Industry

Hyperscalers and AI labs

Compute economics, capex ROI, and grid feasibility analytics for the operators running the AI build cycle.

Framing

Hyperscaler capex crossed half a trillion dollars in 2025 to 2026. The next dollar of compute is increasingly bottlenecked not by silicon but by the megawatt curve, the interconnection queue, and the token-pricing dynamics of the inference market. Internal economics teams have the data; what they do not always have is the time to run a defensible workload-weighted ROI model against a procurement decision that has to be defended to a CFO this quarter.

The work here is the bench capacity for that math. Effective FLOPs per dollar across chip generations, normalized for utilization. Token cost curves under named model-class assumptions. Regional capacity bidding analyses across PJM, ERCOT, and the Phoenix corridor. AI policy economics around CHIPS, DOE loan guarantees, and state incentive stacking. Public-data, replicable, and tied to MLPerf, EIA, FERC, SemiAnalysis, and Epoch AI inputs.

The consultancy is small, which makes it useful for the targeted analytical questions the in-house team needs answered in two weeks, not the multi-month engagements that compete with the team's own roadmap.

Decisions we inform

  • Generation-over-generation compute purchase decisions (H100 vs B200 vs GB200 NVL72).
  • Cloud versus owned versus colo decisions across workload mixes and procurement postures.
  • Regional siting of new training or inference capacity against grid and policy constraints.
  • Inference unit-economics under sustained price compression in the API market.
  • CHIPS Act, DOE, and state incentive stacking for new capacity announcements.
  • Internal pricing and chargeback models for AI compute across product lines.

Named offerings

Compute Curve Analysis

Workload-specific effective FLOPs per dollar across two to four chip generations under named utilization, procurement, and useful-life assumptions. Memo plus model.

Token Cost Curve

Inference unit economics for a frontier or open-weight model class. Per-million input and output token decomposition across hardware, kernel maturity, batching, and quantization. Repricing scenarios.

Regional Siting Diagnostic

Capacity expansion siting analysis: ISO interconnect queue posture, PPA market, permitting timeline, water and tax incentive stack across two to five candidate metros.

Policy and Incentive Map

Stackable incentive landscape for a planned investment: CHIPS, DOE loan programs, Section 48D, state grants, utility riders. Deltas under named legislative scenarios.

AI Productivity Memo

Sector-specific TFP and labor-augmentation analysis tied to BLS productivity, BEA TFP, and adoption survey data. Useful for AI policy and earnings-call positioning.

Sample questions

  • We are sizing a 200 megawatt training cluster expansion. GB200 NVL72 versus extended H100 capacity at our utilization profile, what is the honest ROI delta?
  • Our inference platform is repricing every 90 days. Where does the cost curve break if we drop quantization to FP4 across half the workload?
  • Northern Virginia or Phoenix or West Texas for the next 500 megawatt build? Interconnect, PPA, and water in one memo.
  • What does the CHIPS Section 48D plus DOE Title XVII stack look like for the advanced packaging line we are scoping in Arizona?
  • Our policy team needs a defensible TFP model for AI in financial services. Can you build the decomposition against BLS and BEA data?