{
  "asOf": "2026-09-13",
  "classification": "New house research scenario; numerical assumptions prepared for this edition, not previously published author forecasts or consensus",
  "headline": "Cheaper intelligence expands use. The bottleneck moves to delivery and monetization.",
  "basis": "Illustrative AI inference workload, 2026 = 100. These are fixed research assumptions, not measurements of the whole AI economy.",
  "years": [2026, 2027, 2028, 2029],
  "cases": [
    {"id":"base","name":"Base case","thesis":"Useful task growth outpaces efficiency, but compute prices fall and power limits the speed of commissioning.","tasks":[100,180,300,480],"computePerTask":[100,65,46,34],"computePrice":[100,95,88,82],"utilization":[70,72,74,75]},
    {"id":"upside","name":"Adoption upside","thesis":"Agents become reliable enough for longer workflows; growing paid use more than offsets lower unit prices.","tasks":[100,220,430,800],"computePerTask":[100,68,50,38],"computePrice":[100,100,96,92],"utilization":[70,76,80,82]},
    {"id":"downside","name":"Monetization downside","thesis":"Efficiency improves faster than paid adoption, leaving expensive capacity underutilized and pressuring service prices.","tasks":[100,140,190,250],"computePerTask":[100,58,36,24],"computePrice":[100,88,76,65],"utilization":[70,64,60,58]}
  ],
  "equations": [
    "Required compute index = useful task volume index × compute per task index ÷ 100.",
    "Service revenue index = required compute index × compute-price index ÷ 100. Task volume already describes consumed work; utilization is not multiplied into revenue a second time.",
    "Installed capacity index = required compute index ÷ (utilization ÷ 70). Lower utilization requires more installed capacity for the same consumed workload.",
    "No EPS or stock return is mechanically inferred from these indices. Hardware mix, depreciation, operating cost, reinvestment and entry valuation remain separate underwriting questions."
  ],
  "calls": [
    {"layer":"Applications & agents","view":"Select the workflow owner","reason":"Distribution, proprietary context and verification can retain value as model access becomes cheaper.","watch":"Paid conversion, repeat usage and accepted-task cost; usage alone is not revenue.","exposures":["MSFT","CRM","NOW","ADBE"],"invalidator":"Usage expands while customer retention, willingness to pay or contribution profit deteriorates."},
    {"layer":"Compute & custom silicon","view":"Volume up; platform mix changes","reason":"Our base case requires more compute despite lower compute per task. Internal chips can change who captures the spending.","watch":"Deployed throughput, custom-chip adoption, gross margin and customer concentration.","exposures":["NVDA","AMD","AVGO","TSM"],"invalidator":"Efficiency exceeds task growth, or capacity utilization and service pricing fall together."},
    {"layer":"Memory & equipment","view":"Content and qualification matter","reason":"The original memory thesis focuses on HBM content and the subsequent equipment cycle; stronger demand still has to become qualified shipments and cash.","watch":"HBM4 qualification, bit supply, realized pricing and actual WFE orders.","exposures":["MU","AMAT","LRCX","KLAC"],"invalidator":"Competing supply weakens pricing or customer equipment budgets turn down.","originalSource":"/research#semicap"},
    {"layer":"Power, cooling & construction","view":"Delivery is the constraint","reason":"Electrical and thermal capacity must arrive on time. Announced megawatts are not energized megawatts.","watch":"Order conversion, commissioning dates, project margins and free cash flow.","exposures":["GEV","ETN","VRT","EME"],"invalidator":"Backlog cancellations, delivery delays or lower cash margins overwhelm order growth.","originalSource":"/newsletters/issue-7.pdf"}
  ],
  "externalContext": [
    {"label":"Data-centre electricity","value":"485 → 950 TWh","period":"2025 to 2030","basis":"IEA 2026 outlook · global data centres, not AI-only","source":"https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary"},
    {"label":"NVIDIA data-centre revenue","value":"$89.0bn","period":"Quarter ended 26 Jul 2026","basis":"Company reported · fiscal Q2 2027","source":"https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-second-quarter-fiscal-2027"}
  ],
  "sources": [
    {"label":"Original memory and equipment thesis","url":"/research#semicap","date":"2026-08-25"},
    {"label":"Original U.S. power research","url":"/newsletters/issue-7.pdf","date":"2026-04-24"},
    {"label":"IEA · Key Questions on Energy and AI","url":"https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary","date":"2026"},
    {"label":"NVIDIA fiscal Q2 2027 results","url":"https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-second-quarter-fiscal-2027","date":"2026-08-26"},
    {"label":"Frontier efficiency and scaling research","url":"/ai/frontier","date":"2026-09-13"}
  ]
}