Hyperscalers and AI labs

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

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?