The first phase of the artificial-intelligence infrastructure boom was easy to describe: the largest technology companies were racing to secure chips, data centres and power. The second phase is harder. Investors now have to decide whether the extraordinary amount of capital going into compute will produce revenue and operating profit quickly enough to justify the depreciation and financing burden that follows.
Amazon's 2025 filing put cash capital expenditure at $128.3 billion, up sharply from 2024, with technology infrastructure a central driver and the majority of that infrastructure tied to AWS growth. Meta spent $69.69 billion on property and equipment in 2025 and said it expected 2026 capital expenditure of roughly $115 billion to $135 billion. Microsoft reported $64.55 billion of additions to property and equipment in fiscal 2025, around $20 billion more than the prior year.
The accounting lag matters
Data centres do not hit the income statement in the same way as an operating expense paid today. Servers, buildings and network equipment are capitalised and depreciated over time. That means the cash leaves before the full expense burden reaches reported earnings.
For investors, this creates a lag between the buildout and the margin test. A company can show strong current operating profit while also committing itself to much higher future depreciation. The more aggressive the infrastructure cycle becomes, the more important it is to track both free cash flow and depreciation rather than relying on earnings per share alone.
Utilisation is the variable that separates productive capex from stranded capacity
The bullish case is straightforward. If AI demand continues to rise and the hyperscalers keep scarce accelerators and data-centre capacity highly utilised, the assets can support cloud growth, inference demand and new software products for years. Scale can also lower unit costs as infrastructure is shared across many customers and internal products.
The risk is that capacity arrives faster than monetisation. In that scenario, depreciation rises while pricing or utilisation disappoints. That does not require AI demand to collapse. It only requires the return on incremental capital to fall below what investors have embedded in valuations.
Our view: the next AI winner is the company that proves capital efficiency
Global Markets Review's view is that AI spending should now be analysed as an industrial investment cycle. The relevant questions are capacity, utilisation, pricing, depreciation and return on invested capital.
The companies with the deepest pockets won the first race for compute. The next stage is more discriminating. The market will increasingly reward the companies that can show that each new dollar of infrastructure spending produces durable cloud revenue, software revenue or measurable cost savings rather than simply larger depreciation schedules.