Where AI Value Lands

Innovation vs. Diffusion — The GPT Frame

6 min read

Innovation vs. Diffusion — The GPT Frame

Idea

Per Jeffrey Ding (ding-diffusion-marathon), historical economic power transitions driven by General Purpose Technologies (GPTs) are rarely decided by the initial invention. They are decided by diffusion capacity — the ability of a nation to spread and adopt the technology across all productive sectors. The current US framing of the AI race as an AGI sprint is therefore strategically misframed.

Historical analogs

  • Second Industrial Revolution. The US overtook Britain not by inventing electricity or the internal combustion engine but by diffusing them rapidly across a vast industrial base.
  • Postwar USSR. Massive innovation capacity (first satellite, major physics breakthroughs) but a fatal diffusion deficit — scientific triumphs never reached broad economic productivity.

The implication: the GPT winner is whoever absorbs the technology into the macroeconomy fastest, not whoever invents the most advanced version first.

Application to the current US/China divide

United States. Strong absolute innovation, strong relative diffusion capability (deep engineering base, capable absorptive institutions). But policy is over-indexed on the “leading-sector mechanism” — frontier breakthroughs — and on restricting foreign access. This neglects the domestic diffusion mandate. Outcome risk: a top-heavy economy where AI surplus is trapped inside the tech sector itself.

China. Strong innovation capacity (patents, publications) but historically a diffusion deficit caused by uneven STEM education and institutional bottlenecks that isolated breakthroughs in elite institutions. The strategic pivot — described in open-weight-asymmetry and lee-01ai-pivot — is a deliberate attempt to close this gap by forcing AI downstream into traditional industries (banking, insurance, mining, energy).

OECD measurement layer. oecd-sme-ai-adoption-2025 supplies the first systematic G7-wide measurement of diffusion at the SME tail — the literal bottom of the diffusion S-curve. Findings: OECD-wide AI adoption is at 14% in 2024 (firms 10+ employees), with 40% large vs 11.9% small firms adopting and only 29% of generative-AI-using SMEs deploying it in core business activities. Construction, hospitality, transport, retail — the bulk of SME-heavy sectors — sit at 7–12% adoption vs ~45% in ICT (see sectoral-ai-diffusion-pattern). G7 core-business AI adoption ranges from 1.9% (Japan) to 6.1% (US). These numbers make Ding’s diffusion-marathon thesis empirically operational: the diffusion gap is not a theoretical risk but a measurable artifact, and it is widening in absolute terms (sectoral leaders pulled away from laggards by 15 pp in 3 years).

Why this is the key economic concept

The whole “where does value land” question depends on whether AI behaves as a GPT or not. If it does:

  • The current AGI-sprint framing is the wrong mental model.
  • The value will land wherever a bloc successfully diffuses the technology across sectors, not where it pioneers the next model.
  • This pushes the long-run answer toward the application layer, the B2B enterprise integration layer, and the SME absorptive layer — i.e., away from the bottom of the US-shaped U.

Combined with task-based-framework (which caps US TFP gains from AI at <0.71% cumulative if the diffusion does not happen), this is one of the most important economic concepts in the wiki.

Supporting evidence

Worker-level: capability is 3× ahead of deployment

massenkoff-mccrory-labor-market-impacts-2026 supplies the first systematic occupation-level measurement of the diffusion lag. In the most-exposed category (Computer & Math), Eloundou et al.’s theoretical capability covers ~94% of tasks while Claude’s observed work-context usage covers only ~33% — see theoretical-vs-observed-capability-gap. The “red area will grow to cover the blue” framing is the diffusion process itself, at occupation granularity. Critically, theoretical capability alone does not predict BLS occupational growth forecasts; observed exposure does (see ai-exposure-vs-bls-growth, slope −6.07, SE 1.32). This is the strongest single empirical argument for diffusion-aware measures over capability-only ones in labor economics: capability without diffusion is invisible in the labor market.

Firm-level: OECD G7 SME measurement layer

The OECD G7 discussion paper on SME AI adoption (2025) operationalises this thesis at the firm level. Its entire policy half — the oecd-sme-enabler-quartet (connectivity, AI-enabling inputs, skills, finance), the sme-policy-pathway-novice-to-champion matrix, and the g7-sme-ai-policy-pluralism of instruments — is a diffusion playbook applied to SMEs. The paper notes that nearly 70 countries have adopted national AI strategies; the convergence of G7 instruments on compute-access, sectoral subsidies, skills programmes and innovation hubs (despite divergent rhetorics) is itself evidence that diffusion has become a recognised policy frontier alongside frontier-model invention.

Tensions

  • “Diffusion capacity” is hard to measure directly. The historical analogs are post-hoc; whether anyone can identify a diffusion bottleneck in real time is unclear.
  • The US may diffuse AI faster than this critique assumes. Coding agents, copilots, and consumer chat have penetrated end-user workflows at speed.
  • China’s “diffusion pivot” is recent. Whether 01.AI’s B2B traction in banking/insurance/mining generalizes — or whether it remains scattered industry pilots — is an open question.
  • The US/China asymmetry inside the OECD data. OECD ICT survey shows the US at unusually low aggregate adoption (3–6%) but with a strong young-firm / start-up adoption signal. This is consistent with Ding’s claim that the US over-indexes on the leading-sector mechanism (frontier startups) and under-invests in broad SME diffusion. China is not in the OECD survey frame, but the lee-01ai-pivot B2B-diffusion strategy is a deliberate response to a measurement Ding implied — China is operationalizing a diffusion-first agenda from the opposite direction.
  • Binding constraints made concrete. The OECD evidence sharpens one mechanism: if AI diffusion is governed by SME-level absorptive capacity, then the binding constraint is human capital (see ai-skill-shortage-as-diffusion-bottleneck) and to a lesser extent connectivity (see sme-connectivity-divide) and finance (see sme-ai-finance-gap). The Ding-style thesis becomes empirically testable at this level: does a country’s diffusion capacity correlate more with its training and broadband stack than with its frontier-model investment? Italy (4.7× large-vs-small AI-adoption ratio) and the US (1.65×) bracket the spread within the G7.

Referenced by

Auto-generated reverse-link index — pages in the wiki that link here. Maintained by /lint and reverse-link sweeps.

Paper planning

Search

Concepts

Sources

Data

Graph View

Backlinks