Inference has become a recurring operating expense for engineering teams, with estimates cited by Brookings placing 80–90% of AI computing power on inference. The International Energy Agency reports data center electricity demand grew 17% during 2025, with AI-focused facilities surging 50%, and its base case sees consumption roughly doubling to 945 TWh by 2030. The four largest US hyperscalers have guided toward roughly $725 billion in 2026 capital expenditure, up about 77% year over year. Teams are treating inference as a procurement problem with levers covering model breadth, purpose-built silicon, decentralized compute, serving-engine efficiency, and unit price. Open-weight model APIs generally run 50–90% below frontier pricing for comparable workloads. Together AI serves a catalog of more than 200 models, with published rates spanning roughly $0.03 to $9 per million tokens, and adapter-based inference running at standard rates plus a modest overhead. Groq runs open-weight models on custom Language Processing Unit (LPU) silicon, with independent measurements placing throughput between roughly 300 and 1,000 tokens per second, against 50–150 for comparable GPU serving, and Llama 3.3 70B published at $0.59 input and $0.79 output per million tokens. Decentralized networks publish the lowest headline rates, with model coverage and price stability needing validation. OpenAI-compatible endpoints have cut switching costs to a base-URL change