Backgrounder: AI Is Turning Cloud Growth Into a Capital-Allocation Test

Fast demand can hide inefficient allocation. A provider may invest ahead of orders and carry underused equipment, or invest too cautiously and turn customers away. Depreciation continues either way, while electricity, networking, maintenance, and staffing add to the cost of serving each workload.

Customers face a related risk. Committing to large reserved capacity can improve access and pricing, but it can also lock a company into an architecture before its workload is stable. The technical and financial plans therefore need to be evaluated together.

Microsoft's 2025 annual report says additions to property and equipment increased substantially as the company scaled cloud and AI infrastructure. It also describes pressure on cloud gross margin from that expansion even as Azure and other cloud services grew [1].

Alphabet's 2025 filing says it expects to increase investment in technical infrastructure and expects costs including depreciation, energy, equipment, and network expenses to rise [2]. Amazon's 2025 annual report similarly describes infrastructure spending for technology services, including AI and machine learning, and expects technology and infrastructure expense to increase [3].

These filings do not prove that the spending will earn inadequate returns. They do establish the operating equation: demand, build timing, utilization, pricing, energy, and depreciation all have to work together.

Operators buying AI capacity should separate experimentation from steady production. Short-term or elastic capacity is useful while demand is uncertain. Longer commitments become sensible after the workload's latency, reliability, security, and volume are measured.

Investors and managers should watch utilization proxies, cloud backlog, remaining performance obligations, depreciation growth, energy commitments, and commentary about constrained capacity. Revenue growth without those companion measures provides an incomplete view of the return on infrastructure.

The next phase of AI competition will test capital discipline. Providers must decide how much capacity to build, where to place it, which chips and networking systems to standardize, and how to price a service whose hardware can become obsolete quickly.

Customers can benefit from the buildout without copying its risk. The strongest approach is workload portability, clear unit economics, and staged commitments. The bigger signal is not merely that AI spending is large; it is that infrastructure operations are becoming a core part of software strategy.