Building on AI's Unfinished Foundation
Generative AI will only become a truly economy-transforming general-purpose technology once it is "platformed" — surrounded by the technological, industrial, and institutional architecture that lets decentralized organizations build on it with confidence — and many companies are investing as though that architecture already exists.
Summary
Northeastern's Kevin Boudreau argues that despite generative AI's fast adoption (2.4B monthly users, Cursor past a $2B revenue run rate, Salesforce Agentforce at $1.2B ARR), it isn't yet a true general-purpose technology on the scale of electricity or the internal combustion engine, because it hasn't been 'platformed' — surrounded by stable technological, industrial, and institutional architecture that lets decentralized organizations build on it with confidence. He traces electrification's roughly 25-year platforming lag (proven 1882, AC confirmed at commercial scale via Niagara Falls 1895-96, the regulated-utility model settling 1898-1907, the first state public-utility commissions in 1907) as the template, then maps AI's current stack: the lower layers (chips, cloud, frontier models — Nvidia, AWS/Azure/GCP, OpenAI/Anthropic/DeepMind/xAI) are converging, but the application/integration layer, where most companies actually build, remains fluid, and no one — not even the frontier labs — has taken on genuine platform-leadership coordination (as Microsoft/Intel did for the PC, or Apple for the iPhone); companies are substituting vertical integration for that missing coordination instead. His four investment principles for operating before AI is fully platformed: learn faster than you commit (use cheap open-weight models to capture expert corrections while the model is still often wrong); build assets that survive architectural change (watch for commercial lock-in even when a model is technically swappable — e.g., Salesforce's proprietary 'agentic work unit' billing, SAP's data-access restrictions); invest in complements, not rentable intelligence (value accrues at 'seams' — interfaces the layers above can't easily reproduce, like Epic's pre-existing clinical trust in healthcare AI); and build organizational capability (understanding your own knowledge, decisions, and bottlenecks) as the complement a shared model can never internalize.
Editor's opinion
The framework draws on established GPT economics (Bresnahan & Trajtenberg, Brynjolfsson-Rock-Syverson's J-curve), but the AI-specific synthesis and the four investment principles are genuinely sharp — especially the distinction between a model being technically swappable and being commercially locked in (Salesforce metering agent work in units it alone defines), and the idea of investing in 'seams' rather than intelligence itself, since everyone can rent the same models. The electrification parallel isn't decorative: platforming took roughly 25 years last time, split into technological, industrial, and institutional stages, which gives a concrete way to diagnose where AI's real bottleneck sits right now — not the models, but the missing platform-leadership coordination that Microsoft/Intel or Apple supplied in earlier cycles.
Key concepts
General-purpose technology · platforming · complementary innovation · productivity J-curve