What If the AI Boom Stalls? The Assumptions Behind the 2030 Forecasts

The dominant narrative embedded in this week's AI stories — the Anthropic employee survey, Blackstone's chip-financing terms, Disney's per-head acquisition cost — shares a common operating assumption: that AI capability will continue compounding at something close to the current rate through the late 2020s and early 2030s, translating into massive labor substitution on a short timeline. That assumption deserves scrutiny.

Three underlying premises support the optimistic forecast, and each carries meaningful vulnerabilities. The first is that training compute continues to scale economically — that adding more chips yields proportionally more capability. Several AI researchers have argued that diminishing returns are already visible, with the improvement from GPT-4 to GPT-5 costing more per capability gain than the GPT-3 to GPT-4 jump. The second premise is that training data quality does not hit a ceiling. Models are increasingly trained on synthetic data generated by previous AI models, creating a documented feedback loop where model-generated data can amplify errors rather than introducing genuine new information. High-quality human-generated text on the internet — the primary corpus for the first generation of large language models — has largely been exhausted. The third premise is that inference costs fall fast enough for AI to be economically viable as a substitute for professional labor at scale. A lawyer or doctor currently billing at $200 an hour must be replaceable by an AI that costs substantially less while delivering comparable output quality; current inference costs for routine professional work have not yet reached that threshold across most applications.

The strongest counterargument to the slowdown hypothesis is historical: AI has confounded slow-growth predictions before. The transformer architecture introduced in 2017 reinvented what was possible. Reinforcement learning from human feedback changed alignment. Mixture-of-experts architectures improved efficiency. The field's history is punctuated by unexpected methodological leaps rather than smooth linear progress, and betting against another such leap before 2030 requires confidence that current approaches have exhausted the possibility space.

Three indicators will be telling over the next twelve months. First, benchmark improvement rates on genuinely novel, unpublished multi-step reasoning problems — not headline scores, which are now widely suspected to reflect overfitting to known test sets. Second, the energy cost per useful completed task: if inference costs are falling at the rates Nvidia and AMD are projecting, the economic substitution math begins to close. Third, enterprise AI deployment in legal and medical settings — domains with clear productivity measurement frameworks. If AI is not moving the needle in those domains by 2028, the 2030 forecast looks optimistic. If the slowdown hypothesis is wrong and the Anthropic employees' 75 percent figure proves understated, the labor market implications are more severe and arrive sooner than even the most bullish mainstream forecasts suggest. The policy window to prepare is, in that scenario, narrower than most governments are currently acting as if it is.