Compute Forecast Model Sketch

📋Page Status

Page Type:AI Transition ModelStyle Guide →Structured factor/scenario/parameter page

Quality:0 (Stub)

Importance:0 (Peripheral)

Structure:

📊 0📈 0🔗 0📚 0•0%Score: 2/15

LLM Summary:This page contains only a React component import with no visible content, making evaluation impossible. The actual content would need to be rendered or provided separately for assessment.

Issues (1):

• StructureNo tables or diagrams - consider adding visual content

This is a sketch of what a quantitative compute forecasting model might look like.

1. Enriched Data Structure

Instead of just qualitative causal diagrams, each node would have quantitative estimates:

computeForecastModel:
  target:
    id: effective-compute
    label: "Effective Compute for Frontier AI"
    unit: "FLOP/s (peak training)"
    current:
      value: 5e24
      date: "2024-01"
      source: "Epoch AI"
    projections:
      - year: 2027
        p10: 1e25
        p50: 5e25
        p90: 2e26
        notes: "Depends heavily on investment trajectory"
      - year: 2030
        p10: 5e25
        p50: 5e26
        p90: 5e27
      - year: 2035
        p10: 1e26
        p50: 5e27
        p90: 1e29
        notes: "Wide uncertainty; could hit physical limits or breakthrough"

  factors:
    - id: asml-capacity
      label: "ASML EUV Production"
      unit: "machines/year"
      current:
        value: 50
        date: "2024"
        source: "ASML annual report"
      projections:
        - year: 2027
          p10: 60
          p50: 80
          p90: 100
        - year: 2030
          p10: 80
          p50: 120
          p90: 180
      constraints:
        - "Factory expansion takes 3-4 years"
        - "High-NA EUV adds capacity but different machines"
      keyQuestions:
        - question: "Will ASML build a second major facility?"
          impact: "Could add 50% capacity by 2030"
        - question: "Will high-NA EUV be production-ready by 2026?"
          impact: "2-3x improvement in transistor density"

    - id: fab-capacity
      label: "Advanced Node Fab Capacity"
      unit: "wafer starts/month (3nm equivalent)"
      current:
        value: 100000
        date: "2024"
        source: "TrendForce"
      projections:
        - year: 2027
          p10: 150000
          p50: 200000
          p90: 280000
      dependsOn:
        - factor: asml-capacity
          relationship: "~2000 wafers/month per EUV machine"
          elasticity: 0.8
        - factor: power-grid
          relationship: "~100MW per major fab"
          elasticity: 0.3
        - factor: taiwan-stability
          relationship: "Disruption could remove 70% of capacity"
          elasticity: -0.9

    - id: ai-chip-production
      label: "AI Chip Production"
      unit: "H100-equivalents/year"
      current:
        value: 2000000
        date: "2024"
        source: "Estimated from NVIDIA revenue"
      projections:
        - year: 2027
          p10: 5000000
          p50: 10000000
          p90: 20000000
      dependsOn:
        - factor: fab-capacity
          relationship: "~500 chips per wafer, 30% of capacity to AI"
        - factor: ai-compute-spending
          relationship: "Demand signal drives allocation"

    - id: ai-compute-spending
      label: "AI Compute Spending"
      unit: "$/year"
      current:
        value: 100e9
        date: "2024"
        source: "Sum of major lab capex"
      projections:
        - year: 2027
          p10: 150e9
          p50: 300e9
          p90: 600e9
        - year: 2030
          p10: 200e9
          p50: 800e9
          p90: 2000e9
      dependsOn:
        - factor: ai-valuations
          relationship: "High valuations enable equity financing"
          elasticity: 0.7
        - factor: ai-revenue
          relationship: "Revenue enables sustainable spending"
          elasticity: 0.9
      keyQuestions:
        - question: "Will AI revenue justify current valuations by 2027?"
          scenarios:
            yes: "Spending continues exponential growth"
            no: "Pullback to ~$150B/year, slower growth"

    - id: algorithmic-efficiency
      label: "Algorithmic Efficiency"
      unit: "multiplier vs 2024 baseline"
      current:
        value: 1.0
        date: "2024"
      projections:
        - year: 2027
          p10: 2
          p50: 8
          p90: 30
          notes: "Historical ~4x/year, but may slow"
        - year: 2030
          p10: 5
          p50: 50
          p90: 500
      keyQuestions:
        - question: "Will efficiency gains continue at 4x/year?"
          impact: "Difference between p50 and p90"
        - question: "Is there a 'DeepSeek moment' coming?"
          impact: "Could see sudden 10x jump"

  scenarios:
    - id: base-case
      probability: 0.55
      description: "Current trends continue, moderate growth"
      assumptions:
        taiwan-stability: "No major disruption"
        ai-revenue: "Grows but below hype expectations"
        asml-capacity: "Steady expansion"
      outcome:
        effective-compute-2030: 5e26
        effective-compute-2035: 5e27

    - id: bull-case
      probability: 0.20
      description: "AI boom continues, massive investment"
      assumptions:
        taiwan-stability: "Stable"
        ai-revenue: "Exceeds expectations, clear ROI"
        asml-capacity: "Aggressive expansion"
        algorithmic-efficiency: "Continued 4x/year gains"
      outcome:
        effective-compute-2030: 2e27
        effective-compute-2035: 1e29

    - id: bear-case
      probability: 0.20
      description: "AI winter or investment pullback"
      assumptions:
        ai-revenue: "Disappoints, valuations crash"
        ai-compute-spending: "Drops 50%"
      outcome:
        effective-compute-2030: 1e26
        effective-compute-2035: 5e26

    - id: disruption-case
      probability: 0.05
      description: "Major supply shock (Taiwan, other)"
      assumptions:
        taiwan-stability: "Major disruption"
        fab-capacity: "Drops 50-70%"
      outcome:
        effective-compute-2030: 2e25
        effective-compute-2035: 1e26

2. Squiggle Model

Here's what the actual quantitative model might look like in Squiggle:

// === INPUT PARAMETERS ===

// ASML EUV machine production (machines/year)
asmlProduction2024 = 50
asmlGrowthRate = normal(0.08, 0.03)  // 8% ± 3% annual growth
asmlProduction(year) = asmlProduction2024 * (1 + asmlGrowthRate)^(year - 2024)

// Wafers per EUV machine per year
wafersPerMachine = normal(24000, 3000)  // ~2000/month

// Advanced fab capacity (wafer starts/year, 3nm equivalent)
fabCapacity(year) = asmlProduction(year) * wafersPerMachine * 0.7  // 70% utilization

// Taiwan risk - probability of major disruption by year
taiwanDisruptionProb(year) = 0.02 * (year - 2024)  // 2% per year cumulative
taiwanImpact = beta(2, 8)  // If disruption, lose 20-80% capacity
taiwanMultiplier(year) = if bernoulli(taiwanDisruptionProb(year)) then (1 - taiwanImpact) else 1

// AI chips per wafer
chipsPerWafer = normal(400, 50)

// Fraction of advanced capacity going to AI chips
aiCapacityShare2024 = 0.25
aiCapacityShareGrowth = normal(0.03, 0.01)  // Growing 3% per year
aiCapacityShare(year) = min(0.6, aiCapacityShare2024 + aiCapacityShareGrowth * (year - 2024))

// AI chip production (H100-equivalents/year)
aiChipProduction(year) = {
  baseProduction = fabCapacity(year) * chipsPerWafer * aiCapacityShare(year)
  baseProduction * taiwanMultiplier(year)
}

// FLOPS per chip (H100 = 2e15 FLOPS for training)
flopsPerChip2024 = 2e15
chipImprovementRate = normal(0.25, 0.08)  // 25% per year Moore's law continuation
flopsPerChip(year) = flopsPerChip2024 * (1 + chipImprovementRate)^(year - 2024)

// Algorithmic efficiency multiplier
algoEfficiency2024 = 1.0
algoEfficiencyGrowth = lognormal(1.4, 0.5)  // ~4x/year but high variance
algoEfficiency(year) = algoEfficiency2024 * algoEfficiencyGrowth^(year - 2024)

// AI company revenue and investment
aiRevenue2024 = 200e9  // $200B
revenueGrowthRate = normal(0.20, 0.10)  // 20% ± 10% annual growth
aiRevenue(year) = aiRevenue2024 * (1 + revenueGrowthRate)^(year - 2024)

// Investment as fraction of revenue/valuation
investmentRate = beta(3, 7)  // 20-40% of revenue goes to compute
aiComputeSpending(year) = aiRevenue(year) * investmentRate * 1.5  // 1.5x for valuation leverage

// Utilization rate (what fraction of chips are used for frontier training)
utilizationRate = beta(5, 5)  // ~50% utilization

// === MAIN MODEL ===

// Total AI chip FLOPS available
totalChipFlops(year) = {
  // Stock of chips (assume 3 year lifespan, accumulating)
  stock = sum(
    List.map(
      List.range(max(2024, year - 3), year),
      y -> aiChipProduction(y) * flopsPerChip(y)
    )
  )
  stock * utilizationRate
}

// Effective compute (accounting for algorithmic efficiency)
effectiveCompute(year) = totalChipFlops(year) * algoEfficiency(year)

// === OUTPUTS ===

effectiveCompute2027 = effectiveCompute(2027)
effectiveCompute2030 = effectiveCompute(2030)
effectiveCompute2035 = effectiveCompute(2035)

// Training run size (largest single run, ~10% of total capacity)
largestTrainingRun(year) = effectiveCompute(year) * 0.1 * (365 * 24 * 3600)  // FLOP per run

// === KEY METRICS ===

// Years to 10x current compute
yearsTo10x = {
  current = effectiveCompute(2024)
  target = current * 10
  // Find year where we cross threshold
  List.findIndex(
    List.map(List.range(2024, 2040), y -> effectiveCompute(y) > target),
    x -> x
  )
}

3. What This Enables

With this structure, you could:

1. Generate probabilistic forecasts - not just point estimates
2. Run sensitivity analysis - which inputs matter most?
3. Scenario modeling - what if Taiwan is disrupted? What if AI revenue disappoints?
4. Update on evidence - new ASML numbers? Update the model
5. Identify cruxes - where do optimists and pessimists disagree?

4. Key Uncertainties Ranked by Impact

Factor	Impact on 2030 Compute	Current Uncertainty
Algorithmic efficiency	10-100x range	Very high
Taiwan stability	0.3-1.0x	Low prob, high impact
AI revenue/investment	2-5x range	High
ASML expansion	1.5-2x range	Medium
Chip architecture	2-4x range	Medium

5. Integration with Diagram

The causal diagram could become an interface to this model:

• Click a node → see current estimate, distribution, sources
• Hover over edge → see elasticity/relationship strength
• Scenario selector → see how diagram changes under different assumptions
• Time slider → see which bottlenecks dominate when

---