Jeff Ding, a political scientist at George Washington University, lays out two competing accounts of how technological revolutions redistribute economic power in his book Technology and the Rise of Great Powers. The familiar “leading sector” account says dominance in the breakthrough industry—steel, autos, semiconductors—translates into era-defining advantage. Ding’s alternative, diffusion theory, argues that general-purpose technologies confer long-term leadership to whoever embeds them most broadly and quickly across ordinary economic activity.
Why this matters for AI
Historical parallels—such as Paul David’s 1990 paper The Dynamo and the Computer about electrification—show that productivity gains often arrive only after organizations redesign work around new technology. Initially, factories simply replaced steam engines with a single electric motor and kept the old shaft-and-belt layout; decades later, entrepreneurs and engineers reorganized plants around many small motors and new workflow patterns, producing the real gains. Applied to AI, the implication is that it’s not only where the biggest models are built that matters, but how the technology is woven into everyday jobs across finance, legal, sales, operations, and product development.
Skill infrastructure at the national and company level
Ding uses the phrase “GPT skill infrastructure” to describe education, training, and institutional linkages that widen the pool of people who can work with a general-purpose technology. At national scale, this includes engineering education and university–industry ties. At firm scale, it means the internal systems that spread skills and convert local discoveries into reusable corporate assets. Key elements include:
- internal labs that productize local experiments;
- mechanisms to make experiments visible and shareable;
- a standardized enterprise toolchain and governance; and
- incentives and reward systems that encourage sharing and reuse.
Ethan Mollick (Wharton School) summarizes enterprise transformation as needing leadership that models and incentivizes adoption, a lab to make discoveries reusable, and the “crowd” of everyday workers where applied R&D actually happens. Dan Guido, CEO of AI security firm Trail of Bits, provides operational detail on how firms can make AI “native” rather than merely “assisted.”
Concrete practices for corporate AI adoption
The article collects practical recommendations that firms can apply:
- Standardize the toolchain: limit sprawl by choosing enterprise-wide standards and governance, while allowing disciplined experimentation and evaluation of new tools.
- Write down the rules: publish clear policies on approved tools and risk models (e.g., handling of sensitive data, restrictions on model use) so people understand the why behind restrictions and can safely increase adoption.
- Build a capability ladder (AI maturity matrix): map employee progress from no engagement to creating new tools that advance firm capabilities, and make progress measurable and rewarded.
- Run adoption sprints and hackathons: organization-wide events help employees apply new capabilities to real problems and surface ideas for the lab to productize.
- Package organizational learning into reusable artifacts: skills profiles, repositories, configs, and sandboxes let knowledge compound rather than stay trapped in isolated workflows.
- Make autonomy safe: provide sandboxed environments, guardrails, and hardened defaults that let employees experiment without undue risk.
- Fix data access fragmentation: DJ Patil emphasizes the need for a “tidy house”—unified data environments, clean data flows, and solid data engineering so AI can be used immediately when available.
Incentives and mechanism design
A key difference between national and corporate diffusion is incentives. At the national level, skill infrastructure behaves like a public good: more engineers benefit the whole economy. Inside firms, however, employees may hide AI usage if sharing their automations threatens job security or status. Thus diffusion requires deliberate mechanism design: rules and reward systems that make sharing raise, not lower, an employee’s status. Without that, a small group may exploit frontier models while adoption stalls across the organization.
Geopolitics, sovereignty, and standards
Ding’s framework clarifies geopolitics: a foundational general-purpose technology will be adapted widely and will not remain the exclusive instrument of a single company or nation. “Sovereign AI” thus becomes a predictable outcome of diffusion—societies and firms will shape AI to their legal, cultural, and institutional needs. The arms-race framing that treats AI as a scarce strategic asset is therefore misleading if the correct historical analogies are electrification or computing.
The history of electrification also teaches a caution: application can decentralize while generation centralizes (factories stopped generating power and bought it from grids). For AI, that suggests two pathways: diffusing a centralized large model by API calls, or spreading many smaller models embedded throughout the economy. The healthier future likely mixes centralized and decentralized systems, but it requires interoperability standards—protocols that permit different models and institutions to cooperate without imposing a single universal design.
Open source and the architecture of participation
Open source matters beyond licenses: it depends on the architecture of participation—protocols, interfaces, servers, and conventions that let many actors build on common foundations. The Open Source AI ecosystem is already rich, and it can coexist with widely used proprietary APIs such as those from OpenAI or Anthropic. The future will likely be a hybrid technical and economic order in which cooperation and competition coexist.
Conclusion
Ding’s diffusion view, supported by historical studies such as James Bessen’s Learning by Doing and Arthur Herman’s Freedom’s Forge, implies that the next decade’s winners will be patient, adaptive organizations that build internal skill infrastructures and incentives to share learning. Companies that turn ordinary engineers, analysts, marketers and support staff into people who routinely apply AI in their jobs—and which make it attractive to share those improvements—will gain sustained advantage. The long-term value of AI will accrue not merely to inventors but to those who make it usable, adaptable, interoperable, and widely adopted.
Note: the article references Jeff Ding (Technology and the Rise of Great Powers), Paul David (The Dynamo and the Computer, 1990), James Bessen (Learning by Doing), Arthur Herman (Freedom’s Forge), Ethan Mollick, Dan Guido (Trail of Bits), and DJ Patil, and discusses O’Reilly’s corporate AI transformation practice and related ideas.
