The Hidden Costs of AI Data Centres: Power, Depreciation and Financing

A proposed AI data centre can sound impressive: new buildings, high-end processors and long-term customer demand. But the financial story starts before the first model runs. Developers must find a suitable site, arrange power and cooling, buy or lease computing equipment, connect networks and negotiate service contracts. Each step takes money and introduces execution risk. If electricity capacity is delayed, a costly campus can sit below its planned utilization while financing charges continue.

This is why 'more AI infrastructure spending' is not automatically the same as 'more AI infrastructure profit'. The buyer of a chip, the landlord of a facility and the cloud platform selling computing time can all face different returns. Some capture revenue immediately; others must wait for customers to use capacity over several years.

Computer hardware is expensive and does not remain cutting-edge forever. Accounting spreads many equipment purchases across their estimated useful lives through depreciation, but the cash was often spent much earlier. Free cash flow, operating cash flow and accounting profit can therefore move differently as a company expands. A business might report growing revenue and earnings while buying more hardware than its operations currently generate in cash. It might also have valuable contracted demand that is not yet fully reflected in present revenue.

Investors should look for capital expenditure commitments, lease obligations, financing arrangements, asset-life assumptions and disclosed risks of technological obsolescence. Long-lived buildings and short-lived accelerators should not be casually treated as the same kind of asset. Customer concentration matters too: a facility designed around one major buyer may face a difficult transition if that relationship changes.

A useful stress test assumes that the project opens six months late, electricity costs rise and customers use less capacity than planned. Can the operator still pay lenders and maintain equipment? What if a faster chip generation arrives before existing hardware is fully recovered? Not every project will face those outcomes, but the exercise reveals whether projected profits rely on perfect execution.

The cancelled Firmus AI infrastructure IPO discussed elsewhere in RecoupRev is a reminder that financing structures and valuation assumptions deserve scrutiny even when enthusiasm for artificial intelligence remains high. For public companies, use audited filings where available and separate signed contracts from aspirational announcements. A credible infrastructure story connects megawatts and hardware to utilization, margin and cash generation.