Eight year old graphics chips at new prices, the depreciation question

The short version

• The provider CoreWeave rents out eight year old graphics chips from Nvidia until 2029 at prices that match new equipment.
• Microsoft depreciates its AI data centers over 25 years instead of over 15 years.
• One AI company reported a possible stock market listing in October at a valuation of two trillion dollars.

A rental contract with signal effect

One report from the data center business received more attention this week than many a quarterly figure. The provider CoreWeave rents out eight year old graphics chips from Nvidia until 2029 at prices that match new equipment.

That confirms a decision Microsoft had taken before. The group depreciates its AI data centers over 25 years instead of over 15 years.

Both cases concern the same question from two directions. One is an assumption in the accounts, the other a price paid in the market. That they point in the same direction is the reason for the attention.

Why the useful life moves profits

What sounds technical is considerable in accounting terms. It is not about the price of the equipment, because that has already been paid, but about how many years this price burdens the profit account.

Exactly this has been fuelling a debate for months. Critics accuse the operators of setting the useful life too generously, because accelerators age quickly.

The objection carries weight, because the assumption concerns more than one company. If the useful life is set too long across a whole industry, reported profits appear higher without anyone having taken in more money anywhere. The test for this is the question of what a device actually still earns after several years.

The counter argument and its catch

The rental case now supplies a counter argument from practice. If eight year old equipment can still be rented out at new prices, it is evidently not worthless.

The argument is stronger than an assumption in the accounts, because it is not an estimate. A rental contract until 2029 names a price somebody is willing to pay, and for equipment that on the stricter reading should long since have been written off.

The objection to that is obvious. A rental price in a phase of extreme scarcity says little about what could be achieved in a relaxed market. The contract runs until 2029, and the scarcity that made it possible does not have to last that long.

Alongside this, one AI company reported a possible stock market listing in October at a valuation of two trillion dollars. The Dax and the Nasdaq marked new all time highs. Both describe the same mood in which the rental contract also came about.

Assessment

This report is more important than its length suggests. The useful life decides how high the reported profits of the whole industry turn out.

What makes me skeptical about it is the line of proof. A price in a phase of scarcity is no evidence of lasting value over ten years. It only shows that right now everybody wants to compute.

The honest test comes when new hardware is plentiful again. Then it will show what old equipment is actually worth.

Frequently asked questions

What changes when a data center is depreciated over 25 years instead of over 15

Depreciation spreads the acquisition cost over the expected useful life. At 15 years a device burdens the profit account each year with one fifteenth of the purchase price, at 25 years only with one twenty fifth. The reported profit rises as a result, and nothing changes in the cash outflow.

Why can eight year old chips still be rented out at new prices

Because with computing power it is not only the newest technology that counts. Older chips are slower but they work and they are available. If new hardware is scarce, their value rises, because customers would rather compute now than faster later.

Does the rental price prove that the long useful life is set correctly

Not necessarily. A rental price in a phase of extreme scarcity says little about what could be achieved in a relaxed market. The robust test only comes when new hardware is plentiful again. Then it will show what old equipment is actually worth, and with it whether a useful life of 25 years was set too generously.

This text is not investment advice. It reports verifiable figures and puts them in context.

This analysis is for information only and is not investment advice.