AI Pacing Economics Dashboard

Policy π = {s, p, B, ts, te, ω} · outputs 2022–2035 · all $ in $B, compute in M H100e, Z in ECI points
Input data: embedded defaults (matches the Excel file)
Edit AI Pacing Model - Input Data.xlsx, save, then reload it here (or drag & drop the .xlsx anywhere on this box):
① Growth assumptions (the no-policy world)
Sets the growth-rate fields below (you can still tweak any of them individually). “Normal” uses the workbook’s orange 2027–29 projections; Bullish/Bearish override them.
② Policies (stack as many as you like)
“＋ Add policy” adds a blank (BAU-equivalent) policy — drag its s / p / B sliders to make it bite. Or add one of the preset policies from the dropdown. Policies stack: where they overlap in time and ω, the strictest (max) of each of s, p, B applies. ω = country-blocks the policy applies to.
③ Key parameters
Hover any label for a plain-English explanation.
④ Finance & valuation
⑤ Sector growth: market cap & debt
BAU growth of each AI subindustry's market cap (M) and debt (D), by period. These drive enterprise values E — and hence policy costs f — in later policy years.
⑥ Macro / GDP
⑦ Compute & capabilities
What is this? How do I read it? (click to expand)

What this is. A simulator of the economic cost of deliberately slowing AI development. A policy can turn off existing AI chips (s), pause AI training & R&D while letting existing AI products keep running (p), and/or pause construction of new AI datacenters (B), over a time window (ts to te) in a chosen set of regions (ω). The model traces the consequences through a chain: compute stock → AI capability progress (how many months the frontier is delayed) → costs to AI-exposed companies (delayed and destroyed profits, in today's dollars) → a one-time stock-market repricing → GDP (lost construction demand + households spending less after the market drop).

How to use it. ① Pick a growth scenario in the sidebar (Bullish / Normal / Bearish) — this sets what the world looks like without any policy. ② Add one or more policies and set their strength, timing, and regions. ③ Compare the lines: in every chart, dashed grey = business-as-usual (BAU, no policy) and solid navy = with your policy. Use the USA / China / RoW / World buttons to change whose economy you're looking at. Every input number lives in AI Pacing Model - Input Data.xlsx with its source documented — edit it, save, and reload it in the sidebar.

Key terms. H100e = compute measured in NVIDIA-H100-equivalent chips (millions). ECI = Epoch Capabilities Index, a benchmark score for the best AI model available (GPT-4 ≈ 126, frontier ≈ 155 at the start of 2026, rising ~15 pts/yr — so ~1.25 pts ≈ one month of progress). τ (tau) = how many months of AI progress a country has lost vs. no-policy. fi = the cost of the policy to sector i in net present value. The five sectors: PhysicalCompute (chips, fabs, datacenters, power — the NVIDIA/TSMC complex), NonPhysicalInputs (data, labeling, talent), AICompanies (labs like OpenAI/Anthropic + AI divisions of big tech), AIDeployers (ordinary firms profiting from using AI), NonAI (everything else). Only the AI-specific slice of each company is counted.

Main caveats. Sector values and their AI-fractions are estimates (documented in the workbook); the capability model is deliberately crude (log of cumulative training compute); costs assume value is mostly delayed rather than destroyed unless you raise α or θ; tax policy (q) is not yet implemented.

GDP — Yt, YN,t, YA,t (log)

Total GDP with (solid) vs without (dashed) the policy, plus its two pieces: the normal economy YN and the AI investment boom's contribution YA. $B/yr, log axis.

GDP drag — YD,t components

GDP lost to the policy each year: cancelled datacenter investment (direct, m·dj·B·X) + households spending less after the market drop (wealth effect, MPC·ΔW·H over 1.5 yrs).

GDP growth rate — Ẏ/Y, ẎN/YN, ẎA/YA

Annual growth of GDP and its components; the policy shows up as a dip then a rebound. YA grows fast off a small base (right axis).

Stock market — Wt

Total stock-market value, $B. When a policy is announced (ts), the market reprices down by ΔW = (1−h)·Σifi — the sectors' policy costs, unless already priced in (h).

Stock market growth rate — Ẇ/W

Annual growth of market value; the repricing appears as a one-year dip.

Capability delay vs no-policy — τt,π,j

How many months of AI progress each bloc has lost relative to its own no-policy path (RoW = China). Shrinks after the policy ends as compute growth catches back up.

AI capabilities — Zt (ECI)

Best available AI model's capability score (Epoch Capabilities Index: GPT-4 ≈ 126; ~1.25 pts ≈ 1 month of frontier progress). Solid = with policy, dashed = without.

US–China time gap — τgap,t

How many months China trails the US frontier: the time since the US was at China's current capability level (Zt−τ,USA = Zt,China). Epoch measured ~7 months on average since 2023. Clamped at 0 if China catches up.

Effective training compute — Qt,USA

Cumulative compute usable for training/R&D (∫CUSA(1−p)(1−s)df since 2022) — the input that drives the capability model. M H100e·years, log axis.

Rational cost to actors — fi (NPV, $B)

What the policy costs each part of the AI economy, in today's dollars: profits pushed into the future by the capability delay (fZ) + losses during the window itself — idle GPUs, cancelled buildout (fU).

Compute — Ct, It, Tt

Installed AI compute (C, millions of H100-equivalents, log axis) and how it's used: training+R&D (T = (1−p)ψC) vs serving existing products / inference (I = (1−ψ)C). Pauses idle T; buildout pauses flatten C.

Sector financials

AI-specific value of each sector (only the AI slice of each company). M = equity value under the policy; D = AI-attributable debt; E = M + D − cash. Dashed = all-AI total without policy.

Costs during the window — idle GPUs & foregone buildout

Running (discounted, cumulative) tally within the policy window: G = rental value of GPUs sitting idle; a = value of datacenter buildout that didn't happen. $B, for blocs the policy covers.