": "https://costs.lyra-forge.com/reports/the-real-cost-of-ai/#if-you-run-ml-infrastructure",
      "audience_routes": [
        {
          "title": "Platform and infrastructure teams",
          "url": "https://costs.lyra-forge.com/for/platform-and-infrastructure/",
          "use": "The thresholds arranged as the decision you actually have to make."
        }
      ],
      "related_urls": [
        "https://costs.lyra-forge.com/parts/gpu-ownership-claims/",
        "https://costs.lyra-forge.com/parts/budget-ai-spend/",
        "https://costs.lyra-forge.com/parts/twenty-four-month-horizon/"
      ]
    },
    {
      "slug": "budget-ai-spend",
      "sequence": 6,
      "kind": "decision",
      "title": "If you budget AI spend",
      "group": "What this means for decisions",
      "question": "What envelope should the next budget cycle assume, and what would break it?",
      "answer": "Budget for an effective cost above what you currently pay, because a material share of the posted price is investor-subsidized, and price the subsidy-withdrawal scenario separately.",
      "boundary": "The user-side subsidized fraction is unmeasured; the cost-side anchor is lab gross margins below sustainable infrastructure return.",
      "html_url": "https://costs.lyra-forge.com/parts/budget-ai-spend/",
      "source_report_url": "https://costs.lyra-forge.com/reports/the-real-cost-of-ai/",
      "source_heading": "If you budget AI spend",
      "source_anchor_url": "https://costs.lyra-forge.com/reports/the-real-cost-of-ai/#if-you-budget-ai-spend",
      "audience_routes": [
        {
          "title": "Engineering leaders and CTOs",
          "url": "https://costs.lyra-forge.com/for/engineering-leaders/",
          "use": "The envelope, and the subsidy assumption it silently depends on."
        },
        {
          "title": "Procurement and finance teams",
          "url": "https://costs.lyra-forge.com/for/procurement-and-finance/",
          "use": "What to assume when the credit balance runs out."
        },
        {
          "title": "Founders and product teams",
          "url": "https://costs.lyra-forge.com/for/founders-and-product-teams/",
          "use": "The envelope the margin model has to live inside."
        }
      ],
      "related_urls": [
        "https://costs.lyra-forge.com/parts/twenty-four-month-horizon/",
        "https://costs.lyra-forge.com/parts/evidence-boundaries/",
        "https://costs.lyra-forge.com/parts/short-version/"
      ]
    },
    {
      "slug": "twenty-four-month-horizon",
      "sequence": 7,
      "kind": "forecast",
      "title": "The 24-month horizon \u2014 six dated bets",
      "group": "What this means for decisions",
      "question": "Where do costs head, and which dated bets would falsify that?",
      "answer": "Six dated binary bets with explicit probabilities and annulment conditions \u2014 efficiency-driven decline, custom-silicon share, demand elasticity, subsidy contraction, open-weight parity, and edge expansion.",
      "boundary": "These are probability-bearing judgments over a short record, not calibrated forecasts; two of the six rest on structured priors with named missing baselines.",
      "html_url": "https://costs.lyra-forge.com/parts/twenty-four-month-horizon/",
      "source_report_url": "https://costs.lyra-forge.com/reports/the-real-cost-of-ai/",
      "source_heading": "The 24-month horizon \u2014 six dated bets",
      "source_anchor_url": "https://costs.lyra-forge.com/reports/the-real-cost-of-ai/#the-24-month-horizon-six-dated-bets",
      "audience_routes": [
        {
          "title": "Engineering leaders and CTOs",
          "url": "https://costs.lyra-forge.com/for/engineering-leaders/",
          "use": "Six dated bets with probabilities, in place of a curve you cannot defend."
        },
        {
          "title": "Platform and infrastructure teams",
          "url": "https://costs.lyra-forge.com/for/platform-and-infrastructure/",
          "use": "The hardware-cadence and custom-silicon bets that move the threshold under you."
        },
        {
          "title": "Procurement and finance teams",
          "url": "https://costs.lyra-forge.com/for/procurement-and-finance/",
          "use": "The dated subsidy-contraction bet and its probability."
        },
        {
          "title": "Founders and product teams",
          "url": "https://costs.lyra-forge.com/for/founders-and-product-teams/",
          "use": "Which direction the inputs move over a funding cycle."
        },
        {
          "title": "Executives, investors, and broad readers",
          "url": "https://costs.lyra-forge.com/for/executives-and-investors/",
          "use": "Six dated bets, each with the condition that annuls it."
        }
      ],
      "related_urls": [
        "https://costs.lyra-forge.com/parts/budget-ai-spend/",
        "https://costs.lyra-forge.com/parts/evidence-boundaries/",
        "https://costs.lyra-forge.com/parts/run-ml-infrastructure/"
      ]
    },
    {
      "slug": "routing-layer",
      "sequence": 8,
      "kind": "decision",
      "title": "The routing layer itself \u2014 thin fees, thin moats",
      "group": "What this means for decisions",
      "question": "Is a routing layer worth its take rate, and is it a business or a feature?",
      "answer": "Routing savings are real but quality-conditioned, and the take rate charged for them is thin enough that the layer is a feature more than a moat.",
      "boundary": "Reported savings ranges come from vendor-adjacent pilots; the acceptance threshold that produced them travels with the number.",
      "html_url": "https://costs.lyra-forge.com/parts/routing-layer/",
      "source_report_url": "https://costs.lyra-forge.com/reports/the-real-cost-of-ai/",
      "source_heading": "The routing layer itself \u2014 thin fees, thin moats",
      "source_anchor_url": "https://costs.lyra-forge.com/reports/the-real-cost-of-ai/#the-routing-layer-itself-thin-fees-thin-moats",
      "audience_routes": [
        {
          "title": "Developers and ML engineers",
          "url": "https://costs.lyra-forge.com/for/developers-and-ml-engineers/",
          "use": "Whether to buy routing or build it, given how thin the fee and the moat both are."
        }
      ],
      "related_urls": [
        "https://costs.lyra-forge.com/parts/build-with-apis/",
        "https://costs.lyra-forge.com/parts/inference-economics-claims/",
        "https://costs.lyra-forge.com/parts/short-version/"
      ]
    },
    {
      "slug": "finetuning-vs-retrieval",
      "sequence": 9,
      "kind": "decision",
      "title": "If you're evaluating fine-tuning vs retrieval",
      "group": "What this means for decisions",
      "question": "When does the narrow model win, and by how much?",
      "answer": "Parameter-efficient fine-tuning is cheap in compute and expensive in everything around it; RAG and fine-tuning substitute for knowledge-heavy work and complement for behavior-heavy work.",
      "boundary": "Data preparation, evaluation, and MLOps dominate the total, and two named conditions widen the gap.",
      "html_url": "https://costs.lyra-forge.com/parts/finetuning-vs-retrieval/",
      "source_report_url": "https://costs.lyra-forge.com/reports/the-real-cost-of-ai/",
      "source_heading": "If you're evaluating fine-tuning vs retrieval",
      "source_anchor_url": "https://costs.lyra-forge.com/reports/the-real-cost-of-ai/#if-youre-evaluating-fine-tuning-vs-retrieval",
      "audience_routes": [
        {
          "title": "Researchers and data scientists",
          "url": "https://costs.lyra-forge.com/for/researchers-and-data-scientists/",
          "use": "Where the compute cost stops and the real cost starts."
        }
      ],
      "related_urls": [
        "https://costs.lyra-forge.com/parts/build-with-apis/",
        "https://costs.lyra-forge.com/parts/inference-economics-claims/",
        "https://costs.lyra-forge.com/parts/evidence-boundaries/"
      ]
    },
    {
      "slug": "evidence-boundaries",
      "sequence": 10,
      "kind": "boundary",
      "title": "Evidence boundaries",
      "group": "Evidence boundaries",
      "question": "What does this evidence not establish?",
      "answer": "Posted rates are dated, the index is a composite rather than a workload predictor, the subsidy analysis is structural on the user side, and the forecast bets are judgments over a short record.",
      "boundary": "This section is the reopen condition for every other part on the site. Read it before quoting a figure.",
      "html_url": "https://costs.lyra-forge.com/parts/evidence-boundaries/",
      "source_report_url": "https://costs.lyra-forge.com/reports/the-real-cost-of-ai/",
      "source_heading": "Evidence boundaries",
      "source_anchor_url": "https://costs.lyra-forge.com/reports/the-real-cost-of-ai/#evidence-boundaries",
      "audience_routes": [
        {
          "title": "Developers and ML engineers",
          "url": "https://costs.lyra-forge.com/for/developers-and-ml-engineers/",
          "use": "What the index-dependence caveat does to the quality-adjusted figure you are about to quote."
        },
        {
          "title": "Engineering leaders and CTOs",
          "url": "https://costs.lyra-forge.com/for/engineering-leaders/",
          "use": "Which of the six bets rest on structured priors rather than measured baselines."
        },
        {
          "title": "Platform and infrastructure teams",
          "url": "https://costs.lyra-forge.com/for/platform-and-infrastructure/",
          "use": "How far the crossover bands travel with the assumptions behind them."
        },
        {
          "title": "Researchers and data scientists",
          "url": "https://costs.lyra-forge.com/for/researchers-and-data-scientists/",
          "use": "What the fine-tuning evidence does not settle."
        },
        {
          "title": "Procurement and finance teams",
          "url": "https://costs.lyra-forge.com/for/procurement-and-finance/",
          "use": "Exactly which half of the subsidy claim is anchored and which is structural."
        },
        {
          "title": "Executives, investors, and broad readers",
          "url": "https://costs.lyra-forge.com/for/executives-and-investors/",
          "use": "The honest limit of a forecast made over a short record."
        }
      ],
      "related_urls": [
        "https://costs.lyra-forge.com/parts/short-version/",
        "https://costs.lyra-forge.com/parts/twenty-four-month-horizon/",
        "https://costs.lyra-forge.com/parts/inference-economics-claims/"
      ]
    }
  ],
  "not_published": {
    "reason": "The seven research notebooks carry visibility: private, renders_public: false, and source_trail_public: false in their own front matter. Captured sources are rights-gated. The audience ledger and the beat's working state are internal. None of it may enter a release.",
    "path_prefixes": [
      "notebooks/",
      "audience/",
      "threads/",
      "docs/",
      "research/",
      "tools/",
      ".auto/",
      "beat.md",
      "index.md"
    ]
  },
  "discovery": {
    "llms_txt": "https://costs.lyra-forge.com/llms.txt",
    "parts_registry": "https://costs.lyra-forge.com/parts.json",
    "sitemap": "https://costs.lyra-forge.com/sitemap.xml",
    "release_manifest": "https://costs.lyra-forge.com/release-manifest.json",
    "update_log": "https://costs.lyra-forge.com/backstage/updates/"
  }
}