Is nuclear power actually coming back?

The honest answer is: partially, selectively, and less dramatically than headlines suggest — but the signals are real and worsening. Entry-level hiring has deteriorated sharply since 2022–2023, with young workers (ages 22–25) in AI-exposed occupations experiencing employment roughly 19% below expected trend as of mid-2026. However, no major institutional analysis has found evidence of economy-wide AI job displacement. The decline works primarily through reduced hiring rather than layoffs, is concentrated in specific sectors (tech, customer service, finance, creative), and is entangled with confounding factors — post-pandemic labor market correction, rising interest rates, remote work, and employer cost discipline — that make it impossible to attribute the full decline to AI alone.

AI agents are not replacing SaaS — but they are restructuring it, compressing some categories, and expanding others. The "SaaSpocalypse" narrative that dominated markets in early 2026 proved premature: the software sector lost over $1 trillion in market capitalization between January and April 2026, then partially recovered by Q3. The real shift is architectural and commercial, not existential. AI agents are genuinely disrupting customer support, sales development, and certain point-solution categories. They are not meaningfully disrupting ERP, core financial systems, cybersecurity, or complex multi-jurisdiction compliance software. Every major SaaS incumbent — Salesforce, ServiceNow, Microsoft, SAP, Workday, Oracle — has responded by embedding agents directly into their platforms rather than ceding ground. The most consequential change is the pricing model: per-seat subscription is being replaced by hybrid and outcome-based models, which fundamentally alters SaaS revenue economics without necessarily destroying them. Agent-native startups (Sierra, Decagon, Harvey) are growing fast but primarily expanding the market into the services budget rather than cannibalizing existing SaaS spend. The decision-maker's frame: not "replace or survive" but "which categories, on what timeline, at what pricing, with what governance."

Training a frontier AI model in 2026 is one of the most capital-intensive activities in the private sector. The total cost of the most expensive single training runs is now estimated in the hundreds of millions of dollars, with credible projections of billion-dollar runs by 2027. But the headline "training cost" figure that appears in press releases—DeepSeek's famous $5.6M, for instance—typically covers only the final pre-training compute run, excluding hardware acquisition, preliminary experiments, failed runs, data curation, human annotation, and R&D staff. When all costs are included, total development expenditure is typically 3–10× the final-run compute figure.