Episode
Inference Got Cheap. Renegotiate Everything.
- Published
- May 5, 2026
- Duration seconds
- 516
- Processing state
not_requested- Canonical source
- https://rss.com/podcasts/ypo-technology-network-ai-brief/2794418
Actions
POST https://stenobird.com/v1/public/podcasts/ypo-technology-network-ai-brief-7728971/episodes/inference-got-cheap-renegotiate-everything/transcription-requests
Idempotently request low-priority transcript generation for this episode.GET https://stenobird.com/podcast/ypo-technology-network-ai-brief-7728971/inference-got-cheap-renegotiate-everything.md
Read the agent-friendly Markdown representation of this episode resource.
Summary
For eighteen months the story has been the same. AI is expensive, and getting more expensive. That story has inverted. The price of using AI, not building it, is collapsing, and most of your vendors are quietly hoping you do not notice. In this weekday brief, Stephen Forte teaches the single most important distinction in AI economics, walks through four pieces of evidence in eleven days that the price floor is cracking, and gives you three concrete moves for the contracts already sitting in your legal folder. What you'll learn: Training vs. inference. Training is medical school. Inference is every patient visit for the next forty years. Inference is north of ninety percent of what you actually pay. The chip split. Google announced TPU 8t for training and TPU 8i for inference on April 22. Nvidia, AMD, and AWS Trainium/Inferentia are all moving the same direction. F1 cars vs. delivery vans. The Nebius/Eigen deal. On May 1, Nebius paid $643M for a startup that does one thing: makes AI run inference faster and cheaper. Three months earlier they bought Tavily for $275M. Same theme. DeepSeek V4 (April 24). An open-weight Chinese model claims to close the gap with frontier reasoning at a fraction of the cost. Western vendors will discount or explain why they aren't. Anthropic at $900B. A $50B round only pencils if inference economics work at industrial scale. That is the bet. Models are splitting too. Frontier models are neurosurgeons. Distilled models (Haikus, Minis, Nanos) and mixture-of-experts architectures are nurse practitioners — 95% of the visits at 10% of the cost. Three moves for this week: Pull every AI vendor contract signed in the last eighteen months. Find the inference pricing line (per token, per request, per seat). Ask your CIO: what percentage of our AI work…