Why the AI capex cycle is built to persist

The rapid acceleration of artificial intelligence (AI) has pushed capital allocation into uncharted territory. Annual investment in AI infrastructure by hyperscalers is expected to exceed USD 600 billion¹ in 2027. For investors, the natural question is whether this surge represents rational, value‑added investment or the early stages of overcapacity and potential capital destruction.

“...our analysis suggests that this cycle can continue for another two to three years before facing its first true test.”

We believe the AI capex cycle is poorly framed through a traditional boom/bust lens. Instead, it resembles a structurally enforced equilibrium, shaped by competitive dynamics, physical supply constraints, and the persistent economics of scale. Understanding these forces is essential for identifying where long‑term opportunities may emerge as the cycle evolves. Drawing on Nash equilibrium² dynamics, kinked demand and inelastic supply, internal capacity trade‑offs, durable scaling laws, and the gap between value creation and capture, our analysis suggests that this cycle can continue for another two to three years before facing its first true test. The greater risk lies in monetization lagging investment, creating volatility in sentiment despite the persistence of the strategic rationale.

A competitive equilibrium, not a spending frenzy

Hyperscalers face a binary choice: invest aggressively in AI infrastructure or risk falling behind. Crucially, aggressive investment remains the dominant strategy regardless of competitor behavior. Pulling back while rivals accelerate risks ceding monopoly‑like economics, forfeiting strategic control over the platform’s trajectory, and missing the next platform shift—costs that compound materially in tail scenarios where capabilities scale toward more agentic systems. This dynamic creates a Nash equilibrium, in which sustained capital expenditure is rational even if near‑term project‑level returns compress.

Importantly, the large hyperscalers possess the financial capacity to sustain this equilibrium. Hyperscalers are cash‑generative businesses on a scale we rarely see, with net cash balance sheets and, in most cases, high margins in their core franchises. In this context, capital discipline is less about absolute spending levels and more about maintaining strategic relevance in a rapidly scaling ecosystem. The implication for investors is clear: Persistence of capital expenditures is the base case, so the key is to find who can translate spend into defensible economics.

Why supply expansion has not collapsed prices

AI compute pricing has been more resilient than many expected despite massive capacity expansion. A kinked demand structure³ helps explain why (Figure 1). Above a threshold price (in this example, USD 35), demand is relatively inelastic because enterprise and strategic users treat AI as mission critical regardless of cost. Below that threshold, latent demand emerges rapidly, as experimentation, new use cases, and internal model training workloads become more economically feasible.

On the supply side, constraints remain acute. Power infrastructure lead times currently stretch three to five years. Advanced semiconductor fabrication requires multibillion‑dollar investments and extended construction timelines. Meanwhile, memory, networking, land availability, and permitting further restrict how quickly capacity can scale. In this setting, incremental supply often raises volume more than it lowers price, a modern Jevons Paradox.⁴ The kinked demand curve explains why the Jevons Paradox operates so powerfully in AI compute. Once price falls below the threshold (USD 35), the nearly horizontal demand curve (D marked in blue) means any supply increase is immediately absorbed by latent demand. Consumption expands to meet capacity with minimal price impact. For investors, this dynamic suggests that fears of an immediate price collapse may be overstated. The binding constraint is not demand, but the physical and logistical limits of expanding supply.

Capacity allocation: The internal trade‑off

Scarcity introduces a second‑order decision: how to allocate limited compute across competing priorities. Broadly, capacity is divided among three uses: revenue‑generating services, internal product enhancement, and long‑term research and development (R&D). The mix changes by phase, but R&D remains a competitive floor rather than a discretionary choice. The trade‑off is brutal because there is little slack. Every graphics processing unit (GPU) hour allocated to external revenue generation defers capability development, while every GPU hour devoted to model training defers near‑term revenue. Leaning toward R&D can deepen the long‑term competitive moat at the expense of near‑term margins. Those who choose to prioritize revenue may improve short‑term economics, but risk ceding long‑term leadership if R&D investment falls below the competitive floor.

This “Red Queen” dynamic,⁵ running simply to stay in place, helps explain why capital discipline is difficult to enforce across the industry. From an investment perspective, it also highlights why differentiation among winners and losers will increasingly hinge on execution, efficiency, and regulatory positioning rather than on headline spending levels alone.

Scaling laws and the case for persistence

The economic rationale for AI capital expenditure ultimately rests on scaling laws. Empirical evidence from leading research institutions shows that model capabilities have improved in a predictable manner as compute has increased. As long as incremental compute continues to deliver materially better models, the incentive to invest remains intact. Historically, each order‑of‑magnitude increase in training compute has produced meaningful gains in capability.

Three broad scenarios lie ahead:

• If scaling laws continue to hold, the current investment cycle could extend several more years, with expanding enterprise value justifying infrastructure costs.
• If returns diminish sharply, whether from data constraints or algorithmic efficiency gains that reduce compute needs, rationalization could arrive sooner and some infrastructure could be underutilized.
• A third possibility, a change in basic assumptions in architectures, may render some assets less economically viable, while preserving advantages for early investors with scale and ecosystem depth.

Monitoring these pathways will be crucial as durability of scaling relationships is likely to be the single most important determinant of long‑term value creation in the AI ecosystem.

Hidden economics of AI adoption

A defining feature of early technology markets is the gap between value created and value captured. AI may be creating enormous economic benefits that currently flow to users rather than producers, suggesting that the true market opportunity for producers exceeds current and estimated revenue. For example, enterprise pilots currently report 20%–40% efficiency gains in coding, content creation, and analysis tasks, which far exceed current subscription costs, while willingness‑to‑pay surveys indicate substantial latent pricing power.

These benefits are not accidental. Competitive pricing accelerates adoption, helps build ecosystems, and increases switching costs. Over time, however, as markets mature and differentiation deepens, a greater share of these benefits will likely accrue to producers. For investors, this implies that current revenue metrics may materially understate the true economic opportunity embedded in AI platforms.

Investing for the next phase

The AI capital expenditure cycle is neither irrational exuberance nor a guaranteed path to economic success. It is a competitive equilibrium sustained by supply constraints and the belief that scaling drives value creation. In our view, the market remains in the scale race phase, where proof points on monetization are rising but the capability race still sets the spending floor. In such an environment, active investment discipline will matter, and we are focused on the important signposts, such as changes in capital access, evidence that supply constraints are easing, shifts in capacity allocation toward harvest mode, validation or weakening of scaling returns, and early proof that producers are improving surplus capture.

“As with past periods of profound technological change, uncertainty is elevated, but so too is the opportunity set for those investors willing to look beyond near-term volatility and focus on the long-term fundamentals.”

As with past periods of profound technological change, uncertainty is elevated, but so too is the opportunity set for those investors willing to look beyond near‑term volatility and focus on the long‑term fundamentals. Discipline is less about calling the top in the capex cycle and more about underwriting durability, efficiency, and credible pathways from infrastructure advantage to monetizable outcomes.