AI is making software and cognitive work more abundant — while concentrating the infrastructure that produces them

Acceleration is economic, not only technical

Stanford's AI Index 2026 estimates that global corporate AI investment reached $581.69 billion in 2025, more than double the previous year. Private investment totaled $344.66 billion and mergers and acquisitions $214.44 billion. Organizational AI adoption rose from 55% in 2023 to 88% in 2025, while generative AI reached 53% adoption in three years. Stanford AI Index 2026

These figures do not establish proportional economic returns. They establish that firms and investors are reallocating capital and processes at a scale large enough to affect labour markets, infrastructure and product strategy.

Investment is also highly uneven. Private AI investment in the United States reached $285.88 billion in 2025, compared with $20.92 billion in Europe and $12.41 billion in China. The private-investment metric understates part of China's state-backed financing, so it should not be read as a complete measure of national technological capacity. Even with that qualification, US financial concentration is material. Stanford AI Index 2026

The chart compares the same private-investment category. It does not measure public spending, total compute capacity or model quality.

Digital abundance is creating physical scarcity

A lower cost of producing software does not make the AI economy capital-light. The International Energy Agency estimates that capital expenditure by the largest technology companies exceeded $400 billion in 2025 and is expected to rise another 75% in 2026; the 2026 figure remains an estimate. The IEA also notes that capital investment by only five technology firms now exceeds global investment in oil and natural-gas production. IEA — Key Questions on Energy and AI

Physical infrastructure expands because more sophisticated tasks do not follow the same energy curve as a simple query. The IEA estimates that energy per AI task has fallen by at least an order of magnitude annually in recent years, while reasoning, video and agentic tasks can consume hundreds or thousands of times more energy per query than simple text generation. Higher efficiency can therefore coexist with higher aggregate consumption.

Global data-centre electricity consumption was estimated at 485 TWh in 2025 and could reach about 950 TWh in 2030, roughly 3% of global electricity demand that year. Consumption by AI-focused data centres grows even faster. IEA

The same report identifies critical constraints in grid connections, electrical equipment, chips and high-bandwidth memory, alongside community resistance around electricity prices and environmental effects. This changes the political economy of AI: the constraint is no longer only algorithm quality but also grid availability, permitting, finance and local acceptance.

This dual path — digital abundance and physical scarcity — organizes the rest of the assessment.

Labour effects are appearing first in tasks and entry pipelines

The newest evidence does not show broad worker substitution. The August 2026 revision from the Stanford Digital Economy Lab, using ADP administrative payroll data covering millions of US workers through June, finds no evidence of widespread economy-wide displacement. However, workers aged 22 to 25 in AI-exposed occupations were 19% below the counterfactual path they would have followed had they kept pace with less-exposed peers. Experienced workers did not show a comparable gap. Stanford Digital Economy Lab

That pattern is consistent with a hiring mechanism rather than necessarily a layoff mechanism. Firms can retain experienced specialists while reducing the number of junior hires required for structured tasks that AI systems can partly perform.

The ILO estimates that one in four workers worldwide is in an occupation with some degree of generative-AI exposure, but only 3.3% of global employment falls into the highest exposure gradient. The organization emphasizes that task transformation is more likely than complete elimination for many occupations. ILO

The OECD adds a counterintuitive result relative to industrial automation: IT professionals, business professionals, managers, executives and science and engineering professionals are among the occupations most exposed to AI. Cleaners, agricultural workers, food-preparation assistants and elementary manual workers are among the least exposed. OECD — Skills in the AI age

The junior career ladder may become a major social transmission mechanism

Skilled professions are systems for producing experience. Senior developers were once juniors; lawyers who review decisions began by researching cases; analysts who judge models spent years building spreadsheets and presentations; managers learned by doing operational work.

Many of those early tasks are structured, digitized and verifiable — the same properties that make an activity relatively compatible with automation.

This creates a risk of compressed apprenticeship ladders. A firm can gain near-term productivity by letting an experienced worker supervise more output with agents, while the market produces fewer people capable of becoming experienced workers later.

Traditional path
junior performs structured tasks → receives review → accumulates context → handles exceptions → becomes specialist

Automation-intensive path
agent performs structured tasks → specialist reviews exceptions → fewer entry roles → less practical learning volume → possible future specialist bottleneck

The second path is an analysis, not an established outcome. Evidence also points in the opposite direction: Brynjolfsson, Li and Raymond's study of 5,179 customer-service agents found a 14% average productivity gain and 34% among novice and lower-skilled workers, suggesting AI can accelerate learning and diffuse best practices. NBER

The institutional problem is therefore how to automate entry-level tasks without eliminating the supervised learning that creates future judgment.

Social exposure is not concentrated only among less-educated workers

Exposure is also uneven by gender and national income. The ILO estimates that globally, 4.7% of female employment is in the highest exposure gradient versus 2.4% of male employment. In high-income countries the figures rise to 9.6% and 3.5%, respectively. Total exposure reaches 34% of employment in high-income countries and 11% in low-income countries. ILO

The reason is not an intrinsic ability of AI to “replace women”, but occupational composition: clerical and administrative functions, where women have high shares in many countries, are among the most exposed.

At the same time, high-income economies hold more capital, infrastructure and firms positioned to capture productivity gains. Lower-income countries can have lower exposure today and still capture a smaller share of AI-generated income if they depend on foreign models, clouds and chips.

This creates an important distributional possibility: the place where work is transformed and the place where infrastructure income is captured can be different.

Which professions face pressure — and through what mechanism

A list of jobs that “will disappear” would be methodologically weak. The relevant object is the task, its verifiability, its contextual requirements and the cost of error.

The frontier can move. Low current exposure of physical work is not permanent immunity; it means the automation barrier lies in bodies, perception and reliable action in the physical world rather than language.

Software can stop being scarce before it stops being necessary

Software markets provide a live experiment in scarcity migration.

Software Equity Group's SaaS index recorded a fall in median EV/trailing-twelve-month revenue from 5.7x in Q2 2025 to 3.2x in Q2 2026. At the same time, the report describes larger, more profitable and more efficient companies, with relative premiums for DevOps/IT management, ERP/supply chain and security. Software Equity Group

This is valuation compression, not evidence that the sector is collapsing.

Contrary evidence matters. In September 2026, Stripe reported that it had added more new SaaS platforms in the previous three months than in the final six months of 2025, with new platform businesses more than 180% higher year over year. Stripe also said SaaS payment volumes in its own base were above levels seen before the large software selloff early in 2026. Stripe

Both observations can be true: software can continue growing as an economic activity while markets pay less for revenue that is easier to replicate.

The economic unit of classic SaaS is under pressure

The dominant SaaS-era model tied revenue to human users: more employees using a product meant more seats and more recurring revenue.

Agents break that relationship. If five agents and three employees can perform a process that once required ten employees, purely seat-based pricing loses revenue exactly when automation increases.

Andreessen Horowitz has described this as a shift from per-seat pricing toward usage and outcomes; that is an investor thesis, not a universal empirical law. a16z

There is operational evidence of adaptation. Salesforce's Agentforce offers consumption-based Flex Credits and meters agent actions, while retaining user- and conversation-based licensing options. Salesforce

The economic unit begins to move from “how many people use the software?” toward “how many actions, resolutions or outcomes does the software produce?”

That brings software closer to labour and professional-services budgets. An application no longer competes only with another licence; it can compete with human hours.

SaaS can respond by becoming less substitutable, not by becoming less software

Vertical platforms offer one example. Stripe reports that median payments adoption inside vertical SaaS platforms rose from 27% in 2024 to 40% in 2025, while leading platforms exceeded 80%. Stripe — Vertical SaaS

Payments, identity, compliance, historical data, integrations and critical workflows can raise switching costs even when interfaces and individual features become easier to reproduce.

Bessemer argues that traditional systems of record face pressure because AI lowers migration costs, structures unstructured data and lets software take action, while acknowledging that replacing large ERP systems remains difficult. This is again a venture-investor thesis, useful for mechanisms rather than proof that Salesforce, SAP or ServiceNow will be displaced. Bessemer — State of AI 2025

The implication is polarization: generic software can commoditize while systems controlling data, money, identity, compliance and critical workflows become more central.

The application layer can hypercompete while the base concentrates

The US Federal Trade Commission studied Microsoft–OpenAI, Amazon–Anthropic and Google–Anthropic and reported more than $20 billion in cumulative financial investment, alongside cloud commitments, economic rights and information access. The report does not declare an illegal monopoly; it identifies potential effects on access to compute and talent, switching costs and information asymmetry. FTC

In June 2026, the European Commission informed Amazon and Microsoft of its preliminary view that AWS and Azure should be designated as Digital Markets Act gatekeepers for cloud services. The Commission described them as the largest and second-largest cloud services in the EU. This is a preliminary position, not a final finding of market abuse. European Commission

Microsoft's 2026 10-K illustrates the physical intensity of scale. Servers, network equipment and software at cost rose from $132.836 billion to $215.874 billion in one year; total net property and equipment rose from $204.966 billion to $313.076 billion. The company also had $34.6 billion committed for new buildings and improvements, primarily data centres. Microsoft 2026 Form 10-K

Economic power does not disappear when code becomes cheap. It migrates toward complements that did not become cheap at the same speed.

Peter Thiel provides a lens for where income and profits accrue

In Competition is for Losers, delivered in Stanford's startup course and later published by Y Combinator, Peter Thiel distinguished value creation from where income and profits accrue. His “monopoly” strategy emphasizes technological differentiation, network effects, economies of scale and initial dominance of small markets. Y Combinator — Peter Thiel

This is a strategic lens, not evidence about 2026 market structure.

Applied to AI, it exposes a tension. The application layer is moving closer to the condition Thiel views as unattractive: many competitors, copyable features and low entry costs. Advanced semiconductors, cloud, contracted power and global distribution, by contrast, have strong scale economies and capital barriers.

The relevant question becomes less “who can build an app?” and more “who controls an asset that hundreds of apps cannot reproduce?”.

Large investors are making different bets on the same transformation

David Cahn at Sequoia framed the 2024 “AI's $600B Question” around the gap between infrastructure being built and application revenue needed to support it. In his 2026 thesis, he retained a two-speed view — growing adoption alongside physical constraints delaying data centres — and emphasized that transformative technology does not guarantee returns to all invested capital. Sequoia 2024 Sequoia 2026

Bessemer shows the other side of valuations. In its 2025 Cloud 100, 22 AI companies represented $464 billion, or 42% of total list value. AI companies carried a 24x average revenue multiple versus 19x for non-AI peers. Bessemer Cloud 100

Masayoshi Son represents the most concentration-oriented bet. In SoftBank's 2025 annual report he said he believed a small number of ASI leaders could eventually generate value equivalent to at least 5% of global GDP with margins approaching 50%; the group had committed up to $32.2 billion to OpenAI. These are the investor's expectations, not an independent base case. SoftBank

The distribution of productivity may matter more than the average gain

If AI raises a firm's productivity by 20%, that does not determine who receives the gain.

It can appear as higher wages, lower prices, higher margins, fewer employees, more output, shorter hours, or payments to cloud and model providers.

Market structure shapes distribution. When many firms can buy similar models, application-level advantages can be competed away quickly. When a small number of providers control infrastructure, data, channels and capital, more income can migrate toward those complements.

This creates five social mechanisms to monitor:

1. polarization within professions: specialists who supervise systems can expand output while entry roles shrink;
2. changes in bargaining power: workers compete not only against other workers but against combinations of automation, outsourcing and software;
3. wealth concentration: if gains are capitalized mainly in infrastructure owners, asset values can rise faster than labour income;
4. territorial asymmetry: regions hosting data centres can receive investment and tax revenue while also absorbing grid, water, permitting and community costs;
5. international technological dependence: countries can consume cheap intelligence through APIs without owning the infrastructure, models or channels capturing most income.

None is automatic. Competition, interoperability, open models, regulation, taxation, capital ownership, training systems and electricity-market design can change the outcome.

Education faces a training problem, not only a curriculum problem

The obvious educational response is to teach AI. That is insufficient.

When tools can produce first drafts of code, reports and analyses, schools and firms need to distinguish the ability to generate an artifact from the ability to explain, test and take responsibility for it.

Value shifts partly toward problem decomposition, testing, evidence interpretation, causal modelling, security, domain knowledge and judgment under uncertainty.

At the same time, removing all basic tasks from students and junior workers can reduce opportunities to build mental models. The challenge is to use AI to accelerate learning without eliminating the practice that produces judgment.

Evidence is not yet sufficient to establish which pedagogical arrangement works best at scale. This is a major monitoring question.

Emerging markets can gain access and lose income at the same time

AI lowers an important barrier for countries outside the main technology centres: access to cognitive capability. A small firm can use advanced models without building a data centre or training a frontier model.

But that benefit can be accompanied by income transfers to foreign compute, cloud and model providers. The concentration of private AI investment in the United States suggests that ownership and financing are not diffusing at the same speed as usage.

Outcomes depend on the layer in which an economy can compete. Countries that combine reliable electricity, data centres, fibre, local software, sectoral data, engineering talent and domestic markets can capture more value. Countries that remain consumers can gain productivity while increasing technological dependence.

This is a structural analysis, not a predetermined forecast. Open models, lower hardware costs, regional clouds and interoperability policy can reduce the asymmetry.

Four scenarios for 2026–2030

The scenarios below are not forecasts or probabilities. They identify mechanisms that can coexist.

The most plausible outcome may combine several: rapid diffusion at the application layer, concentration in some inputs, periodic valuation corrections and fragmentation in regulated sectors.

What would weaken this thesis

The scarcity-migration thesis should be revised if several signals appear together:

• the young-worker employment gap in exposed occupations persistently closes while AI adoption rises;
• per-seat models remain dominant even where agents execute a large share of workflows;
• generic-software valuations regain a premium without corresponding improvements in growth, retention or pricing power;
• compute and electricity costs fall enough to remove relevant scale economies;
• interoperability makes switching between clouds and models very low-cost;
• independent studies find that productivity gains do not persist in real workflows;
• new competitors gain enough share in cloud, chips and models to materially reduce concentration.

The objective is not to prove one structure inevitable but to keep the assessment falsifiable.

Indicators to watch

Marginal Thinking will monitor in particular:

• hiring of workers aged 22–25 versus experienced workers by AI exposure;
• revenue and output per employee in AI-native companies;
• the mix of per-seat, consumption, action and outcome pricing;
• revenue multiples and margins across traditional software, AI, security, data and infrastructure;
• cloud concentration and switching costs;
• hyperscaler capital investment and the economic return on data centres;
• electricity demand, grid queues and allocation of network costs;
• price and availability of GPUs, HBM and electrical equipment;
• independent productivity evidence at task and workflow level;
• labour-income shares in highly exposed sectors;
• startup formation versus concentration of revenue and value;
• evolution of junior training and apprenticeship structures inside firms.

Conclusion

The analytical error would be to choose between two extreme stories: “AI will destroy all jobs and software” or “AI is simply another productivity tool”.

Evidence available through 5 October 2026 points to a more structural transition.

Applied intelligence and software production are becoming cheaper across a growing set of tasks. That lowers entry barriers, pressures generic features and expands the reach of individuals and small teams. At the same time, the infrastructure that makes this abundance possible is becoming extraordinarily intensive in capital, energy, chips, grids and distribution.

In labour, the first visible effect is less aggregate layoffs than changing task composition and young-worker hiring. In software, the first disruption is less the disappearance of SaaS than the erosion of the seat as a universal economic unit and the compression of products without scarce assets. In industrial structure, competition can rise at the top while cloud, compute and energy remain concentrated at the base.

The central question for the next phase is not whether there will be more intelligence or more software. There will be.

The question is who controls the resources that remain scarce when intelligence and software become abundant — and how the income generated by that combination is divided among capital, labour, consumers, states and territories.

Principal sources

• Stanford HAI — AI Index 2026, Economy
• Stanford Digital Economy Lab — Canaries in the Coal Mine?, 2026 revision
• ILO — Generative AI and Jobs: refined global exposure index
• OECD — Skills in the AI age
• NBER — Generative AI at Work
• METR — 2026 developer productivity experiment update
• IEA — Key Questions on Energy and AI
• Software Equity Group — Quarterly SaaS Report
• Stripe — SaaS platforms are surging despite the SaaSpocalypse
• Stripe — Vertical SaaS insights, 2026
• Salesforce — Agentforce pricing
• FTC — Large AI partnerships and investments
• European Commission — AWS and Azure preliminary DMA position
• Microsoft — 2026 Form 10-K
• Y Combinator — Peter Thiel, Competition is for Losers
• Sequoia — AI's $600B Question
• Sequoia — AI in 2026: A Tale of Two AIs
• Bessemer — The State of AI 2025
• Bessemer — Cloud 100 Benchmarks 2025
• SoftBank — Masayoshi Son, Annual Report 2025
• Andreessen Horowitz — AI and outcome-based pricing