The Ai Capex Supercycle Meets Its First Real Earnings Test

The artificial intelligence capital expenditure supercycle is entering a critical phase: the point at which investors stop accepting "trust us, the returns are coming" and start demanding actual evidence that the hundreds of billions of dollars being committed to data centres, chips, and infrastructure are translating into revenue and margin expansion. The past twelve months have seen the four largest hyperscalers collectively announce capex plans that, if fully executed, would represent the largest coordinated private sector infrastructure buildout in history. The question is no longer whether AI spending is happening — it is whether it is working.

The earnings results from this cycle have begun to reveal a bifurcated picture. On one side are the infrastructure providers: chip manufacturers, data centre operators, networking equipment companies, and energy utilities that power the compute. For these companies, AI capex is unambiguously accretive — orders are strong, backlogs are growing, and pricing power is intact. On the other side are the hyperscalers themselves, who must convince investors that the margin compression from building and operating this infrastructure will be recovered through AI-enabled revenue streams that are, in many cases, still in their early innings.

The Unit Economics Problem

The central challenge for AI economics is that the cost of training and running large language models has been falling rapidly — a dynamic known as algorithmic efficiency improvement — while the revenue models for deploying those models remain nascent. Cloud AI services, enterprise copilots, and API-based developer tools are generating real revenue, but at scale levels that remain a fraction of the infrastructure investment required to produce them. Bears argue that the economics will never fully close; bulls argue that the revenue ramp is simply behind schedule and that the addressable market is large enough to justify the upfront investment.

The stock market's verdict on this debate has so far favoured the bulls. The valuation premiums commanded by AI-exposed equities remain elevated relative to historical norms, supported by the argument that investors are buying optionality on a genuinely transformative technology rather than paying for current earnings. The broader equity rally has provided a supportive backdrop that makes it easier for growth stocks to carry expensive multiples, but that support will look less durable if rates stay higher for longer than markets currently expect.

The Chip Supply Chain Is the Bottleneck That Matters

At the heart of the AI buildout is the GPU, and at the heart of GPU supply is a remarkably concentrated production ecosystem. A single company designs the chips that power the vast majority of AI training workloads, and a single foundry produces the most advanced nodes on which those chips are manufactured. This concentration creates both extraordinary leverage for the incumbents and extraordinary fragility for the industry. Any disruption to the semiconductor supply chain — whether from geopolitical risk, natural disaster, or manufacturing yield problems — has the potential to materially delay the AI deployment timelines that equity valuations are banking on.

The geopolitical dimension is particularly acute. Export restrictions on advanced chips to certain markets have already forced hyperscalers to develop differentiated infrastructure strategies for different geographies, increasing complexity and potentially reducing the network effects that make AI infrastructure economically attractive. Investors in this space need to track both the technology advancement curve and the regulatory environment with equal diligence.

Energy: The Constraint Nobody Wants to Talk About

Data centre electricity consumption is growing at a pace that is beginning to attract serious attention from grid operators, environmental regulators, and long-term investors who care about stranded asset risk. The largest AI data centres consume as much power as a medium-sized city, and the expansion plans being announced suggest the gap between data centre demand growth and renewable energy supply growth will widen before it narrows. Some hyperscalers are making significant investments in nuclear energy — both through power purchase agreements with existing plants and through direct investment in small modular reactor companies — but these solutions are years from meaningful scale.

The energy dimension connects the AI story to the commodities market in ways that are underappreciated in most equity-focused analysis. Natural gas prices, uranium spot contracts, and transmission infrastructure investment are all being influenced by data centre siting decisions. Our commodity desk's coverage of energy markets and supply dynamics provides context for how the traditional energy complex is adapting to this new and fast-growing demand category.

Practical Implications for Equity Investors

For investors trying to express a view on AI without concentrating in a handful of mega-cap names, the infrastructure layer offers compelling alternatives. Power management companies, cooling system manufacturers, fibre optic network operators, and specialised real estate investment trusts that own data centre campuses have all participated in AI-driven revenue growth with lower valuation risk than pure-play software companies. The diversification logic here is straightforward: somebody has to build and operate the physical layer, regardless of which AI application layer models ultimately win.

Tax treatment of AI-related investments — including R&D expense deductions, accelerated depreciation for qualifying infrastructure assets, and capital gains treatment on tech holdings — is worth reviewing before year-end. Our comprehensive tax planning checklist covers the specific line items that tech-heavy portfolios should address before the calendar turns.