AI infrastructure has a timing problem. A company may spend cash on servers today, report strong demand, and only later feel the full accounting cost of those servers through depreciation. That does not make depreciation fake or non-cash hand-waving. It makes useful analysis more careful.

The short answer

Depreciation spreads the cost of a long-lived asset across the period in which it is expected to help generate revenue. For AI infrastructure, the risk is that hardware becomes less productive or less valuable before the planned useful life ends. Investors should compare utilization, revenue per accelerator, maintenance, power, and cash commitments instead of looking at capex alone.

Why GPUs create a timing mismatch

Buying a server uses cash immediately. Accounting usually records the server as an asset and recognizes depreciation over its estimated useful life. If that life is 4 years, a simplified $40,000 asset might create $10,000 of annual depreciation before considering residual value and tax treatment.

The example is only an explanation, not a forecast. Real companies group assets, choose estimates, and disclose policies in filings. The important point is that cash flow and operating income can move differently.

AI hardware adds another layer: a newer accelerator can change the economics of an older one. An older system may still work, but customers may pay less for its tokens, or the operator may need more electricity to deliver the same useful output.

What can go wrong

The danger is not depreciation by itself. The danger is low utilization combined with a large fixed asset base. If a facility is half-empty, the company still pays for power contracts, staff, financing, and equipment ownership or leases.

Another risk is an estimate change. If management decides hardware will be useful for a shorter period, depreciation expense can rise. If equipment is impaired, the accounting loss can be larger and faster. Neither event alone proves fraud, but both deserve an explanation.

I would build a simple monthly table with capacity available, capacity used, revenue, power cost, depreciation, and maintenance. Even 3 months of consistent measurements is more useful than a single utilization claim.

The investor checklist

Ask whether capex is owned, leased, or funded by a customer commitment. Then check the depreciation life and compare it with the company’s hardware-refresh cycle. A 6-year commercial commitment supported by equipment that may be replaced in 2 years needs a clear economic explanation.

Also separate gross bookings from recognized revenue. A multi-year agreement can sound huge while revenue is recognized over time. The SEC revenue-recognition guidance is a useful starting point for that distinction.

What this means for developers

Infrastructure accounting eventually reaches the API. Providers may retire a model, limit a region, change a price, or add a cheaper backend to improve utilization. Applications should record latency, error rates, model version, and provider in traces so an unexpected change can be explained.

For a production system, I would keep a fallback provider and a small benchmark suite. Run it for 2 hours with representative traffic before changing the default route. Cheap compute is not cheap if it increases retries or lowers answer quality.

Sources and limits

NVIDIA’s 2026 annual report discusses infrastructure expense, indirect customers, and concentration considerations. Its 2027 Q2 filing discusses commitments to AI cloud partners. These filings provide context, not a complete forecast for every operator. The accounting examples above are simplified to make the mechanism understandable.

Depreciation is best understood as a test of whether yesterday’s hardware can keep earning enough tomorrow. That is why utilization and refresh speed belong beside every AI capex headline.

Keep the conclusion tied to operating evidence: utilization, pricing, power efficiency, and refresh cadence. Those measures show whether the asset is still productive after the excitement of the purchase has passed.

The accounting line deserves the same attention as the hardware announcement. A fleet that generates strong gross demand but weak contribution after power, depreciation, maintenance, and financing may need a different refresh plan. Compare the trend over several quarters, because one quarter can include a delivery spike or a one-time impairment. The goal is not to reject investment in compute. It is to ask whether each new server expands useful output faster than it expands fixed cost.