Compounding Waves — How Each Tech Era Built the Substrate, and the Skills, for the Next

Narrative

The foundational dependency chain across all three technology waves is most visible in the scaling-law literature. Kaplan et al. (2020, arXiv) established that language model performance scales as a power law with compute, data, and parameters across seven orders of magnitude, while Hoffmann et al.'s Chinchilla laws (2022) demonstrated that data must grow proportionally with model size for compute-optimal training. These results made the open web — assembled by wave-one internet infrastructure — the load-bearing input for wave-three AI, not merely a convenience. METR's empirical benchmarking programme then provides the clearest view of how rapidly that AI wave is closing the gap with human software labour: the HCAST and RE-Bench task suites show AI task-completion time horizons have doubled every seven months since 2019, and as of early 2025 Claude 3.7 Sonnet achieves 50% success on tasks that take human experts roughly 50 minutes.

The data-access transition from open externality to contested resource is now the best-documented constraint in the field. Longpre et al.'s NeurIPS 2024 paper 'Consent in Crisis' audited 14,000 web domains and found robots.txt and Terms of Service restrictions on AI crawling rose more than 500% between mid-2023 and April 2024, with 28%+ of the most critical C4 sources now fully restricted. A subsequent arXiv paper on crawler restrictions (2025) extends this finding to political-bias consequences, showing moderate news sites withdrawing first and leaving hyperpartisan material over-represented in training sets. Together, these papers frame wave three as the last generation able to consume the open web at minimal cost: the Wikimedia Foundation reported in 2025 that 65% of its most expensive traffic now comes from AI crawlers, raising the question of who funds the next data substrate.

The labour-economics literature is in productive disagreement about net job effects. Acemoglu and Restrepo's task-displacement framework, synthesised in Oxford Open Economics (2024), shows automation systematically reduces labour share unless offset by new task creation. Applied to LLMs, the recent 'Crashing Waves vs Rising Tides' preprint (arXiv, 2026) finds that AI capability gains improve performance broadly across task durations rather than arriving in abrupt occupation-level bursts, while the 'Augmenting or Automating Labor' paper (arXiv, 2025) shows displacement effects outpace productivity gains for low-skilled workers from 2015 to 2022. The Harvard Business School displacement-complementarity study finds a 24% reduction in automatable skills per firm per quarter post-ChatGPT against a 15% increase in augmentation-exposed roles — but this positive complementarity figure is concentrated in mid-to-high-skilled occupations.

On the infrastructure concentration question, van der Vlist et al. in Big Data & Society (2024) and the Internet Policy Review cloud-platform study both document how hyperscaler cloud credits, API lock-in, and strategic investment create a circular capital structure. The 'Hype, Sustainability' arXiv paper (2024) quantifies this: three major cloud providers contributed two-thirds of the $27 billion raised by AI startups in 2023, and up to 80–90% of early-stage AI startup capital flows back to cloud providers. The 'externalised harvest' framing thus has strong empirical support — value created by wave-one open-web content and wave-two cloud infrastructure accrues disproportionately to current-wave AI labs and their hyperscaler backers, while the original creators receive neither compensation nor attribution.