The Uneven AI Machine

AI, the macro economy, and the politics of adjustment

Most discussions on artificial intelligence and the economy focus on whether AI will raise output. It almost certainly will. But the key questions are when, through which prices, to whose benefit, and how the political system reacts to the answer. Growth that is real but unevenly captured is not a benign outcome. It is the precise condition under which democracies become difficult to govern.

Growth: two distinct stories that are often confused

There are two AI growth stories running simultaneously. The first is the capital expenditure boom: data centres, semiconductors, power generation, grid connection, cooling, land. This is demand-side, immediate, measurable, and highly concentrated in a handful of firms and geographies. It is flattering GDP now. The second is diffusion or how firms reorganise their work given a new general-purpose technology. This is where the productivity gains unfold. Electrification took roughly four decades to impact factory productivity, because the gains required rebuilding the factory, not merely replacing the steam engine. The complementary investment - i.e. the process redesign, data infrastructure, management capability and retraining - is far larger than the technology spend itself and is mostly intangible, so it depresses measured productivity before it lifts it.

The implication may be uncomfortable for markets. The Capex story and the productivity story do not arrive together. This gap between them will test valuations.

Inflation: a relative price shock, not a level shock

The general assumption is that AI is disinflationary. More precisely, it is disinflationary in what it substitutes for and inflationary in what it requires.

On the disinflationary side, the marginal cost of cognitive output including drafting, analysis, translation, coding, first-pass research and routine client service, is collapsing. Anything whose price reflected the scarcity of trained human attention faces sustained downward pressure. On the inflationary side, electricity, grid infrastructure, turbines, transformers, copper, industrial land near substations, construction labour, and some engineering talent are increasing. These are physically constrained inputs with long lead times and inelastic short-run supply. Power in particular has moved from a commodity assumption to a binding constraint.

On a net basis, this looks less like a shift in the aggregate price level and more like a large, persistent change in relative prices with cheap intelligence, expensive energy and physical capacity. That could be a genuinely difficult context for central banks, because the aggregate index conceals two opposing forces of uncertain relative magnitude and impact on businesses and people. From a labour perspective, when wage growth softens, is that slack, or is it technology raising output per worker? In such an environment monetary policy errors can become more recurrent. During the Capex boom phase, the risk is to tighten into what is really a relative price shock in energy and industrial capacity, while underestimating how much potential output has risen. Later, as diffusion displaces labour, the risk inverts as structural disinflation looks like demand deficiency, and easing into it inflates assets rather than activity.

Winners and losers

Owners of capital appear to be the most likely winners. Energy producers, grid infrastructure, semiconductor and hardware supply chains are benefiting and so are incumbents holding proprietary data and distribution as their moats deepen rather than erode. A portion of workers who direct AI systems rather than compete with them will also likely benefit.

On the other end, mid-tier cognitive labour, where the work is structured, documented and therefore learnable are most likely to be disrupted. Entry-level roles in law, accounting, consulting, research and coding, where the tasks used to structure junior's apprenticeship will be impacted along with outsourcing-dependent economies whose comparative advantage was cheap, English-speaking, process-driven cognitive work. Firms whose competitive position rested on process efficiency rather than on data or customer relationships will also face difficult moments.

The asset owner's problem. For institutional investors there is a subtler risk. The gains are accruing to an extraordinarily concentrated set of firms, which means index exposure is now a concentrated technology bet whether or not that was the intention. Diversification has quietly become harder to achieve at exactly the moment it matters most.

The social adjustment:

Previous automation waves fell on manufacturing communities that were geographically concentrated, which made the resulting dislocation visible, locatable, and eventually impossible for policy to ignore. AI is likely to invert that pattern. It falls on cognitive work that can be dispersed across offices rather than clustered around plants and spread thinly through every regional economy rather than concentrated in a few. That dispersion matters because dislocation that is spread thinly is harder to measure, slower to register in the data, and correspondingly slower to attract a policy response. It arrives as a gradual thinning of opportunity rather than as an event.

As AI erodes the entry-level opportunities for many, the effect is not a redundancy statistic but an absence of roles that are no longer created. Nothing appears drastically in the unemployment figures as what changes is the rate at which careers begin.

From a social perspective, one feedback loop that is worth watching closely is how the physical footprint of AI is becoming a local planning question, and local planning questions have a way of scaling. Data centre demand raises local electricity prices and competes for water and land. Household energy bills are among the most politically sensitive prices in any developed economy, and several jurisdictions have already moved to constrain data centre siting.

A second important feedback loop is how entry level workers train to enter the workforce successfully. Professional competence has historically been built through apprenticeship with juniors learning judgement through training. AI substitutes for that work without substituting for the learning embedded in it, which puts the mobility ladder itself in question. The response is not only to teach the tools, which change faster than any curriculum can follow but also to teach judgement under uncertainty and ambiguity. When producing an answer costs nothing, the scarce skill is recognising which answer is wrong. That requires depth in a discipline alongside breadth across them, and a degree of agility that most professional formation has never had to develop.

Conclusion: While technology is an important variable to watch, adoption speed, the physical constraints on energy, and the political tolerance for concentrated gains are critical. With dislocation dispersed and benefits concentrated, it is the unevenness rather than the aggregate that will determine how this is experienced. The real question is therefore not only whether the pie grows, but how narrow the hands are that hold it, and how long the rest are prepared to wait.

For organisations, that argues against treating AI as a procurement decision. The gap between those firms who rebuild process, decision rights and talent models around the technology and those who bolt it onto existing structures will widen faster than most planning cycles assume. The same asymmetry applies between economies: those with power, compute and capital will pull away from those without, and the gap will widen inside economies as much as between them.

The binding constraint, though, is not only capital or compute. It is human capital or the far scarcer capacity to exercise judgement over systems that produce fluent answers at no cost. Organisations have spent decades optimising for process efficiency in exactly the layer that AI now performs, and rather less on building the discernment that sits above it. The firms that pull ahead will be those that treat this as an investment question rather than a training question, and that accept they must now build deliberately the judgement their apprenticeship structures used to produce as a by-product.