Are we in an AI bubble? Signals I actually watch

Signals I actually track when people say “bubble” Capital intensity is unprecedented; payback depends on real usage, not announcements. I watch whether users return unprompted, whether p95 latency holds under load, and whether unit costs fall with scale instead of creeping up with every new feature. The discipline that cuts through noise Run a ten‑task benchmark that mirrors your customers’ work. Publish the rubric. Measure activation, completion, abandonment, and support contacts per 1,000 tasks. If those trend the right way, valuation debates are background static. Operating stance Keep burn low, ship tight loops, and treat GPUs like perishable inventory. Winners will tell a simple story: users stay, latency stays low, and unit margins survive real usage. Reading the macro noise Hardware lifecycles are short, so payback windows are tight. The practical answer is discipline at the product layer: stable retention, low abandonment, and margins that don't collapse when usage spikes.