How to read an AI capex disclosure

Capital expenditure disclosures for AI infrastructure have become one of the most closely watched numbers in technology reporting and one of the most frequently misread. Four questions cover most of the interpretive difficulty.

1. What is in the number?

Most disclosures aggregate AI-specific infrastructure with general datacentre, network and office capital. Segment-level breakouts are the exception. A rising figure may reflect AI deployment, or a datacentre refresh cycle, or a property purchase. Where the breakout does not exist, the number cannot answer the question most readers are asking of it.

2. What is off balance sheet?

Leased facilities, colocation with embedded capacity commitments and special-purpose vehicles do not appear as capital expenditure. Capacity held this way has grown substantially. A firm can materially increase deployment while reported capex is flat.

3. What useful life is assumed?

Accelerator depreciation schedules have lengthened across the sector. This is disclosed, usually in the accounting policy notes rather than in the headline. A change from three years to six halves annual depreciation on the affected assets, which flows directly to reported operating income with no change in cash flow.

4. Is it committed or spent?

Announced investment programmes are frequently multi-year commitments reported in aggregate. Cash actually deployed in a given period is a different figure and is often several times smaller. Both circulate under the same headline framing.

None of these are irregularities. They are ordinary accounting, disclosed in the ordinary places. The problem is that the summarised versions that reach general readers rarely carry the qualifications, and the qualifications are where the meaning is.