Wed 25 March 2026

The venture capital landscape in 2025 reveals a stark concentration of funding in AI startups, capturing 41% of the $128 billion raised via Carta, with just 10% of companies—led by giants like Anthropic, OpenAI, and xAI—securing half of all VC dollars through massive rounds exceeding tens of billions at valuations approaching trillions. This K-shaped market favors a select cadre of investors betting big on AI's computational demands, yielding promising internal rates of return for 2023-2024 funds, outpacing earlier vintages, largely due to rapid up-rounds in AI-native portfolios. Yet, these metrics reflect paper gains rather than realized exits, raising skepticism about sustainability amid hype-driven valuations; while IPO teases excite markets, the true test lies in whether this fervor translates to blockbuster liquidity events or exposes an overinflated bubble vulnerable to economic shifts or technological plateaus. Key points: • AI startups claimed 41% of $128B in VC funding on Carta in 2025, a record high. • Top 10% of startups, including OpenAI, Anthropic, and xAI, absorbed 50% of total VC dollars via multi-billion rounds. • Recent VC funds (2023-2024) show highest IRR due to AI portfolio up-rounds, contrasting declining returns from 2017-2020 vintages. • Market bifurcation concentrates capital in few firms backing high-cost AI models, with larger but fewer rounds. Expert question (counterfactual): What if the elevated IRRs from AI up-rounds fail to materialize into viable exits, leaving investors with illiquid assets in a post-hype correction?

The article astutely applies the historical 'productivity J-curve' to AI, positing an initial economic dip as firms invest in intangible assets like worker retraining and workflow redesign, before eventual surges in output—much like electricity's delayed impact via Ford's assembly lines. While this framework explains current modest productivity gains despite AI's hype, it warrants skepticism: past technologies transformed industries over decades, but AI's rapid scalability might compress timelines, amplifying disruptions. Economist Levy Yeyati's concern about fast adoption overwhelming retraining pipelines is compelling, potentially leading to labor force exits and unequal outcomes. Yet, the piece's call for proactive policies—expanding retraining and mobility—assumes government can nimbly intervene without bureaucratic drag, a optimistic view given historical policy lags in tech transitions. Overall, it reframes AI's bottleneck as organizational, not inventive, urging preparation over panic. Key points: • AI adoption follows a productivity J-curve: initial dip from intangible investments, then surge. • Rapid AI integration risks clogging retraining pipelines, causing worker displacement and labor force exits. • Historical precedents like electrification show tech impacts lag invention by decades due to business reorganization. • Public policy should preemptively boost retraining and labor mobility to mitigate transition bumps. Expert question (counterfactual): What if AI's software-driven nature allows faster organizational adaptation than hardware-based technologies like electricity, potentially flattening the J-curve and minimizing the predicted dip?

This NBER working paper, drawing on a survey of 750 corporate executives, reveals a mixed picture of AI's impact on productivity and labor markets, underscoring significant adoption—over half of firms have invested, though smaller ones lag—yielding positive but uneven labor productivity gains, particularly in high-skill services and finance, projected to intensify by 2026. Notably, these stem from total factor productivity enhancements via innovation and demand channels rather than mere capital deepening, yet a 'productivity paradox' emerges where executives perceive greater benefits than measured outcomes, possibly due to lagged revenue effects. On employment, aggregate declines appear minimal in the near term, but larger firms foresee reductions while smaller ones anticipate growth, alongside labor reallocation from routine clerical to skilled technical roles. Skeptically, the self-reported perceptions may inflate optimism, warranting caution against overhyping AI's transformative potential without robust longitudinal data. Key points: • Over 50% of firms have invested in AI, with gains concentrated in high-skill sectors like finance and services. • Productivity paradox: perceived gains outpace measured ones, linked to delayed revenue realization. • Minimal near-term aggregate job losses, but reallocation from clerical to technical roles expected. • Larger firms predict workforce reductions; smaller firms expect modest employment increases. Expert question (counterfactual): What if the productivity paradox reflects measurement flaws in capturing AI's intangible benefits, such as improved decision-making, rather than mere revenue delays—how might this alter projections for long-term economic growth?

While headlines amplify fears of AI-driven job apocalypse, a NBER survey of 750 U.S. CFOs reveals a more tempered reality: only 44% anticipate AI-related layoffs this year, projecting 502,000 job losses—a ninefold increase from 2025's 55,000 but a mere 0.4% of the workforce. This contrasts sharply with alarmist predictions from AI luminaries like Mustafa Suleyman and Dario Amodei, suggesting executives' private assessments prioritize fiscal caution over hype. Skepticism arises from the persistent productivity paradox, where AI investments yield perceived but not yet realized gains, echoing Solow's 1987 observation on computers. Reports of AI increasing workflow strain for workers further undermine claims of immediate efficiency. Nonetheless, small firms may offset losses by hiring technical roles, hinting at uneven sectoral impacts rather than wholesale disruption. Key points: • CFO survey indicates 502,000 AI-attributed job losses in 2026, up 9x from 2025 but only 0.4% of U.S. workforce. • 44% of CFOs plan AI-related cuts, with half affecting white-collar jobs. • Perceived AI productivity gains exceed actual results, mirroring Solow's paradox. • Small firms plan to increase technical hiring, potentially offsetting some losses. • AI leaders' doomsday predictions contrast with executives' subdued private views. Expert question (counterfactual): What if the lagged productivity gains from AI materialize unevenly across sectors, leading to concentrated job displacements in vulnerable industries rather than the diffuse 0.4% impact projected?

Martha Gimbel's analysis in Nature tempers the hype surrounding AI's labor market impact, asserting that despite CEO announcements and layoff attributions, empirical data reveals only modest effects since ChatGPT's 2022 debut. Aggregate employment patterns show no accelerated shifts in job distributions or prolonged unemployment for AI-vulnerable roles, contrasting with fears of a rapid Industrial Revolution-style disruption. Adoption remains low—merely 18% of US businesses report recent AI use—suggesting executive rhetoric often outpaces practical implementation, potentially inflating public anxiety. While acknowledging historical precedents like computing's gradual workforce reconfiguration, Gimbel urges enhanced data systems to track transitions, highlighting policy risks from speculative interventions. Skeptically, one wonders if this calm masks uneven sectoral impacts or underreported displacements in non-traditional metrics, underscoring the need for nuanced, forward-looking scrutiny beyond current aggregates. Key points: • No significant employment shifts observed since 2022 AI advancements, per Yale and Brookings research. • AI adoption in US businesses is low at 18%, with executives' announcements often exceeding actual use. • Fears of job apocalypse driven more by poor data and hype than evidence of structural disruption. • Better data collection essential to identify at-risk workers and inform targeted policies. Expert question (counterfactual): What if AI's effects are manifesting unevenly across sectors or demographics not captured by aggregate data, potentially delaying visible disruptions until a tipping point?

The article compellingly argues that the entry-level job market's deterioration, evidenced by a 37-year high of 13.3% new workforce entrants among the unemployed in 2025, stems from structural failures rather than generational shortcomings in Gen Z. While data on job losses in finance and information sectors—down 9,000 monthly since 2023—and the narrowing unemployment gap for college graduates underscore a fracturing ladder for young workers, the piece's dismissal of attitudinal critiques feels somewhat absolute, potentially overlooking how economic pressures might amplify perceived entitlement. The rise in side hustles (57% for Gen Z vs. 21% for Boomers) signals resilience but also precariousness, and AI's threat to entry-level roles, with a 13% employment drop for young AI-exposed workers, demands scrutiny. Policy prescriptions for apprenticeships and worker voice in AI deployment are apt, yet their feasibility amid entrenched corporate hesitancy warrants skepticism, as broader economic retooling remains elusive. Key points: • Share of new workforce entrants among unemployed reached 13.3% in 2025, highest in 37 years, exceeding Great Recession levels. • Finance and information sectors lost 9,000 jobs monthly since 2023, reversing pre-pandemic gains of 44,000. • 57% of Gen Z engage in side hustles for income, compared to 21% of Baby Boomers. • College graduates now face higher unemployment than average workers, with skilled trade associates outperforming them in 2025. • AI has caused 13% employment drop for 22-25-year-olds in exposed occupations since 2022.

The assertion that reliability in AI agents must be viewed as a systemic property, drawing from high-reliability organizations (HROs) like aviation or nuclear facilities, underscores a critical gap in current AI development. While individual agents may perform competently in isolation, the snippet highlights vulnerabilities in interconnected operations—such as handoffs between agents, escalation protocols, and human intervention triggers—which are foundational to HRO resilience. This perspective is compelling yet warrants skepticism: AI builders often prioritize agent autonomy for scalability, potentially underestimating the complexity of systemic safeguards. Without robust orchestration layers, agentic systems risk cascading failures, echoing past tech incidents where siloed components amplified errors. Policymakers should scrutinize whether self-regulation in AI firms adequately addresses these interdependencies, lest hype around autonomous agents overshadows the need for holistic reliability frameworks. Key points: • Reliability emerges from system-wide design, not isolated agent performance. • Current AI tools falter in agent-to-agent handoffs and escalation mechanisms. • Human involvement remains essential for high-reliability AI operations. • Lessons from HROs like aviation can guide AI agent builders toward systemic robustness. Expert question (counterfactual): What if enhancing individual agent intelligence alone could achieve reliability without overhauling systemic handoffs and escalations—would that undermine the HRO analogy?

The EU's AI Omnibus proposal represents a pragmatic recalibration of the ambitious but cumbersome AI Act, addressing business complaints about overlapping regulations and unrealistic timelines by extending deadlines for high-risk systems to 2027-2028 and prioritizing sector-specific laws to minimize duplication. While this 'stop-the-clock' mechanism offers breathing room—potentially easing compliance costs in sectors like healthcare and finance—it introduces prolonged uncertainty amid ongoing trilogues, with final agreement targeted for spring 2026. The addition of a ban on 'nudifier' AI underscores retained safeguards against harmful uses, yet the spin of 'simplification' warrants skepticism: without finalized technical standards, companies may still face fragmented enforcement, and the grace periods could inadvertently slow innovation if not paired with robust guidance. Overall, this balances regulatory ambition with feasibility, but success hinges on institutional alignment to avoid further delays. Key points: • Extended compliance deadlines for high-risk AI to December 2027 or August 2028. • Sector-specific regulations take precedence over AI Act to reduce overlap. • New deadline of November 2026 for generative AI labelling requirements. • Prohibition on AI tools generating non-consensual intimate images. • Support extended to small mid-cap enterprises for AI compliance. Expert question (counterfactual): What if trilogue negotiations fail to align on key details, forcing businesses back to the original AI Act timelines and amplifying rather than alleviating compliance chaos?

AI-driven data centres are outpacing the capacity of the U.S. power grid, with global investment projected to hit US$6.7 trillion by 2030, US$2.7 trillion of that in the U.S. alone. AI workloads have pushed rack power demand from 5–10 kW to 50–100 kW, creating major strain on grid capacity, cooling, and siting. In Texas, where data centre consumption could reach 78 GW by 2031 (around 36% of statewide electricity demand), interconnection delays have forced operators to explore off-grid ‘island-mode’ power systems. To meet “five nines” (99.999%) uptime, many new centres rely on natural gas plants—typically 1 GW per 40 acres—since renewables lack consistent output without large-scale storage. However, both gas equipment supply chains and regulatory oversight are tightening. Senate Bill 6 in Texas adds disclosure rules for facilities above 75 MW, while federal proposals under President Trump aim to ease regulations for fully off-grid projects. Meanwhile, public resistance is mounting: by early 2025, US$64 billion in data centre projects were delayed over environmental and community concerns. As legal, technical, and social pressures converge, power availability—not cost—is now the key constraint shaping AI infrastructure in America.