Athena

Summary

Athena is a four-engine observatory tracking the economics of silicon, frontier compute models, AI markets, and policy across roughly 195 economies. Around 60 collectors feed a multi-entity warehouse that resolves observations by series, entity, and date, with explicit compute-unit discipline for FLOPs, token prices, energy, and carbon. Every series is sourced and every brief claim cites a row in the warehouse.

What it is

Athena is a working research platform built to answer the economic questions raised by AI buildout: what compute costs, who controls it, how fast it scales, what it draws from the grid, and which policies bend the curve. The system is organized into four engines. Silicon covers chip families, foundries, and trade controls. Compute models covers frontier labs, training runs, and inference economics. Markets covers hyperscaler capex, cloud regions, and venture funding. Policy and macro covers regulation, talent flows, and national strategies. The four engines never import from one another; cross-engine signal moves through a single bus, which keeps the architecture auditable and the data lineage clean.

For client engagements, Athena functions as private research infrastructure rather than a product. Teams use it to size training and inference cost curves, stress-test grid and water exposure for new data center footprints, benchmark fab and packaging capacity against announced demand, and pressure-test capex assumptions in hyperscaler and AI-adjacent diligence. Outputs are reproducible: each figure, claim, and chart is traceable to a row in the observations or results tables, with method tags for training cost estimates, separate token prices for input and output, energy in kWh, and carbon in tCO2e with scope. Engagements typically combine a custom data extract, a written brief, and a working session with the underlying warehouse so client teams can extend the analysis on their own.

Methodology

• Four-engine framework separating silicon, compute models, markets, and policy or macro signal, with no cross-engine imports and a single internal bus for shared state.
• FLOPs and token-pricing methodology: base-10 FLOP floats, training cost estimates carry a method tag, and token prices are tracked separately per one million input and one million output tokens.
• Multi-entity warehouse schema keying observations by series, entity, and date, with entity types spanning country, model, chip, fab, region, hyperscaler, and provider.
• Around 60 collectors run on staggered UTC schedules through APScheduler, writing to a single SQLite warehouse in WAL mode for reproducible point-in-time reads.
• Energy and carbon discipline: kWh for energy, tCO2e for carbon with scope tagging, and ISO3 country codes normalized to uppercase before any cross-country query.
• Series-level analytics including AR(1) persistence and half-life, year-over-year growth, volatility and shock detection, summary CAGR, and cross-country comparator rankings.
• Coverage of roughly 195 economies, with country-level and entity-level extracts available as long-format panels for downstream modeling.
• Reproducible briefs assembled with Jinja2 templates, where every claim cites a row id in the observations or results tables.

Data sources

• MLPerf training and inference benchmark results
• SemiAnalysis chip, fab, and hyperscaler intelligence
• Epoch AI compute and frontier model database
• U.S. Energy Information Administration (EIA) electricity and fuels data
• Federal Energy Regulatory Commission (FERC) interconnection and capacity filings
• U.S. Bureau of Labor Statistics (BLS) labor and wage series
• UN Comtrade bilateral trade flows for semiconductors and equipment
• Hyperscaler capex disclosures from quarterly filings and investor materials

Deliverables when used in engagements

• Custom data extracts as long-format panels across countries, chips, models, and hyperscalers.
• Written briefs with every figure and claim cited to a warehouse row id.
• Compute and inference cost models with method tags and sensitivity ranges.
• Grid, water, and siting risk reads for proposed or expanding data center footprints.
• Working session and handoff so the client team can re-run and extend the analysis.