concept

AI Economic Activity Index

A global index tracking AI compute activity in real-time, measuring token volumes, spend, and energy consumption across major providers—modeled on how the IMF tracks currency reserves.

Type: concept · Role: Creator · Duration: 2025–present · Scale: 64 models tracked daily · Status: Live

There’s no reliable way to measure how much AI activity is happening globally. Hyperscalers don’t disclose token volumes. Energy consumption gets lumped into “datacenter” categories. Investment analysts are guessing.

So I built one.

The AIU

The AI Economic Activity Index tracks global AI economic activity through a synthetic unit called the AIU (AI Unit). The concept borrows from the IMF’s Special Drawing Rights—a synthetic currency basket that combines USD, EUR, GBP, JPY, and CNY to measure something more stable than any single currency.

The AIU works the same way, but for AI compute:

Component	Weight	What it measures
Token Volumes	60%	Weekly tokens processed across 64+ models
Inferred Spend	30%	Calculated from volumes × blended pricing
Energy Proxy	10%	Estimated from volumes and model efficiency tiers

The baseline is February 2025 = 100. As activity grows or shrinks, the index moves.

Why this matters

The AEAI is a volume index, not a price index. It answers: “How much AI work is happening?” not “What does AI cost?”

This makes it complementary to the Compute CPI (which tracks price deflation). Together they tell you:

Compute CPI falling = AI getting cheaper per unit
AEAI rising = More AI work being done
Both together = Total spend = AIU × CPI

If AIU grows 50% and CPI falls 20%, total spending grows ~20%. That’s the kind of analysis that wasn’t possible before.

The data pipeline

Everything runs automatically via GitHub Actions:

Rankings data pulls from OpenRouter’s public API (64+ models, updated daily)
Pricing data aggregates from OpenRouter, LiteLLM, llm-prices.com, and pricepertoken.com
Energy factors apply tier-based estimates (Budget → Frontier → Reasoning)
Calculator runs daily at 6 AM UTC, commits results to the data directory

The output is a JSON snapshot with full methodology transparency:

{
  "aiu_index": 100.0,
  "components": {
    "token_index": 100.0,
    "spend_index": 100.0,
    "energy_index": 100.0
  },
  "activity": {
    "tokens_weekly": 13909678380565,
    "spend_usd_weekly": 46629717.81,
    "energy_gwh_weekly": 0.000591
  }
}

What I’m measuring (and what I’m not)

Current coverage captures activity visible through OpenRouter—roughly 5-10% of estimated global AI inference. That’s a significant sample but not comprehensive.

Not captured:

Direct API usage (OpenAI, Anthropic, Google direct customers)
Enterprise deployments and private instances
Regional providers not routed through OpenRouter

This is by design. The goal isn’t perfect measurement—it’s creating a consistent, automated baseline that can track changes over time. When activity doubles, the AEAI will show it, even if the absolute number is an undercount.

Planned extensions

Geographic attribution: Track AIU by region (US, EU, APAC) to see where AI activity is concentrating.

Subindices: $ACTIVITY-CODE for coding workloads (Copilot, Cursor), $ACTIVITY-CHAT for conversational AI, $ACTIVITY-AGENT for agentic workloads.

Better energy data: Integrate IEA datacenter baselines and hyperscaler sustainability reports as they become available.

Use cases

Investment analysis: Gauge AI infrastructure demand independent of stock prices
Policy planning: Understand AI’s economic footprint for energy and compute policy
Market sizing: Track AI compute market growth with real data
Trend analysis: Identify acceleration or deceleration in AI adoption