AI & Cloud Infrastructure Finance

Why: a company's position in the stack determines its COGS structure. Learn what's actually being bought.

• The stack: IaaS → PaaS → SaaS — what the customer manages at each layer, and why position determines cost structure
• The three cost primitives: compute (instance families, GPU vs. general), storage (hot→archive tiers, ~20× spread), and network/egress (the deliberate lock-in asymmetry)
• Pricing mechanics: on-demand vs. Reserved Instances/Savings Plans (30–60% off for 1–3-yr commits) vs. spot (60–90% off, interruptible) vs. enterprise agreements (EDP/CUD/MACC) — and FP&A's role in sizing commitments
• Fixed vs. variable in practice: the two-layer model (committed base + variable burst) that separates credible COGS forecasts from naive ones
• Multi-cloud & repatriation: when owning beats renting (~60–70%+ steady utilization over 3+ yrs) and why most startups rent while the biggest players buy

• AWS EC2 pricing, S3 pricing, and data transfer pricing — one sentence per pricing dimension
• a16z — The Cost of Cloud, a Trillion Dollar Paradox (read critically)
• One hyperscaler earnings call: Amazon IR (AWS segment) or Microsoft IR (Azure segment)
• Cloudflare on egress economics — the contrarian view of cloud pricing

Why: FinOps sits exactly at the FP&A × engineering intersection — fluency here is a differentiator in almost every AI/SaaS finance interview.

• The operating model: Inform → Optimize → Operate; making cost a first-class engineering metric rather than policing it after the fact
• Cost allocation: tagging strategy (<5% untagged at maturity), shared-cost allocation methods (proportional, per-seat, per-request) and their weaknesses, showback vs. chargeback
• Unit economics: cost per customer/transaction/API call as the number that makes spend interpretable; the volume-vs-unit-cost-vs-mix decomposition
• Optimization levers, ranked: commitment coverage (70–85% of steady state, finance-led) → rightsizing (10–25%) → storage lifecycle → architecture changes — each with an owner and an ROI
• Forecasting cloud spend: driver-based on units, step changes, and the asymmetric cost of commitment forecast error

• FinOps Foundation — Framework and unit economics
• State of FinOps survey — benchmarks for coverage, tagging, team structure
• AWS Well-Architected — Cost Optimization Pillar (or CloudZero methodology posts)
• GitLab public finance handbook — real documented allocation methodology

Why: for any AI application company, inference is the COGS that scales with revenue — and FP&A must model it per unit.

• The AI cost taxonomy: training (capex-like, frontier labs only) vs. fine-tuning vs. inference vs. per-customer customization
• Inference mechanics: tokens, input vs. output pricing (output ~3–5× input), cost per interaction, and the ~100× price spread across models
• The cost levers: model routing (the biggest single lever), prompt caching, batching, context discipline, quantization/distillation
• The deflation curve: ~10×/yr cost decline per unit of capability ("LLMflation") — and the strategic choice of banking it as margin, reinvesting in capability, or passing it through in price
• API vs. self-hosted: pure-variable COGS vs. GPU capacity planning and utilization risk — a repatriation problem one level up
• GPU economics: per-GPU-hour markets, the utilization math (40% utilization = 2.5× effective unit cost), and the live depreciation controversy
• The multimodal stack: STT + LLM + TTS + telephony + human escalation; automation rate as the master margin lever in AI-plus-human businesses

• a16z — LLMflation and their application-layer AI economics essays
• Sequoia — AI's $600B Question
• Anthropic pricing and OpenAI API pricing — build live-estimation fluency
• One inference-optimization explainer (batching, caching, quantization) — conceptual fluency only
• SemiAnalysis — one or two GPU-cloud economics articles; don't fall down the rabbit hole

Why: gross margin is where infrastructure meets the P&L — this module turns Modules 1–3 into the artifact boards actually ask for.

• COGS methodology: what goes above the line, AI additions (inference, HITL labor, fine-tuning), and the classification judgment calls worth 10+ margin points (ML team? free tier? customer success?)
• Benchmarks to memorize: pure SaaS 75–85% · AI/infra-heavy 40–60% with a credible path up · usage-based infra 45–60% structural · services 20–40% — always segment product vs. services
• The margin bridge: a waterfall where every point of change is attributed to exactly one named driver with an owner and a plan
• Cost-to-serve: customer-level margin, the 10–20% margin-negative customer discovery, and subsidized-margin traps (credits, discounts, parent funding)
• Margin as a system: revenue mix × unit costs × utilization × efficiency roadmap, built driver-based

• Two or three S-1s/10-Ks with AI or usage-heavy COGS via SEC EDGAR — read the COGS and margin discussions closely
• Bessemer — State of the Cloud and the BVP Cloud Index metric definitions
• Jamin Ball — Clouded Judgement — 5 recent issues for current benchmarks
• Meritech public comps for margin benchmarking by category

Why: AI is pushing pricing from seats toward usage and outcomes — forecasting these is a different sport from classic SaaS ARR.

• The pricing spectrum: seats → platform + usage → pure consumption → outcome-based, and why AI agents accelerate the shift (they replace labor, not seats)
• Metrics that change: the three ARRs (run-rate, committed, reported) and their games; NRR as master metric (120–150% possible, fragile downward); cohort ramps and the MRR waterfall
• Forecasting consumption: bottoms-up cohort ramps, the bookings → implementation → ramp → steady-state lag chain, and leading indicators that predict revenue 1–2 quarters out
• Unit economics of growth: CAC payback and LTV complications under usage pricing; burn multiple bands (<1 great, 1–2 good, >2 concerning); Rule of 40 for AI-native companies

• Investor materials of two public usage-based companies — how they define NRR, RPO, consumption metrics
• OpenView / ICONIQ Growth / Bessemer usage-based pricing benchmarks
• David Sacks — The Burn Multiple
• One deep piece on outcome-based pricing for AI agents (a16z or Bessemer)

Why: this is where the models start making the company money.

• Pricing AI products: anchoring on labor displaced, price-metric selection, the deflation-pass-through decision, and discount discipline (commit-based vs. quarter-end leaks)
• Capacity & commit sizing: expected-value math of over- vs. under-commit against a probabilistic forecast; commit to the P80 base, stay flexible on the rest
• Build vs. buy vs. rent: one framework for APIs vs. self-hosting, cloud vs. hardware, tools vs. internal builds — fixed-cost absorption, utilization risk, strategic control, switching costs
• Investment cases & resource allocation: efficiency work as NPV-ranked projects; headcount planning as the real budget; M&A screening basics

• Madhavan Ramanujam — Monetizing Innovation (core chapters on price-metric selection)
• Pricing pages of 3–4 AI products across the spectrum — reverse-engineer the metric choice
• Mostly Metrics and OnlyCFO — practitioner takes on AI-era finance decisions
• Re-read the a16z cloud-cost essay through a decision lens: as CFO, what would you do?

Why: core FP&A craft, adapted to growth rates that break normal planning cadences.

• The integrated operating model: pipeline → bookings → revenue → unit-cost COGS → headcount-driven opex → cash; every assumption a labeled input
• Scenario architecture: base/bull/bear as coherent narratives that move hiring, commits, and runway together; rolling forecasts over dead annual plans
• Runway & burn management: the hiring plan as the main throttle; spending as a milestone-gated sequence, not a level
• Board reporting: the standard growth-stage pack, driver-based variance commentary, and metrics governance (one definitions doc, one source of truth)
• AI-augmented FP&A practice: concrete workflows for commentary drafting, documentation, scenario narratives — with a crisp view of where human judgment stays

• GitLab finance handbook — planning and reporting sections; the best public artifact of real FP&A process
• Board deck templates and teardowns (Sequoia and VC blogs)
• Mostly Metrics / OnlyCFO archives on operating models and board reporting
• Your own past models and reporting packs — inventory and repackage with better vocabulary and benchmarks

Why: not to become a software engineer — to read, modify, debug, and direct AI-generated code with confidence.

• What vibe coding actually is: specify precisely → AI writes → you verify, debug, iterate; the honesty bar of never shipping a number from code you can't explain
• Minimum viable Python: control flow, functions, error handling, lists/dicts, file I/O and JSON, and pandas — the workhorse that replaces half of manual Excel work
• Your environment — Anaconda + Jupyter: the Anaconda distribution as one-installer setup (Python + pandas + Jupyter, zero dependency headaches); notebooks as your analysis workspace — code cells, inline results, markdown workpaper notes, and the "Restart & Run All" trust discipline; notebooks for exploration, scripts for automations, with AI refactoring the working notebook into a clean script
• The pandas exercise ladder (done in Jupyter, on finance data): load & inspect → select & filter (.loc/.iloc) → groupby aggregations → pivot_table (the Excel-brain bridge) → merge to a chart-of-accounts map → time-series resampling & MoM deltas → cleaning → export to Excel. Eight rungs, 15–30 min each, climbed twice: once with a tutorial, once from scratch
• The AI-assisted loop: context-rich code prompts, read-run-fix debugging, spot-check verification for anything touching financial figures, basic git
• The finance toolkit: Excel programmatically (openpyxl/pandas), REST API basics, and how a script becomes a scheduled recurring process

• DataCamp — Data Manipulation with pandas (the intro pandas course; subscription, first chapter free) — interactive reps on real-world data, then re-do everything locally in Jupyter so the skill transfers
• Anaconda distribution (free) + 10 minutes to pandas — environment setup and first local session
• Automate the Boring Stuff with Python — basics, files, Excel, and web/API chapters (free online)
• Real Python — pandas fundamentals, Jupyter notebooks, and working with APIs
• Anthropic — prompt engineering docs
• Your AI assistant as primary teacher: build, then ask "explain this line" until explanations stop surprising you

Why: scripts are tactics; the stack is how a finance engineer builds durable infrastructure instead of one-off hacks.

• LLM APIs as building blocks: messages, system prompts, structured output (the hinge skill), tool use / function calling, model routing in your own builds, and evaluation basics (golden sets, confidence thresholds)
• Workflow engines: n8n / Make / Zapier / Power Automate; the anatomy of a finance workflow (trigger → fetch → transform → route with approval gates → deliver); when engine vs. script vs. native ERP automation
• MCP: the open standard connecting AI to tools and data — how assistants get safe, governed access to ERPs, warehouses, and docs; conceptual fluency plus one hands-on afternoon
• Data foundations: just-enough SQL (SELECT → JOIN) and the source-of-truth discipline that keeps automated pipelines from producing wrong numbers faster

• Anthropic API docs — messages, tool use, structured output (read + experiment)
• Model Context Protocol — intro docs + one connect-a-server walkthrough
• n8n docs + 2–3 finance-adjacent template workflows, reverse-engineered
• KNIME free self-paced courses — the no-code analytics counterpart; double-duty setup for the MIT-equivalent track
• Any SQL crash course covering SELECT through JOIN (e.g. SQLBolt) — a few hours max

Why: the finance engineer's end state is also where the risk lives — this module separates demos from automation a controller will sign off on.

• What an agent is: an LLM in a loop with tools vs. a fixed workflow; agents for genuine case-by-case judgment, workflows for repetition — and why most finance automation should be workflows with AI steps
• The automation map: close & reconciliations, reporting, AP/AR, FP&A itself — prioritized on four axes: frequent, rule-describable, data-accessible, low-blast-radius
• Governance: every automation as a control — owner, logic spec, validation, exception routing, audit trail; human-in-the-loop gates with real spot-check protocols; failure-mode detection (staleness, hallucination, schema drift, permissions)
• The ROI discipline: hours × frequency × error-cost avoided, minus build and maintenance — track realized savings and be walking proof of the value proposition

• Anthropic — Building Effective Agents (the workflows-vs-agents distinction especially)
• One or two close/reconciliation automation case studies — read for architecture and controls, skip the marketing
• A Big 4 ITGC/SOX primer — change management, access controls, audit trails
• Practitioner writing on human-in-the-loop LLM system design

• Analytical (judgment proof): cloud cost model · allocation & unit-cost model · AI unit-economics model ★ · margin bridge · consumption revenue model · build-vs-buy memo · operating model + board pack ★
• Finance engineer (building proof): three scripts · flash-reporting pipeline + runbook · reconciliation agent + controls memo ★ · platform blueprint (elective). A 2-minute screen recording of each automation running beats any description.

8–10 two-sentence bridges: real experience → generalized skill → their problem — including at least two that connect finance judgment to automation builds ("I know what variance commentary should say because I've written hundreds; that's why my automated version routes low-confidence drafts to review").