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

Narrative

The dominant analytical frame across independent blogs is Ben Thompson's Stratechery, which provides the most coherent through-line across all three waves. His 2015 Aggregation Theory piece identified how zero marginal distribution costs rewired industry power during wave one; his 2024 AI and enterprise piece traces how Salesforce's SaaS model bridged wave one's internet infrastructure to wave two's cloud platforms; and his 2026 Agents Over Bubbles argues that wave three breaks the aggregation model itself by reintroducing marginal compute costs at scale. The structural logic is explicit: Amazon could fund Anthropic because AWS generated the capital and the infrastructure, and Anthropic can train frontier models because AWS built the hyperscale compute layer — a three-generation compounding chain made concrete in a single corporate dependency graph.

Simon Willison's Weblog provides the practitioner counterpart. His annual LLM reviews (2023, 2024, 2025) document the 100x inference cost drop between 2022 and 2025 and the diffusion of GPT-4-class capability to 18 separate organisations, empirically tracing the commoditisation dynamic that Thompson theorises. Willison's 2026 Jevons paradox framing — that cheaper cognition generates more cognitive demand rather than less — is the clearest independent statement of the compounding dividend thesis, while his detailed attention to training data provenance and model limitations keeps the analysis honest about what open-web corpora actually contain and what happens when they thin.

On the labour-economics side, Substack writers engaging with Acemoglu, Autor, and the Frey–Osborne tradition form a coherent cluster. A February 2026 Great Leadership post applies the displacement-productivity-new-task-creation trifecta to current AI vacancy data and finds that the balance varies sharply by task type: routine cognitive tasks show clearly net-negative displacement while strategic and creative roles show job growth. Tom Bewick's Substack engages the newer Gans–Goldfarb O-ring automation model, which challenges the task-separability assumption underlying most exposure indices and suggests that automating easy tasks concentrates human effort on harder bottleneck tasks — a finding that cuts against simple displacement narratives without fully validating the compounding dividend story.

The externalised-harvest framing receives its clearest independent articulation in the Augmented Mind Substack, which characterises the AI platform stack as a structural repeat of wave-two platform dependency: cheap initial access to build user and developer lock-in, followed by pricing power once dependency is established. LessWrong contributors add the data exhaustion dimension, with multiple posts converging on an estimated 2026–2032 window when general-purpose internet training data runs dry — directly implicating wave-one open-web content (Common Crawl, Reddit, GitHub, Stack Overflow) as a finite and now-contested substrate, not an infinitely renewable externality.