## How the Battle for Control Could Crush AI’s Promise

## Source details

**Canonical URL:** [How the Battle for Control Could Crush AI’s Promise](https://www.imf.org/-/media/files/publications/fandd/article/2025/09/frey.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/fandd/article/2025/09/frey.pdf.md)
- [Structured JSON version](/-/media/files/publications/fandd/article/2025/09/frey.pdf.json)

---

### Historical lessons on institutions and technological change
- Mid-20th century expectation: centrally planned economies (USSR) might outperform market-driven ones after early technological successes (Sputnik, Yuri Gagarin).
- USSR failure attributed not to scientific talent shortages but to institutions “inhospitable to exploration” and central orchestration that prevented decentralized experimentation.
- Japan’s postwar model: tightly coordinated production and dense knowledge networks produced efficiency gains but emphasized incremental process improvements over frontier product innovation.
  - Japanese autoworkers were 17 percent more productive than their US counterparts by 1980.
  - Roughly two-thirds of Japanese R&D targeted process improvements (Edwin Mansfield).
- Europe’s coordinated capitalism supported catch-up growth but struggled when the computer revolution created technological uncertainty; some Southern European nations experienced “two lost decades.”
- Core insight: institutions that serve catch-up growth and stable mass production become liabilities when the challenge is pioneering frontier innovation.

### Frontier innovation and why centralization is limited
- Frontier breakthroughs arise from exploring the unknown, not from perfecting formalized processes.
- Large language models (LLMs) scaled massively but show limited advanced reasoning performance:
  - LLMs grew 10,000-fold in scale between 2019 and 2024 yet still scored only about 5 percent on the ARC reasoning benchmark.
  - Program search approaches topped 20 percent on ARC.
- Embodied, sensorimotor human experience remains absent from large, internet-trained language models; human-style exploration and embodied knowledge are essential for many breakthroughs.
- Historical analogy: models trained on past consensus would reproduce outdated beliefs (example: a model trained in 1633 would uphold geocentrism), underscoring the limits of statistical-consensus defaults.

### Control, concentration, and current risks to AI-driven innovation
- Centralized authority and corporate concentration are rising in both China and the US, threatening the competitive dynamism that fosters frontier innovation.
- China:
  - Recent recentralization: licenses, credit, and contracts favor politically reliable conglomerates; antitrust law is applied selectively.
  - Provincial experimentation has withered; crude indicators (e.g., patent counts) are being chased, producing low-value filings.
  - Firms without political connections (example: DeepSeek) are often the most innovative but remain vulnerable without robust legal protections; firms divert resources toward political alliances.
- United States:
  - Increased industrial concentration and extensive non-compete clauses hamper labor mobility, reducing tacit knowledge flow and discouraging new firm formation.
  - Incumbent lobbying has hard-coded regulatory advantages (patent extensions, sector-specific licensing hurdles), contributing to concentration.
- Market concentration in AI:
  - Microsoft’s deep alliance with OpenAI controls about 70 percent of the commercial LLM market.
  - Nvidia provides about 92 percent of the specialized GPUs used to train these models.
  - Alphabet, Amazon, and Meta have been acquiring stakes in promising AI start-ups.
- Consequence: Without policies that protect the competitive arena itself (not the fortunes of particular firms), the next generation of transformative innovators may be stifled.

### Key findings and implications
- Centralized bureaucracies excel at exploiting accessible, established technologies for catch-up growth but falter at frontier innovation where decentralized experimentation is crucial.
- Technological paradigms can shift the institutional mix that optimally supports growth; failure to adapt governance and competition policy risks prolonged stagnation.
- Empirical indicators and examples:
  - 17 percent productivity advantage of Japanese autoworkers by 1980.
  - LLM scale growth: 10,000-fold (2019–2024).
  - ARC benchmark performance: LLMs ~5 percent; program search >20 percent.
  - Market shares: Microsoft/OpenAI ~70 percent of commercial LLM market; Nvidia ~92 percent of specialized GPUs.

*Carl Benedikt Frey; F&D, SEPTEMBER 2025*

---


_Source: https://www.imf.org/-/media/files/publications/fandd/article/2025/09/frey.pdf_