Tiny Little Lab · AI Infrastructure Intelligence
Microsoft's $190B Infrastructure Bet: Largest single-company AI capex in history. 6% of market cap. Azure pre-selling compute before buildings are powered. The new bottleneck isn't GPUs — it's HBM3E memory bandwidth + chip packaging.
GPUs are available. HBM3E supply is the real choke point — not raw GPU wafers. Azure vertically integrating the entire chain: custom silicon (Maia 100), power purchase agreements, chip fab to customer workload.
Gemini 3.5 Flash default in AI Mode. Spark 24/7 personal agent. Spark Agent SDK commoditizes the wrapper layer. Pichai: " We'd rather have a billion users at breakeven than a million at margin."
Claude Opus 4.8 abstains more, guesses less — 4x fewer confident wrong answers. Anthropic positioning for high-stakes enterprise. The "I don't know" reflex is now a feature, not a bug.