Sky-high AI spending masks a system in distress: unprecedented capital flows into AI coincide with operating losses and speculative valuations, suggesting AI is functioning as a vehicle of financial expansion rather than the engine of capitalist renewal.

Main Finding

The paper argues that the recent surge in AI investment—despite its scale and technical novelty—functions primarily as a crisis response within financialised capitalism rather than as the foundation of a new, self-sustaining productive regime. AI accumulation exhibits the hallmarks of overaccumulation: rising capital intensity, speculative valuation disconnected from realised profits, fragile revenue models, and heavy dependence on credit and equity finance. These features make AI an advanced expression of hegemonic autumn (financial expansion) and limit its capacity to resolve capitalism’s structural profitability problems or to produce an inclusive development pathway for peripheral economies.

Key Points

• Scale and concentration of investment

  • By 2025 private AI investment in the U.S. reached US$285.9 billion; global private AI investment was US$344.7 billion with generative AI ≈ US$163.6 billion. Global private funding surged ~127.5% in one year (Stanford HAI 2026).
  • The U.S. hosts an extreme concentration of computational infrastructure (≈5,500 operational data centres as of 2025).

• Capital-intensity and the organic composition of capital

  • AI development is anchored in large fixed-capital outlays (hyperscale data centres, specialised semiconductors, energy infrastructure). Labour costs are comparatively small relative to compute/infrastructure.
  • Training costs scale steeply (GPT-3: millions; GPT-4: estimated US$50–100M; frontier projections up to US$500M–US$1B per run), implying a rapidly rising organic composition of capital.

• Valorisation and realisation problems

  • Trained models act like fixed capital with near-zero marginal reproduction cost; value realisation depends on deployment and revenue capture rather than repeated production.
  • Fragile revenue models, operational losses at frontier firms, and speculative valuations (e.g., OpenAI and Anthropic valuations cited in the paper) indicate that market prices often price future expectation, not current surplus extraction.

• Financialisation and hegemonic dynamics

  • Patterns mirror Marxian overaccumulation and Arrighi-style financial expansion in hegemonic autumn: capital flows from productive to speculative/financial channels when profitable productive outlets decline.
  • AI investment mixes material expansion (infrastructure) with strong financial-expansion features (inflated valuations, dependence on finance, uncertain profitability).

• Institutional and geopolitical constraints (SSA & hegemonic instability)

  • The institutional coherence required for stable accumulation (labour relations, regulatory frameworks, global financial order) is fragile; neoliberal configurations and securitisation of technology produce tensions that hinder steady valorisation.
  • Geopolitical securitisation (export controls, IP protection) contradicts prior hyper-globalisation, constraining peripheral capability-building.

• Unequal exchange and core–periphery dynamics

  • AI value chains concentrate high-value activities (IP, platform control, revenue capture) in core economies while hardware manufacture, data annotation, and resource extraction concentrate in peripheral/semi-peripheral countries.
  • These dynamics reproduce unequal exchange, limiting developmental prospects for peripheries even as AI is framed as transformative.

Data & Methods

• Approach: theoretical synthesis combining (1) Marxian crisis theory (falling rate of profit/overaccumulation), (2) Social Structures of Accumulation (SSA) theory, and (3) Arrighi’s systemic cycles of accumulation / hegemonic transition.
• Empirical documentation (primarily 2022–2025):
  • Sources: corporate disclosures, industry reports, independent research (notably Stanford HAI / AI Index), Reuters reporting, International Energy Agency (IEA) projections, and secondary literature.
  • Metrics and evidence used: private AI investment flows, generative-AI share of investment, data-centre counts and energy forecasts, reported and estimated model training costs, firm valuations and reported revenues/ARR, observed operational losses and financing patterns.
• Limitations acknowledged by the paper:
  • Lack of audited financial accounts for leading frontier firms (OpenAI, Anthropic, xAI) forces reliance on industry estimates, press reporting, and extrapolations.
  • Some cost and valuation figures are contested or speculative in secondary sources; inferences about the organic composition of capital and valorisation risks are therefore qualified.

Implications for AI Economics

• Reframe expectations about AI’s macroeconomic role

  • Economists and policymakers should not treat AI investment automatically as a self-sustaining productivity growth engine. Models must incorporate realisation risk, high fixed costs, and the possibility that large-scale AI deployment will not translate into proportional surplus-value extraction or broad-based productivity gains.

• Incorporate capital-intensity and financialisation into analyses

  • Growth and welfare models should explicitly model extreme fixed-cost structures, network effects, non-rival outputs, and the role of speculative finance in setting valuations and directing resources.

• Energy and infrastructure constraints matter

  • AI’s dependence on energy-intensive infrastructure creates material limits (grid capacity, long-term energy contracts, environmental externalities) that condition accumulation and should be integrated into macro/sectoral projections.

• Institutional and geopolitical factors are binding

  • Institutional coherence (regulation, labour policy, industrial policy, international financial order) shapes whether AI can be channelled into productive, inclusive accumulation. Geopolitical securitisation (trade controls, IP regimes) constrains peripheral catch-up and global diffusion.

• Development policy implications for peripheral economies

  • Peripheral states should be cautious about strategies that assume easy technology transfer or replication of core-based AI models; capital intensity and geopolitical barriers reduce the feasibility of latecomer catch-up without distinct industrial/state strategies.
  • Policies should focus on alternative developmental paths: skills and labour upgrading, data governance, local niche applications, energy planning, and negotiating more equitable global arrangements for value capture.

• Regulatory and redistribution considerations

  • Given the speculative and concentrated nature of gains, policymakers might consider taxation of large AI rents, stricter oversight of valuations and financing practices, stronger antitrust enforcement on platform control, and policies to channel AI investment toward socially productive outcomes (public infrastructure, climate mitigation, public-sector AI).

• Research agenda for AI economics

  • Empirical work needed on realised returns to AI investments, audited profitability of frontier firms, total social energy/resource costs, and the distributional flows along AI value chains (to quantify unequal exchange).
  • Theoretical work should integrate classical value-based crisis frameworks with modern theories of non-rival digital goods and financialisation to better model systemic risks.

In sum: the paper urges treating the AI boom as historically situated within capitalist crisis dynamics—highly capital‑intensive, financially driven, geopolitically fraught, and distributionally skewed—rather than as an unambiguous resolution to stagnation or an automatic engine of inclusive growth.