Five anchor patterns. One $/M-token.

VP Infrastructure at a Series B neocloud, lease-financed, going live in 18 months.

Committing $400M of B200 and GB300 silicon across the next four quarters. CFO wants $/M-token sensitivity at 60%, 75%, and 90% sustained utilisation. Sales wants quotable rates for the top five model SKUs the platform will serve. Procurement needs BOM specificity to lock vendor contracts before the tariff window closes.

• Procurement-ready BOM down to leaf parts, scoped to the MW envelope.
• Quotable $/M-token sheet by model, with utilisation-sensitivity bands the CFO will sign.
• Lease vs CapEx break-even per chip family, decision-ready by category.

Head of Platform at a Series C AI SaaS, ~200 employees, $6M/month cloud bill.

Inference traffic growing 30% quarter over quarter. Board has approved a partial in-house build; the team has to recommend which workloads to migrate, on which chip, and on what financing structure. Cost of getting it wrong is a 36-month lease against the wrong silicon.

• Eight-quarter token-throughput forecast per active workload, with redeployment hooks for retired silicon.
• Head-to-head sizing across H100, H200, and B200 on the actual model mix and precision profile.
• Break-even timeline per workload across cloud, 36-month lease, and CapEx — the financing decision tree, not just a TCO.

Programme director for a national AI initiative; budget approval cycle on a Q3 deadline.

Vendor list is restricted by sovereignty and export-control constraints — half the commercial chip market is unavailable. The team has to size a 20MW research cluster within those constraints and quantify the cost of constraint relative to an unrestricted baseline.

• Optimal sovereign-compliant BOM, scoped to the approved vendor list.
• Tokens-per-MW under restriction vs the unrestricted reference architecture — the explicit cost of constraint.
• Procurement spec ready to attach to the budget submission.

Director of Strategic Sourcing buying AI infra across 5+ chip vendors annually.

Every vendor shows their own benchmarks on their own workload mix. The team needs an apples-to-apples comparator using internal workload distribution and internal MFU assumptions, defensible enough to set negotiation floors that procurement can hold across a multi-year buying cycle.

• Vendor-neutral RFP scorecard across the full approved-chip set.
• $/chip-hour and $/FLOP comparison under internal workload mix and MFU assumptions, not vendor defaults.
• Negotiation-floor memo with the workload-anchored evidence to back it.

Head of Commercial at a compute-as-a-service aggregator, six models on the public price card.

Every customer conversation ends with a request for a sharper price. Without a clean view of the floor across cloud / lease / CapEx delivery, the team is either leaving margin on the table or quoting under cost. They need a structured margin map by chip × model × contract, refreshed whenever silicon or rates move.

• Margin matrix by chip × model × contract — the explicit floor across delivery modes.
• Sensitivity to spot vs reserved capacity, surfaced as a price-band rather than a single number.
• Refresh-on-change: matrix recomputes when chips, contracts, or model presets change.