AI Infrastructure Unit-Economics Director — Multimodal Platform

The mandate

The platform cannot reconcile reported product margin with the economics of reserved GPUs, burst cloud capacity and colocated inference clusters. Shared training runs, scheduler idle time, failed generations and customer-specific serving configurations disappear into broad infrastructure pools, leaving the board unable to distinguish scale investment from structural margin leakage.

The commissioned deliverable is an AI Infrastructure Unit-Economics Decision Book comprising a ledger-reconciled GPU-hour baseline, workload cost curves, customer contribution model, capacity-placement rules and pricing guardrails. It must allow Finance and Engineering to trace a quoted gross margin from customer invoice through tokens, accelerator occupancy, storage, networking, orchestration and committed-capacity exposure.

The engagement begins on 14 December 2026. Milestone one, due 8 January 2027, is the signed source reconciliation and data-gap register; milestone two on 12 February is the workload cost atlas; milestone three on 19 March is the placement and commercial intervention case; and milestone four on 16 April is the board decision book and controlled calculation model.

Acceptance requires the CFO’s team to reperform sampled cost allocations within two percentage points, the CTO’s capacity leads to reconstruct contribution margins for the twenty largest accounts within a five-percent variance, and both sponsors to approve every unresolved estimation convention. Board use of the model in the financing plan completes acceptance; procurement execution remains outside the statement of work.

The client will provide twelve months of invoices, reservations, power and colocation costs, scheduler telemetry, model-serving logs, billing records, product catalogues and customer discount schedules inside a secure environment. A finance controller, capacity engineer, data analyst and executive sponsor will be available weekly, with named owners responsible for clearing access failures within two business days.

Why this is external work

Engineering, Finance and Sales each own a valid fragment of the economics, but none can arbitrate the allocation choices without defending its current decisions. The board also needs a specialist view of accelerator utilisation and inference cost behaviour that the internal FP&A model does not contain. External authorship gives the financing case a neutral evidence trail.

What you will own

• Reconcile cloud, colocation, energy, networking and accelerator invoices to consumed and reserved compute units, documenting every allocation convention and residual variance.
• Segment training, fine-tuning, batch inference, low-latency serving and failed workloads into cost behaviours that remain stable under volume changes.
• Construct workload curves showing utilisation breakpoints, memory constraints, queue penalties, model-specific token economics and the cost of service-level commitments.
• Quantify reserved-capacity, cloud-burst, hardware-purchase and managed-inference scenarios under demand, price, failure-rate and obsolescence sensitivities.
• Design customer and product guardrails for minimum contribution, context-window pricing, premium latency, multimodal processing and bespoke deployment configurations.
• Draft the board decision pack connecting each infrastructure or pricing intervention to cash use, gross margin, execution dependency and downside exposure.
• Maintain a reproducible evidence room containing source extracts, transformation logic, reconciliations, assumption approvals and instructions for monthly model refresh.

Candidate qualifications

• Led a board-visible unit-economics programme for an AI platform spending at least US$250 million annually on training and inference infrastructure.
• Reconciled GPU scheduler or Kubernetes telemetry to cloud, colocation and hardware ledgers without treating contracted capacity as fully utilised consumption.
• Modelled the distinct economics of training, fine-tuning, batch workloads and latency-constrained inference across more than one accelerator architecture.
• Changed pricing, workload placement or reservation strategy using an analysis whose realised margin effect was measured after implementation.
• Built a controlled financial model that FP&A could refresh and auditors or transaction advisers could reperform from retained source evidence.
• Advised executive teams through infrastructure decisions where hardware scarcity, customer service promises and rapid model obsolescence produced materially different answers.

Non-negotiables

• The named engagement lead must perform the core analysis and attend every sponsor and board review rather than delegating delivery to a junior team.
• On-site availability in San Francisco is required for the opening two weeks and each milestone review, with travel to Santa Clara and Phoenix.
• No reseller margin, cloud-provider incentive, hardware referral fee or success payment may influence the placement recommendation.
• All calculations must remain inside the client’s controlled environment, with no production telemetry copied into external benchmarking tools.

1. 49 words maximum. Describe how you would establish a ledger-to-GPU cost baseline during the first three weeks when scheduler and invoice identifiers disagree.
2. 49 words maximum. Give one example where utilisation data changed an AI infrastructure purchase or reservation decision and quantify the result.
3. 49 words maximum. Disclose any current commercial relationship with a cloud, accelerator, colocation or FinOps provider relevant to this engagement.