Is There an AI Bubble? (www.deeplearning.ai)

Andrew Ng argues there isn’t a single “AI bubble” but differing risk profiles across the stack: the application layer is underinvested and ripe for growth, inference infrastructure is currently supply‑constrained and needs more capacity, while model‑training infrastructure is the riskiest—vulnerable to overinvestment and to open‑weight competition plus rapid algorithmic/hardware cost declines. He highlights rising demand from agentic workflows (notably coding assistants), which is driving urgent need for inference throughput and capacity; if investors overbuild training or inference capacity, returns could suffer and negative sentiment might spill over even though fundamentals remain strong. Concretely, Google’s launch of Gemini 3 Pro and Nano Banana Pro illustrates these dynamics: Gemini 3 Pro is a multimodal mixture‑of‑experts transformer with huge context (up to 1M tokens in, 64k out), tool use (search, code exec, function calling), and strong leaderboard wins, but it’s expensive to run (high per‑benchmark token costs) and showed a high “hallucination” rate in one test. Nano Banana Pro leverages Gemini’s reasoning for image generation/editing, multi‑step refinement, consistent character retention, and grounding via search. Broad deployment across Google products gives Google distributional advantage and may shift developer defaults — underscoring why infrastructure and application investments will be decisive for who captures value next.