AI Engels' pause, bubble risks & tabular models

Leading global economists, multilateral policymakers, and financial executives gathering at the 2026 Bund Summit in Shanghai issued warnings regarding generative AI's macroeconomic trajectory, cautioning that rapid adoption risks creating severe income disparities and structural labor shocks^[1]. Held from October 19 to 22 under the banner of "Technology, Geopolitics and the Global Economy" - jointly organized by the China Finance 40 Forum (CF40) and Tsinghua University - the summit centered on the friction between rapid productivity gains and equitable wealth distribution^[1][2]. Multiple senior analysts highlighted that while generative AI is entering an unprecedented expansion across scientific discovery, knowledge creation, and enterprise operations, its current macro benefits are heavily skewed toward capital owners rather than labor^[1].

A central focus of the deliberations was an economic parallel drawn by Huang Yiping, Dean of the National School of Development at Peking University^[1][3]. Huang warned that the global economy may be entering a modern "Engels' pause" - a phenomenon first observed during the early decades of the Industrial Revolution when corporate output and gross domestic product surged, yet worker wages stagnated for nearly half a century^[1]. Huang emphasized that if generative artificial intelligence disproportionately enriches the holders of compute infrastructure and proprietary algorithmic models at the expense of wages, it will trigger wider societal inequality and strain sovereign fiscal and tax systems^[1]. Reinforcing this view, Cornell University professor and former World Bank senior vice president Kaushik Basu stated that AI-driven labor demand suppression cannot be resolved by domestic policies alone, urging the United States and China to initiate coordinated global tax policies to prevent an unmanageable concentration of wealth^[4][5].

Empirical labor data presented at the summit underscored these structural shifts. Krishna Srinivasan, Director of the Asia and Pacific Department at the International Monetary Fund (IMF), revealed that relative labor demand in industries with high generative AI exposure has dropped by 10% compared to sectors with low exposure^[1][6]. Although Srinivasan observed that the decline has recently begun to stabilize - indicating workplace realignment rather than outright worker obsolescence - the threshold of required technical skills is rapidly rising^[1][6]. Former U.S. Treasury Secretary Robert Rubin echoed these labor market vulnerabilities, stressing that knowledge workers face displacement pressures as frontier models advance^[1]. Rubin urged Beijing and Washington to institute bilateral, Cold War-style risk containment and de-risking mechanisms to prevent catastrophic labor and technological disruptions^[7].

The geopolitical and macro distribution of AI returns remains similarly uneven. Claudio Irigoyen, Global Head of Economics at Bank of America, noted that the current economic impulse from AI remains fundamentally driven by an upstream capital expenditure and investment boom rather than downstream corporate profit transformation^[1][6]. Consequently, the immediate financial upside is largely restricted to a small cohort of hardware suppliers, chip manufacturers, and countries deeply integrated into the semiconductor and power supply chains, leaving consumer-facing sectors and non-aligned developing markets searching for tangible returns^[1][6].

Central banking authorities and multilateral institutions raised urgent concerns regarding financial stability risks tied to hyperscale AI capital expenditure and widening gaps in regulatory oversight^[1][2]. Speaking on October 20 at the Bund Summit, Zhou Xiaochuan, former Governor of the People's Bank of China (PBOC), warned that the sheer scale of global capital flowing into AI physical infrastructure - encompassing mega data centers, custom silicon, advanced liquid-cooling facilities, and dedicated power grids - risks fueling an investment bubble^[1][3]. Zhou observed that because the long-term dependency of future AI applications on massive centralized data centers remains uncertain, any systemic mismatch between infrastructure outlays and practical application revenues could jeopardize financial stability, urging monetary authorities to stockpile policy intervention tools^[1][3].

Zhou classified AI spending into three distinct tiers: foundational model training compute, downstream enterprise applications, and supporting infrastructure including energy and semiconductors^[1]. He argued that durable productivity gains will ultimately emerge from lightweight, localized application models deployed directly on edge devices such as laptops, automobiles, and smartphones, rather than perpetual hyperscale data center expansion^[1]. Should end-user software monetization fail to cover the high debt loads and capital investments incurred by data center builders, the financial system could face balance-sheet stress across tech hardware and specialized credit lenders^[1][3].

Simultaneously, the World Bank sounded an alarm over the institutional capacity of emerging markets to govern these systemic shifts^[2]. Zhang Wencai, Managing Director and Chief Administrative Officer of the World Bank, presented survey findings showing that only 11% of financial regulators in emerging market economies currently possess or deploy AI-enabled supervisory tools^[2]. Approximately 50% of the surveyed agencies acknowledged an acute absence of technical infrastructure and specialized personnel necessary to supervise automated trading, AI credit assessments, and cross-border financial algorithms^[2]. Zhang called for immediate international cooperation to build digital foundations and urged developing markets to prioritize small-scale, cost-effective supervisory tech rather than costly proprietary platforms^[2].

The monetary policy dilemmas created by this bifurcated economic environment were underscored on October 20 at the 2026 Beijing Financial Street Forum by Edward S. Robinson, Deputy Managing Director and Chief Economist of the Monetary Authority of Singapore (MAS)^[4][4]. Robinson outlined how generative AI introduces new complexities to central banking, as supply chains face reconfiguration and traditional productivity models struggle to capture tech-driven growth^[4]. Echoing classic economic debates on the modern productivity paradox, Robinson emphasized that central banks must decipher whether generative AI's capital boom is delivering true total factor productivity gains or merely generating temporary demand shocks that complicate long-term inflation forecasting^[4][5].

A growing divide between individual employee productivity and corporate enterprise earnings has sparked an operational pivot among software architects and enterprise strategists^[1][1]. According to an analysis published by SAP on October 19, despite widespread workforce adoption - where 80% of generative AI users report significant boosts in day-to-day task speed - only 6% of enterprises report a measurable increase in their operating profit attributable to generative AI^[1][1]. This stark divergence has exposed structural limitations in standard large language models (LLMs) when deployed to resolve complex, revenue-critical business challenges^[1][1].

Industry researchers point out that while text-centric LLMs excel at drafting reports, synthesizing research, and generating code, they struggle with core enterprise tasks centered on structured data^[1][1]. Benchmarks conducted across enterprise systems in 2026 reveal that even frontier general-purpose language models experience performance deficits when executing table-reasoning tasks, evaluating multidimensional rows and columns, or predicting balance sheet fluctuations^[1]. Because industrial economies - such as mechanical engineering, complex logistics, and supply chain manufacturing - rely heavily on relational databases rather than unstructured text, pure LLM deployments have largely failed to mitigate corporate risks, optimize liquidity, or reduce supply chain defaults^[1][1].

To capture projected macroeconomic value - such as the €440 billion in additional gross value that the German Economic Institute projects AI could generate for Germany by 2034 - enterprises are moving away from monolithic language wrappers toward hybrid architectures^[1][1]. This emerging architectural trend couples LLMs (serving as natural language interfaces) directly with specialized tabular foundation models, including architectures such as SAP RPT-1 and TabPFN^[1]. Analysts stress that achieving genuine total factor productivity growth will require autonomous decision agents trained specifically to process numeric relational schemas, turning static corporate metrics into predictive operational actions^[2][1].

This transition reflects a broader recalibration across the AI economy as the market moves beyond initial experimental rollouts^[3]. With corporate software budgets under heightened scrutiny, enterprise leaders are demanding that AI investments demonstrate clear, audit-ready operational cost reductions and balance-sheet impact^[1][1]. Rather than treating generic generative conversational agents as an end-state solution, businesses and technology developers are refocusing their investments on specialized decision intelligence, domain-specific models, and explainable data systems capable of driving tangible operating margins^[1][3].

The U.S. Department of Homeland Security has ordered federal civilian agencies to remove DeepSeek's software and related services, citing concerns that sensitive government information could be exposed through the company's China-based infrastructure. The restrictions also prohibit agencies from procuring new DeepSeek products.

DeepSeek has attracted global attention for developing AI models that are widely accessible and can be used by developers and businesses. The U.S. has also raised concerns about the company's alleged use of distillation techniques to reproduce capabilities associated with American AI systems, an accusation Beijing disputes.

China's Foreign Ministry condemned the restrictions, arguing that Washington is politicizing technology and undermining Chinese companies. The dispute comes amid broader restrictions on advanced semiconductor exports and growing scrutiny of AI systems used in sensitive government environments.

According to a Goldman Sachs report cited by Axios, global data-center electricity demand is expected to rise by 170% between 2025 and 2030, with more than 60% of the increase occurring in the United States. The expansion reflects the growing computing requirements of AI model training, inference and cloud services.

However, power grids, electricity generation and related infrastructure are not expanding quickly enough to meet projected demand in some areas. Regulatory complexity, lengthy construction timelines and community opposition are contributing to delays. Axios reports that 86 infrastructure projects have been canceled or stalled since January, although the report's figures cover infrastructure projects rather than AI models themselves.

The constraints could affect where technology companies build data centers and how much they pay for electricity and computing capacity. They also increase the importance of utilities, grid upgrades, energy generation and long-term power agreements in determining which AI projects can move forward.

Anthropic has proposed committing 5% of its shares to a new Anthropic Public Benefit Fund LLC. The arrangement is designed to direct resources toward public-benefit initiatives, including areas such as AI safety and biomedical research. The initial commitment remains subject to shareholder approval.

Reporting on the announcement has put the potential value of the allocation at tens of billions of dollars, depending on the company's valuation. The proposed structure could also allow the fund to receive additional shares as Anthropic's valuation increases, although the precise value and eventual distributions will depend on the final terms and company performance.

The announcement comes as Anthropic prepares for a potential IPO, reportedly targeted for November 2026. The timing places the company's public-benefit commitments alongside questions about its valuation, financing requirements and long-term approach to AI development.

South Korea's KOSPI index dropped nearly 4% on October 20, according to Reuters reporting, as technology shares came under pressure. Elice Group, an AI-focused cloud startup, fell substantially below its IPO price during its market debut, adding to concerns about investor appetite for AI infrastructure businesses.

The market reaction comes as major technology companies commit enormous sums to data centers, chips and computing capacity. Investors are increasingly examining whether AI-driven revenue can grow quickly enough to cover capital expenditure, financing costs and the operating expenses associated with these investments.

The pressure is not limited to South Korea. Investors are monitoring upcoming earnings from major technology companies for evidence that AI spending is translating into sustainable profits and cash flow. The latest moves indicate greater selectivity toward AI investments rather than a uniform loss of confidence in the technology.

World, the identity technology company co-founded by Sam Altman, is extending its digital identity infrastructure toward enterprise applications involving AI agents. The initiative is intended to help organizations establish whether an automated agent is associated with a verified human identity, particularly when the agent participates in sensitive communications or financial workflows.

World's existing identity system uses iris scanning through its Orb device to establish a digital identity. The company has also pursued integrations that allow people to verify their identities in digital interactions. Its agent-focused initiative applies that identity infrastructure to a growing challenge: determining which AI systems are acting on behalf of real, authorized users.

The development comes as AI agents become more capable of interacting with websites, software and business systems. Identity verification could be one component of a broader security approach, but it does not by itself guarantee that an agent is safe, authorized for every action or protected against misuse.