Hardware Depreciation

When a company buys a very expensive machine, that amount does not show up as the cost of a single day. After all, the machine will be used for years. That’s why companies spread the purchase price over its estimated useful life. At a price of 6 million euros and six years of use, that comes to 1 million euros in costs each year. This exact allocation is called amortization or depreciation. Hardware depreciation concerns the computers, chips, cooling systems, and networking equipment in data centers — that is, in the large halls where AI services run.

Why the useful life of AI chips became a point of contention

The estimated useful life is an assumption, not a measured fact. And this assumption dramatically changes the reported profit. If an operator calculates with six years instead of three, the annual costs for the same purchase are cut in half. On paper, the business then looks significantly more profitable, even though not a single cent more has been earned.

With AI, this is especially contentious because the sums involved are enormous. Large technology companies invest double-digit billions annually in data centers. Several of them have extended their assumed useful life for servers in recent years. Critics consider this too optimistic, because new chip generations can quickly render old hardware economically worthless.

For investors, this is not merely an accounting question. If chips actually need to be replaced after three years, the real costs are much higher than reported. That’s why analysts now scrutinize closely which depreciation period a company applies in its reports.

How a purchase price becomes annual costs

The simplest method is straight-line depreciation. You divide the purchase price by the number of years of use and book the same amount as an expense each year. Some companies first subtract a residual value, meaning the price at which the hardware can still be sold at the end. For specialized AI chips, this residual value is difficult to estimate.

It’s important to distinguish between cash flow and accounting. The hardware is paid for immediately, often even before it goes into operation. But the costs only appear spread out over years in the profit and loss statement. As a result, a company can report a profit while simultaneously losing a great deal of money.

A common misconception: amortization does not mean that an investment automatically pays for itself. The term merely describes the allocation of costs. Whether the hardware generates enough revenue is an entirely separate question. If no one uses the purchased computing power, the depreciation continues regardless.

Where the depreciation period appears in quarterly reports

In the quarterly reports of major cloud providers, there is usually a sentence about the assumed useful life of the servers. Phrases like “extension of the useful life to six years” are common there. Such sentences regularly make headlines and move stock prices, because they affect reported profit.

The calculation also matters for providers that rent out computing power. They must set the hourly rental price so that it covers depreciation. If the market price for computing power falls, operators with expensively purchased hardware run into trouble.

By the way, this concept is not limited to AI. Airlines depreciate aircraft over decades, while shipping companies depreciate their trucks over just a few years. What’s new with AI is only the pace of technological progress. Anyone buying chips today cannot be certain whether they will still be competitive in four years.