Every technological revolution forces investors, operators, and policymakers to answer the same question: where will the next trillion dollars of value come from? For AI, the honest answer is not that a new economy is being built from nothing. It is that a very old economy — the global stock of about $500 trillion in accumulated wealth — is about to be rewired through a new stack. The critical question is not whether AI is valuable. It is how AI creates new wealth, and who captures it.
Critics argue the AI boom looks like every previous bubble — hundreds of billions of CapEx chasing a demand curve that has not yet arrived. Supporters argue AI is a foundational technology in the class of electricity, the internal combustion engine, and the internet. The truthful position sits between them. The infrastructure spend is real, the demand is real, and the wealth transfer is real — but the mechanism by which value accrues is not evenly distributed across the stack. Understanding where it lands is the entire investment thesis.
This briefing does three things. It sizes the pool AI is entering. It identifies the three mechanisms by which AI creates new wealth. And it names the layer where the value is most likely to accumulate — the deployment layer, controlled by whoever owns data, distribution, and industry-native operating context.
The map: what AI is actually entering.
To understand AI's potential, we first have to understand the size of the economic system it is entering. Below is the global wealth pool as it stands today, sorted by scale. What becomes immediately obvious is that AI does not need to invent an entirely new asset class. Capturing even a small percentage of what already exists creates enormous value.
The economy AI is rewiring.
A 1% productivity improvement applied across the annual output line — the $115 trillion of global GDP — represents $1.15 trillion of new economic output every year. That figure is larger than the GDP of most countries. It is also a floor, not a ceiling; the historical impact of general-purpose technologies on productivity has run into double digits over sufficient time horizons. AI does not need to create magic. It needs to compound modestly against a very large base.
Five eras. Five different definitions of wealth.
Every major economic era in modern history has been built on a different wealth creation mechanism, and each has produced a different set of winners. The pattern is worth internalising — because the AI era is not a new species of transition. It is the fifth in a recognisable sequence, and the shape of the winners is knowable from the pattern of the previous four.
Every era has an asset, a driver, and a winner.
Agriculture
Industrial
Oil
Internet
AI
Read the row across the bottom. Every era's winners share a structural property: they sit at the intersection of the era's dominant asset and the customer who needs it deployed. Landowners did not invent the wheat. Industrial conglomerates did not invent the steam engine. Energy producers did not invent the piston. Platforms did not invent the packet. In each case, the winner owned the bridge between the raw asset and the commercial application. That is the pattern to hold in mind for AI.
The three mechanisms of AI wealth creation.
AI does not create value from nothing. It creates it through three distinct mechanisms, each with a different economic character and a different set of winners. Any coherent thesis on AI wealth creation has to name which of the three it is betting on.
The first mechanism is the amplification of knowledge work. Software engineering, legal research, customer support, sales operations, marketing execution, financial analysis, procurement, and internal communications are all seeing measurable productivity gains in the range of 20–40%. Apply that to the global stock of cognitive labour and the number is denominated in trillions. This is the AI equivalent of what mechanisation did for physical labour during the Industrial Revolution — and the historical pattern is that the winners of a productivity wave are rarely the companies that build the tooling. They are the companies that deploy it fastest against a large operational base.
The second mechanism is the collapse of transaction costs. Historically, firms have spent enormous resources identifying buyers, qualifying demand, personalising outreach, and running long sales cycles against imperfect information. AI compresses that cost curve dramatically. Markets become more efficient because buyers and sellers find each other faster, and with higher signal. Most of the value created in modern B2B go-to-market systems over the next five years will come from this mechanism — and it is the reason categories like industry intelligence, buyer-intent data, and decision-maker networks are structurally undervalued relative to where they will sit in 2030.
The third mechanism is entirely new economic activity. AI-native software firms. Agentic service providers. Autonomous commerce. Personal AI advisors. Digital workers. AI-generated products and media. The internet created Google, Amazon, Salesforce, Shopify, and Uber — companies that did not merely digitise an existing business but invented a new category. AI will do the same, and by historical precedent, this is the mechanism from which the largest single fortunes will emerge. The dot-com equivalents of the AI era are being incorporated right now; most of the eventual winners are not yet visible.
Each mechanism has a different risk profile and a different time horizon. Production gains are the safest and fastest, but the value tends to leak to the deployers, not the builders. Distribution gains are the highest-quality compounding return and the least understood. New business creation carries the highest variance but also the highest ceiling. A serious portfolio thesis has to be explicit about which of the three it is underwriting.
Semiconductors: the oil infrastructure of intelligence.
Every economic revolution requires infrastructure. The Industrial Revolution required railways. The oil age required pipelines and refineries. The internet age required fibre. The AI age requires semiconductors — and the semiconductor industry has quietly become the strategic equivalent of oil infrastructure for the intelligence economy.
NVIDIA now powers the majority of frontier AI training and inference globally. Taiwan manufactures the majority of advanced AI chips through TSMC. The four US hyperscalers are collectively committing hundreds of billions of dollars per year to AI data centre CapEx — Alphabet alone is guiding to $180–190B in 2026. Without this substrate, there is no AI economy. This is not a controversial claim; it is the observable ground truth of the last three years.
The historical analogy is instructive but incomplete. Oil producers captured enormous rents in the twentieth century, but the largest fortunes were made downstream — by the refiners, the distributors, and the companies that put oil-derived products into the hands of consumers. Semiconductors will follow a similar arc. The chip layer will remain strategically indispensable and profitable, but the disproportionate value will accrue to the layers that translate that raw compute into industry-specific commercial outcomes.
The capital concentration effect.
One of the most remarkable structural features of today's economy is the extraordinary concentration of capital at the top of the market. A handful of companies now command balance sheets larger than the aggregate value of entire national exchanges.
The six largest technology companies collectively represent more market value than most national stock markets. This concentration is not incidental to the AI transition — it is causal. Large firms with fortress balance sheets can invest tens of billions in AI infrastructure before competitors can react, before regulators can respond, and before the market has priced the outcome. That structural advantage is why AI adoption is accelerating faster than any previous technology cycle in modern economic history.
The strategic consequence for anyone building on top of that substrate is direct: the hyperscalers will win the compute-and-model layer, and probably by a wider margin than they won the cloud layer. The interesting question is not whether they win there. It is what layer sits above them — and who captures the returns from routing global industry activity through that layer.
Infrastructure without monetisation — the historical warning.
History teaches a specific and uncomfortable lesson: infrastructure booms do not always create proportional shareholder returns. Railways transformed the nineteenth-century economy and bankrupted a generation of investors along the way. Telecommunications transformed how the world communicated and destroyed a trillion dollars of capital in the dot-com crash. The internet created a handful of generational winners and thousands of losers, most of whom had built genuinely useful technology.
AI may follow the same path. The infrastructure layer — data centres, chips, models — is receiving hundreds of billions in investment. The open question is whether the demand and monetisation curve will keep pace with the CapEx curve. If it does, the returns compound spectacularly. If it does not, we are watching a repeat of 1999–2001, in which the ultimate winners are visible in retrospect but not in real time.
The prudent read is that the demand is real, the timing is uncertain, and the biggest returns will accrue not to the layer receiving the CapEx but to the layer that turns that CapEx into industry revenue. This is where the deployment argument becomes the load-bearing part of any AI investment thesis.
The three D's: Data, Distribution, Deployment.
Most discussions of AI focus on models. A smaller number focus on chips. Very few focus on deployment. Yet history suggests, with unusual clarity, that the greatest value in any technological revolution accrues not to those who invent the technology but to those who deploy it at scale into specific commercial contexts.
The organisations that will control the AI economy of 2030 own three assets. Each is more difficult to build than a foundation model. Each is more defensible than a chip design. And each is closer to the customer than anything the hyperscaler stack can offer.
Not the public internet, which the hyperscalers have already indexed and which trains every frontier model equally. The valuable data is the first-party operational signal that only flows through a real industry network — buyer behaviour, deal patterns, churn signals, workflow telemetry, decision-maker movement. This is the raw material that turns a general-purpose model into a useful operator inside a specific vertical. Whoever owns the pipe owns the layer above it.
Boring, expensive, slow to build, and structurally impossible for a horizontal platform to replicate at economics that make sense. Editorial authority, peer networks, events, communities, channel partners, decision-maker access. These are the mechanisms by which a new product actually reaches the buyer who needs it. The hyperscalers will not build them. The AI labs will not build them. The system integrators do not have the product instinct to build them. Whoever does build them owns a compounding position.
The layer where horizontal AI capacity becomes vertical operational outcome. Sector-specific workflows, compliance regimes, procurement cycles, unit economics, and organisational tempo. Deployment is the layer that translates a $300B+ platform tier into $30T+ of industry revenue over the next decade. Just as oil required refineries and distribution networks to become gasoline in a car, AI requires deployment networks to become revenue in a business. This is the third phase — and it is where the next generation of winners will be built.
Phase one of the AI economy has been dominated by infrastructure — chips, data centres, energy. Phase two, playing out now, is dominated by foundation models. Phase three, still ahead, will be dominated by deployment. The next generation of winners will not simply build AI. They will operationalise AI across industries, workflows, and revenue systems — packaging horizontal capacity into vertical outcomes and selling them into buyers the hyperscalers do not know how to reach.
The rewiring, priced.
The opportunity is not merely to create better technology. The opportunity is to create new wealth — and in a world containing roughly $500 trillion of accumulated wealth and $115 trillion of annual output, even a small shift in productivity, distribution, and decision-making creates trillions of dollars of new value. The maths does not require heroic assumptions.
The critical question was never whether AI is valuable. It was how AI creates new wealth, and who captures it. The answer is now visible in the shape of the stack: three mechanisms of value creation, one substrate that is already priced, and one layer above it that is not yet built. The AI economy has only just begun. The infrastructure is in the ground. The models are in production. The deployment layer — data, distribution, deployment — is the trillion-dollar category that still sits open.
Naming the mechanism is the first piece of work. Building the deployment layer is the rest of the decade.