CrowdMood Column
How to read an AI capex announcement: spending, depreciation, and the power bill
Each earnings season the large cloud companies report how much they spent on data centers and chips, and the number keeps rising. Headlines read a bigger number as proof that AI demand is strong. It may be. But capital spending is cash going out. Whether it creates value depends on how fast revenue comes back to cover it. This guide explains how to read the announcement the way this desk does.
Capex is cash now, depreciation is cost later
Capital expenditure appears in the cash flow statement when the money is spent. It reaches the income statement gradually, as depreciation over the asset’s assumed useful life. A company that spends heavily this year can still report strong operating margins this year, because most of the cost has not yet been expensed. The bill arrives over the following years. Comparing capex growth with operating income growth today understates what margins will face later.
Useful-life assumptions move reported profit
The depreciation charge depends on how many years the company assumes its servers and network gear will last. Several large cloud companies have changed those assumptions in recent years, some lengthening and at least one shortening them for certain equipment. Lengthening the assumed life lowers annual depreciation and lifts reported profit without changing a dollar of cash. Shortening it does the reverse. When a company changes its useful-life policy, read the footnote: it tells you how management thinks the hardware will age, and AI accelerators may age faster than older servers.
Free cash flow is the honest number
Operating cash flow minus capex is free cash flow. For the largest companies, strong advertising, software, and cloud businesses have funded the build so far. The question to track is whether free cash flow keeps pace as capex rises. If capex grows faster than operating cash flow for several years, the build is being funded by shrinking free cash flow, by debt, or by off-balance-sheet commitments.
Leases and commitments
Not all data center capacity shows up as capex. Companies also sign long-term leases and capacity agreements with third-party developers and specialized cloud providers. Those obligations appear in lease disclosures and commitments notes rather than in the headline capex figure. A company can keep reported capex steady while its future obligations grow. Read the commitments section next to the capex number.
Power is the physical constraint
A data hall needs electricity before it earns anything. Developers sign long-term power purchase agreements, and new grid connections can wait years in interconnection queues. That is why this desk reads AI through watts as well as chips. Power demand from data centers can keep electricity and gas bid even if the equity story weakens. Empty racks on a live grid still pay for capacity.
Circular demand
Some AI demand is circular: a chip supplier invests in or finances a customer, and the customer uses that money to buy the supplier’s chips or rent capacity built with them. Each step is reported as revenue somewhere, but the end user paying for the service may be much smaller than the chain of orders suggests. When a large order comes with an investment or financing arrangement attached, separate the two before reading it as end demand.
Where Korea sits in the chain
Korean memory makers supply the high-bandwidth memory that sits next to AI accelerators, so hyperscaler capex shows up quickly in their orders and earnings. That makes the pace of AI spending a direct input to Korea’s export cycle. The same checks apply in reverse: if cloud companies slow capex because revenue is not covering depreciation, memory orders are among the first places it appears.
A checklist for the next announcement
Compare capex growth with the growth of revenue that is directly tied to AI or cloud services. Check whether depreciation is growing faster than operating income. Read any change in useful-life assumptions. Check free cash flow, lease commitments, and new debt. Look for signs of utilization, such as statements about capacity constraints or idle capacity. This desk’s view is that AI infrastructure is being overbuilt relative to final demand. These checks are how that view is tested quarter by quarter.
Market context
These series provide context. Co-movement does not establish causation.