The Capital-Intensity Paradox: When Software Economics Breaks

01 A Paradox With a Mechanism Inside It

The AI boom has produced a genuine analytical paradox, not a rhetorical one. Software was supposed to require little physical capital. The entire investment thesis of the software era — the reason software firms commanded premium valuations relative to their book assets — was that the marginal unit of software output costs nearly nothing to produce and distribute, that scaling required hiring rather than building, and that capital intensity was something other industries did. Frontier AI inverts this. It requires data centers, electricity, chips, cooling, and transmission — the full apparatus of heavy industry. The paradox is that the most advanced expression of information technology has become one of the most capital-intensive activities in the economy.

The software service model made the separation explicit: Software as a service is a cloud computing service model in which a provider delivers application software to clients while managing the required physical and software resources, separating the possession and ownership of software from its use, a model that began around 2000 and by 2023 was the main form of software application deployment (source: Wikipedia summary — Software as a service). SaaS outsourced the physical layer to someone else — a cloud provider — so the software firm's own balance sheet could stay light. The paradox now is not that AI firms need physical capital; it is that the physical layer itself became the scarce, strategic, competition-defining asset, which drags the economics of the whole stack back toward the capital-intensive model that software was supposed to have escaped.

The framing for this analysis: why software economics broke (training compute and inference scale); the capital cycle and depreciation now embedded in the AI stack; and the consequences for entry barriers — the opposite of software's low-capex promise. The analytical standard throughout: observed facts, reported claims, causal inferences, and scenarios are labeled, and no precise investment or capacity figures are asserted as sourced data.

02 Why Software Economics Broke

Software economics rested on a specific production function: development cost was the dominant cost, marginal delivery cost approached zero, and distribution was essentially free at scale. Two properties of frontier AI break it. First, training compute: building a competitive frontier model requires computation at a scale and cost that resembles industrial capital formation rather than product development — the model is, in effect, a manufactured artifact produced by an expensive physical process, not a written artifact produced by an expensive labor process. Second, inference scale: unlike traditional software, where the finished product costs nothing to serve, every user interaction with a large model consumes measurable computation, electricity, and silicon. The marginal cost of AI output is small per unit but strictly positive and cumulative at scale — and it scales with usage rather than flattening at maturity.

The SaaS reference point makes the break precise. In SaaS, the provider manages the required physical and software resources (source: Wikipedia summary — Software as a service), but those resources are general-purpose commodity compute whose cost trended down for two decades — outsourcing the physical layer was cheap because the physical layer was abundant. Frontier AI's requirements are not abundant: specialized accelerators, high-density power delivery, advanced cooling, and grid-scale electricity are all currently capacity-constrained. When the outsourced layer becomes the bottleneck, the economic logic of outsourcing weakens: firms that could rent commodity compute indefinitely now have a strategic reason to own the constrained layer, which is why the capital expenditure has migrated from rental operating expense into owned balance sheets. The software industry's light-asset equilibrium was conditional on the physical layer being cheap and abundant; the condition failed, and the equilibrium broke.

03 The New Capital Cycle and the Depreciation Problem

Once the physical layer is owned rather than rented, the AI industry inherits the classic problems of capital-intensive industries, and depreciation is the sharpest of them. Capital-intensive businesses live and die by the relationship between the asset's life, its utilization, and the revenue it must generate before replacement. Software-era assets — code and brand — depreciated slowly and mostly through obsolescence of relevance. Frontier AI's physical assets depreciate through three channels at once: physical wear (the data center, cooling plant, and power infrastructure age on ordinary industrial timelines), technological obsolescence (a compute generation is superseded on a much faster cycle — a machine that is competitive today may be inefficient relative to successors in a small number of years), and demand-side uncertainty (the utilization of the asset depends on the continued growth of AI usage, which is precisely what is in question in the boom-versus-bubble debate).

The combination is analytically distinctive. In traditional heavy industry, long physical lives support long financing structures; in AI infrastructure, the short technological life sits inside the long physical life, which means the economic depreciation of the asset is much faster than its accounting life, and the financing must be repaid on the technological clock rather than the physical one. This creates a structural pressure on the whole complex: revenue per unit of installed compute must rise, or the effective utilization must rise, or the write-offs must eventually be recognized. The capital cycle question — whether current investment is being amortized against real demand or against projected demand — is the single most consequential uncertainty in the current episode, and it is the same question every capital-intensive boom in economic history has faced.

04 Entry Barriers: The Inversion of Software's Promise

The second-order consequence is the inversion of the entry-barrier structure. Software's low-capex promise was a competitive promise: low capital requirements meant low entry barriers, which meant continuous disruption — the startups of each cycle could unseat incumbents because competing required developers, not factories. Frontier AI reverses the direction. When the competition-defining asset is a capital-intensive physical layer, the firms that can fund it — those with the deepest balance sheets and the cheapest cost of capital — gain a structural advantage over those that cannot. Entry barriers in the AI stack are now measured in capital formation capacity, power procurement, multi-year supply contracts, and the organizational capability to build and operate industrial facilities.

The transmission into industrial structure follows from standard industrial-organization logic: high fixed costs plus constrained capacity favor concentration, because scale spreads the fixed cost and long-term supply contracts lock in the constrained inputs. Where software-era markets tended toward fragmented, contestable structures, capital-intensive markets tend toward oligopolies. The consequence for the broader economy: the gains from AI, if the technology delivers, may accrue to a narrower set of firms than the software era conditioned observers to expect — not because of any single firm's conduct, but because the production function's shape has changed. This is a causal inference from the capital structure, labeled as such; whether concentration in fact materializes is an empirical question the indicators below will track.

05 Second- and Third-Order Effects Across the Stack

The paradox propagates. Upstream: demand for the physical layer's inputs — advanced chips, power equipment, cooling systems, grid capacity — transmits into those industries' own investment cycles, with their own long lead times, which means supply responses are slow precisely when demand signals are strongest. This is the classic recipe for cyclical overshoot in capital goods: capacity arrives in waves, years after the demand that justified it. Downstream: application-layer firms face a changed cost structure too — if inference is metered and positive at the margin, then AI-native products carry a usage-linked cost of goods that pure software products did not, raising the bar for unit economics in the application layer.

Third-order, and most systemically interesting: electricity. Data-center demand growth transmits into power markets, grid planning, and generation investment — decisions made by regulated utilities and grid operators on multi-decade horizons that are far slower than AI investment cycles. The mismatch between the tech industry's clock and the power system's clock is a genuine structural tension: the AI complex can decide to build in months what the grid needs years to deliver. How that mismatch resolves — through price, through siting constraints, through generation technology shifts — will shape both the AI industry's geography and the power system's own investment cycle. The companion analyses in this series treat the macroeconomic transmission in detail; the point here is the mechanism's direction: the software industry's economics have been dragged into the economics of power systems, and both sides' planning horizons now matter to each other.

06 Scenarios: How the Paradox Resolves

Scenario A — Demand validates the capital. AI usage grows into and past the installed base, utilization rises, and the depreciation clock is comfortably outrun by revenue growth. Trigger: sustained evidence that inference demand is absorbing installed capacity. Mechanism: positive marginal-cost economics become manageable at scale, and the capital cycle looks, in retrospect, like disciplined capacity ahead of demand. Indicators: rising utilization of installed compute; power-purchase and grid-interconnection queues converting into delivered capacity; application-layer revenue per unit of inference improving. Consequence: the paradox becomes a settled cost structure — a capital-intensive technology industry, concentrated but growing — and the entry barriers consolidate around the incumbents who built first.

Scenario B — Efficiency outruns the build. Algorithmic and hardware efficiency improve fast enough that the same capability requires less compute than the installed base provides, and utilization stagnates. Trigger: a credible efficiency discontinuity — the historical pattern in computing has been that capability per unit of physical resource improves persistently. Mechanism: the depreciation mismatch bites — assets are economically superseded before their financing matures. Indicators: declining effective utilization of installed capacity; a widening gap between announced capacity and contracted demand; early recognition of write-downs. Consequence: the paradox resolves through the classic capital-cycle bust, with overbuilt capacity absorbed at low prices — which, notably, would be the mechanism by which AI capability becomes cheap and widely available even as the builders bear the losses.

Scenario C — The physical layer stays binding. Neither demand disappointment nor efficiency discontinuity dominates; power, chips, and grid capacity remain the binding constraints on AI growth for the foreseeable period. Trigger: interconnection queues, power prices, and chip supply continuing to gate expansion. Mechanism: the AI industry's growth rate is set by the slowest-moving dependency — the power system's clock. Indicators: sustained power-price premiums in data-center-concentrated regions; multi-year waits in interconnection queues; data-center siting migrating toward power-rich geographies. Consequence: the paradox becomes permanent industrial structure — AI remains a heavy industry, and competitive advantage accrues to those who control energy and supply chains rather than those who write the best software.

07 Indicators to Watch

First, announced-versus-under-construction-verses-operational data-center capacity: the conversion rate between announcement and delivered capacity is the cleanest measure of whether the capital cycle is real or aspirational. Second, utilization of installed compute — the depreciation question's central observable. Third, power-purchase agreements and interconnection-queue statistics: the physical layer's most binding constraint made visible. Fourth, chip supply commitments and delivery lead times: the upstream capital-good cycle. Fifth, effective compute cost per unit of capability — the efficiency trend that Scenario B rides on. Sixth, balance-sheet composition shifts across the software industry: capital expenditure migrating into software-sector balance sheets is the paradox itself, measured. Seventh, market-structure observables — market-share stability and entry rates in the AI infrastructure layer — to test the concentration inference of section four. Eighth, accounting behavior: depreciation schedules and any write-down recognition in AI-linked infrastructure, which will register the mismatch's arrival with the lag that accounting always has.

08 The Bottom Line

What we know: Software's light-asset model was conditional on a cheap, abundant physical layer, per the SaaS model's own structure of separating software ownership from use while the provider manages physical resources (source: Wikipedia summary — Software as a service); frontier AI's requirements — data centers, electricity, chips, cooling, transmission — sit on the other side of that condition.

What we think we know: The production function has changed shape: training is capital formation and inference carries a positive, usage-scaling marginal cost, which reintroduces depreciation risk, concentration pressure, and the classic capital-cycle overshoot dynamics that software's economics had exempted the industry from. Entry barriers have inverted — from low-capex contestability to capital-formation capacity.

What we do not know: Whether installed capacity will be validated by demand (Scenario A) or economically superseded by efficiency (Scenario B); how long the power system's planning clock can gate the AI industry's investment clock; and whether the industry's accounting treatment of the depreciation mismatch will surface the problem early or late.

What to watch next: Capacity conversion rates, compute utilization, interconnection queues, effective compute cost, software-sector capital expenditure, and the first significant write-down cycle in AI-linked infrastructure — the moment at which the paradox's resolution becomes legible in reported numbers.