AI Economics Series Article 1 | The AI Economy Has Arrived. Now Comes the Hard Part

For most of the past three years, the public conversation about artificial intelligence has been a technology conversation — which model is best, how quickly the benchmarks are moving, when agents will finally be reliable enough to do complex work.

Those questions still matter. But AI is now large enough, capital-intensive enough and economically consequential enough that it needs to be analysed as an emerging economic system rather than a software feature. Whether the technology works is no longer the interesting question. Whether the economics work is.

That is the lens through which I read Exponential View’s The State of the AI Economy. Plenty of reports tell us AI is big. This one tries to quantify whether the demand underneath it is real, pulling the debate away from anecdotes, launches and market sentiment towards realised revenue, capital intensity, value capture and return on invested capital.

On that question the evidence is getting hard to dismiss. Exponential View puts the generative AI economy at an annualised run-rate of roughly $175 billion, after stripping out double-counting across applications, foundation models and infrastructure hosting, with trailing 12-month revenue of around $110 billion. Chip manufacturing is excluded; the figures cover deduplicated app, model and hosting revenues.

None of this means the AI economy is mature, or that every investment will pay off, or that bubble risk has gone away. It does mean we should be careful about describing AI as purely speculative. Real customers are paying real money for these capabilities. The live question is whether that demand can grow fast enough, and profitably enough, to support the infrastructure being built in anticipation of it.

AI has crossed from technology adoption into capital allocation

Every major technology shift eventually turns into an economics story. Nobody judged the internet on the elegance of TCP/IP; they judged it on the industries it enabled. Cloud mattered because it changed the economics of infrastructure, software delivery and enterprise technology investment — the virtual machines themselves were never the point. AI is entering the same phase.

The scale of investment is already extraordinary. McKinsey estimates that data centres globally could require around $6.7 trillion in cumulative capital outlays by 2030 to meet demand for compute, with AI workloads accounting for the majority of that requirement. (McKinsey & Company) Goldman Sachs has highlighted the growing role of private capital in financing the build-out, citing hyperscaler plans to spend more than $5 trillion on technology and data centres by 2030. (Goldman Sachs)

Once an industry needs trillions of dollars of infrastructure, product innovation stops being the right vocabulary for assessing it. Utilisation, depreciation, financing, pricing, margins, payback — that is the language it now has to be judged in, and it makes for a very different conversation from the one most organisations are currently having.

Today’s AI economy is still lower in the stack

One of the report’s more careful points is that the AI economy is not a single undifferentiated market. Revenue flows through several layers — applications, foundation models, hosting and chips — and the same customer dollar can show up as revenue for an application provider, a model provider and a hosting provider unless the analyst deduplicates properly. Exponential View distinguishes its deduplicated GenAI figure from stack-level revenue, which is not adjusted in the same way.

Even allowing for that caution, the direction of travel is clear: today’s AI economy remains heavily weighted towards compute, hosting, data centres and semiconductors. Applications and models are growing and gaining share, but the economic centre of gravity is still infrastructure.

There is nothing unusual about that. Railways needed track before they could transform logistics; electricity needed grids; the internet needed fibre, data centres and protocols; cloud needed hyperscale infrastructure before SaaS could reshape enterprise software. The first wave of visible value in AI is accruing to whoever controls the scarce enabling resource — compute. How long that lasts, and whether value gradually migrates upwards into models, applications and enterprise workflows, is the more interesting question.

The bubble question is too simplistic

Is AI a bubble? I find the framing too binary. A bubble can exist in one part of an economy while genuine value is being created elsewhere. In previous build-outs, infrastructure investment often ran ahead of near-term demand, and the assets later became the foundation for large new markets. The dot-com cycle is the obvious example: many companies were overvalued or disappeared, yet the internet was hardly a failed technology.

The same distinction applies here. Some valuations are probably excessive, some data centre plans will prove overbuilt, some model companies are structurally unprofitable, and plenty of enterprise AI programmes will fail to earn their keep. AI may also become one of the most important general-purpose technologies of the coming decades. Both can be true at once.

The more useful question is which parts of the AI economy are priced for perfection, and which remain underdeveloped. The lower stack is where today’s revenue and capital are concentrated, but much of the future opportunity in the upper stack is unresolved. The application layer is immature, enterprise workflows are still being redesigned, agentic systems remain unreliable in many production settings, and most organisations have not yet worked out how to turn AI capability into repeatable economic value. If there is bubble risk, it sits alongside a risk that gets far less attention — underestimating how much value in the upper stack is still untapped.

The depreciation test is the hard economic test

The most important financial question in AI is not whether models become more capable — that seems likely — but whether revenue, utilisation and pricing can support the capital base being created.

Here the Exponential View analysis earns its keep. The report forecasts a 2026 depreciation charge approaching $111 billion, based on assumptions around IT equipment and buildings. GenAI revenues now cover quarterly depreciation of AI infrastructure, but the coverage is thin: depreciation absorbs roughly 81% of hyperscaler and neocloud GenAI revenue, and roughly 68% of total GenAI revenue, before any other operating costs.

Revenue may be real, but the headroom is not comfortable. The industry is building assets that must be depreciated, powered, cooled, financed and kept busy. If the revenue base keeps compounding, the economics improve. If utilisation disappoints, or pricing falls faster than usage expands, they get much harder.

This is why I assess the next phase of AI through what I call the depreciation test: can falling AI prices unlock enough new demand to keep infrastructure utilisation and revenue growth ahead of the rising depreciation base? That single question matters more than most model benchmarks, because it decides whether the build-out becomes a sustainable economic platform or a capital cycle that overshoots demand.

Falling prices are the business model

At first glance, falling token prices look like a problem for the industry: cheaper inference, less revenue per unit. In technology markets, though, price declines routinely expand the market by making new use cases viable. Everything turns on elasticity — whether cheaper AI creates enough additional demand to more than offset lower unit prices.

Lower inference costs could make AI viable in workflows where the economics do not yet work: customer operations, software engineering, compliance review, research, sales support and knowledge work more broadly. The report’s central open question is whether cheapening intelligence can generate enough volume, and enough margin, to service the build-out.

Enterprise adoption is where this gets tested. Consumers may prove that people value the tools; enterprises decide whether the demand is durable. They will not keep funding AI at scale because it is impressive — they will fund it if it improves revenue, reduces cost, manages risk, accelerates delivery or opens up new operating models. Most are still early, experimenting with copilots and productivity tools rather than redesigning end-to-end workflows. The upper stack — applications, agents, domain-specific systems, AI-native operating models — is the demand-generation engine the lower stack is waiting for.

The physical economy now matters to the digital economy

AI also differs from earlier software waves in that its constraints are increasingly physical. Compute demand turns into data centre demand, then into power demand, and from there into questions of grid capacity, generation, transmission and permitting — which is where the AI economy connects directly with energy policy, industrial strategy and infrastructure planning.

The International Energy Agency projects that global electricity consumption from data centres could roughly double to around 945 TWh by 2030, growing far faster than overall electricity demand. (IEA) Goldman Sachs Research estimates that US data centre power demand could rise from 31 GW in 2025 to 66 GW by 2027, on rapid expansion of capacity and utilisation. (Goldman Sachs)

AI, in other words, is becoming an infrastructure and energy system rather than just a digital product category — with dependencies many business cases still underplay: grid capacity, energy availability, cooling, water, land, construction capability, skilled engineers, transformers, substations, permitting. Software metrics alone cannot capture this. The economics of AI sit at the intersection of software, infrastructure, energy and capital markets.

The AI Stack Economic Maturity Path

My first working framework for this series is the AI Stack Economic Maturity Path. The idea is simple: in major technology shifts, investment tends to start lower in the stack, but durable economic value frequently migrates upward over time.

Phase 1: Infrastructure scarcity. Value accrues to those who control the scarce resource — chips, compute, hosting, data centres and power.

Phase 2: Model differentiation. Model providers capture value while frontier capability remains scarce and meaningfully differentiated.

Phase 3: Application formation. Products that package intelligence into useful workflows begin to take share.

Phase 4: Enterprise transformation. The largest gains emerge when AI changes how organisations operate, make decisions, serve customers and allocate labour.

Phase 5: Industry reinvention. Finally, AI stops being a tool and becomes a structural change in cost curves, competitive advantage and business models.

The risk today is that too much of the market is valuing Phase 1 as though it permanently owns the economics of Phases 3 to 5; the opportunity is that many of the most valuable AI businesses may not have been built yet. That is why I see the current AI economy as more real than the sceptics suggest and less mature than the optimists imply.

What executives should take from this

For enterprise leaders, the implication is that the AI discussion has to move beyond technology selection. Choosing a model is not a strategy, and running pilots is not value creation.

The real executive questions are economic. Where does AI change the cost structure of the business? Which workflows can be redesigned rather than merely assisted, and what is the full operating cost beyond tokens? What happens to the business case when model prices fall, usage rises or human review stays in the loop? Which capabilities commoditise, and which create durable advantage?

The organisations that answer these well will not necessarily be the ones running the most experiments. They will be the ones with the clearest grasp of AI unit economics, value capture and operating-model change.

The hard questions begin now

AI has arrived as an economy. The revenue is too large to ignore, the capital investment too significant to dismiss, the infrastructure dependencies too material to treat as background detail. The hardest questions are still ahead, though: whether prices can fall fast enough to unlock mass adoption, whether usage can outrun the depreciation burden, whether applications and agents can create enough business value to justify the infrastructure beneath them, and whether the industry can manage the physical constraints — power, water, engineering capacity, environmental impact — without undermining the economics it is trying to create.

Those are economics questions, not technology questions. The next phase of AI leadership will belong to the people who understand both — but it will be decided by the economics.

The technology has done its part. What happens next comes down to the economics.

Sources

Exponential View — The State of the AI Economy (June 2026); accompanying essay

McKinsey & Company — The cost of compute: A $7 trillion race to scale data centers

Goldman Sachs — Private markets expected to have growing role in data center financing

Goldman Sachs — US data center power demand projected to double by 2027

IEA — Energy and AI: Energy demand from AI