AI and compute economics

Summary

AI infrastructure is the largest private-sector investment cycle since the buildout of the electrical grid. This practice produces the upside math: who captures the productivity gains, what the ROI looks like across chip generations, which regions translate investment into durable growth, and how to structure the energy infrastructure that makes it possible.

What we do

Hyperscaler capital spending crossed half a trillion dollars in 2025 and 2026. Effective compute cost per FLOP is halving roughly every 18 months. CHIPS Act awards, DOE loan guarantees, and state incentive packages are rewiring where semiconductors and data centers get built. Decisions about capex, procurement, siting, and pricing in this environment require workload-specific math, not spec-sheet ratios.

The Athena platform is the technical backbone: 60 collectors covering chip and GPU manufacturers, cloud providers, training labs, MLPerf benchmarks, and policy databases. Engagements deliver effective FLOPs per dollar curves under named utilization, procurement, and useful-life assumptions. Token cost curves for inference. Regional siting analyses against grid and policy constraints. Sector-specific TFP and labor-augmentation analyses tied to BLS productivity and BEA TFP decompositions.

Methods

• Effective FLOPs per dollar normalization across chip generations
• Token cost decomposition (input, output, batching, quantization)
• Workload-weighted utilization analysis
• MLPerf result normalization to dense FP8 throughput
• TFP and labor-augmentation decomposition (BLS, BEA)
• ISO interconnect queue parsing
• PPA pricing and behind-the-meter analysis
• Industrial policy stacking analysis (CHIPS, DOE, IRA)

Public data partners

• MLPerf training and inference results
• SemiAnalysis chip and cloud pricing
• Epoch AI compute trends
• BLS labor productivity, BEA TFP and capital stock
• EIA hourly generation and demand
• FERC, NERC, and ISO/RTO data (PJM, ERCOT, MISO, CAISO, SPP, ISO-NE, NYISO)
• CHIPS Act awards, DOE loan program data
• BACI HS 8541 and 8542 trade flows

Deliverables

• Compute cost-curve benchmark pack across GPUs, TPUs, and custom accelerators
• Productivity impact memo tied to BLS labor productivity and BEA TFP decompositions
• Capex ROI model with scenario trees for chip cost, utilization, and cloud pricing
• Regional growth dashboard for a specific state, metro, or utility service area
• Industrial policy brief on incentive stacking

Sample engagements

• GB200 NVL72 versus extended H100 ROI for a hyperscaler training cluster.
• Token cost curve for a frontier model class under named procurement scenarios.
• Regional siting diagnostic across PJM, ERCOT, and Phoenix corridor.
• CHIPS plus DOE plus state incentive stacking for a planned advanced packaging line.
• Quarterly AI compute brief subscription for a sovereign wealth fund.