IA

The MIT Iceberg Index, built with Oak Ridge National Laboratory and reported by CNBC in November 2025, attaches a figure to the substitution: AI can already replace 11.7% of the US workforce, $1.2 trillion in wages, with only 2.2% of that visible in announced tech and IT layoffs. The remaining 9.5% sits in routine HR, logistics, finance and administration roles, across all 50 states rather than coastal tech hubs.

That macro number has a small-scale counterpart. Anthropic's December 2025 study of its own 132 engineers found Claude covering 60% of their work and a self-reported productivity gain near 50%, together with two effects the interviewees named themselves: atrophy of the deep skills needed to write and critique code, and a senior engineer who noticed juniors no longer come to him. Marco Mornati's market figures point the same way. 93% of developers use AI assistants and 27% of production code is generated, yet delivery is only 8-10% faster.

David Thomas compressed it into one anecdote. Adding features to a graphics side project with Claude, he produced unmaintainable code in 18 hours spread over five evenings, a state teams used to take 18 months to reach.

The open question is the training pipeline. Philippe Ensarguet and Mornati both ask where the next architects come from once the entry-level coding that formed them is delegated. Bill Gates, writing in August 2026, says no plan for the transition exists. Dario Amodei expects large effects within one to five years while conceding the timing is very hard to forecast.