Cloud Repatriation Architect

AI COGS (Cost of Goods Sold) refers to the variable costs directly attributable to delivering AI-powered features to customers. Unlike traditional SaaS (near-zero marginal cost per user), AI features have significant per-interaction costs. **Components of AI COGS:** - LLM API fees (OpenAI, Anthropic, Google per-token charges) - Embedding generation and vector database queries - GPU compute for inference or fine-tuning - Data retrieval and processing pipeline costs - Monitoring, logging, and observability infrastructure - Error handling, retry logic, and fallback model costs - Human-in-the-loop review costs **Impact on SaaS economics:** Traditional SaaS enjoys 80%+ gross margins. AI-heavy SaaS products can see margins compress to 40-60%, fundamentally changing valuation multiples and capital requirements.

AI inference is the process of running a trained model to generate predictions or outputs from new input data. Unlike training (which is done once), inference happens every time a user interacts with an AI feature — every chatbot response, every code suggestion, every image generation. Inference cost is the dominant variable cost in AI features. Training GPT-4 cost an estimated $100M, but inference costs across all users dwarf that number. Each inference call consumes GPU compute proportional to model size and input/output length. Inference optimization is a critical engineering discipline: model quantization (reducing precision from 32-bit to 8-bit or 4-bit), batching (processing multiple requests simultaneously), caching (storing common responses), and distillation (creating smaller student models from larger teacher models). For product leaders, inference cost is the unit cost that determines whether your AI feature has positive or negative unit economics. Richard Ewing's AUEB tool calculates Cost of Predictivity — the true per-query cost including inference, retrieval, verification, and error handling.

Technical debt is the implied cost of future rework caused by choosing an expedient solution now instead of a better approach that would take longer. First coined by Ward Cunningham in 1992, technical debt has become one of the most important concepts in software engineering economics. Like financial debt, technical debt accrues interest. Every shortcut, every "we'll fix it later," every copy-pasted function adds to the principal. The interest comes in the form of slower development velocity, more bugs, longer onboarding times for new engineers, and increased fragility of the system. Technical debt exists on a spectrum from deliberate ("we know this is a shortcut but ship it anyway") to accidental ("we didn't realize this was a bad pattern until later"). Both types compound over time. Organizations that don't actively measure and manage their technical debt risk reaching what Richard Ewing calls the Technical Insolvency Date — the specific quarter when maintenance costs consume 100% of engineering capacity.

The Cost of Predictivity is a framework coined by Richard Ewing that measures the variable cost of AI accuracy. Unlike traditional software with near-zero marginal costs, AI features have costs that scale with usage and accuracy requirements. The key insight: as AI correctness increases, cost scales exponentially. Moving from 80% accuracy to 95% accuracy often requires a 10x increase in compute and retrieval costs. Moving from 95% to 99% may require another 10x. This creates margin compression that traditional engineering metrics don't capture. A feature that works beautifully at 100 users may be economically unviable at 100,000 users because AI inference costs scale linearly with usage while accuracy improvements require exponentially more resources. The AI Unit Economics Benchmark (AUEB) calculator at richardewing.io/tools/aueb helps companies calculate their Cost of Predictivity and identify their AI margin collapse point.