The Depreciation Clock

Consider an accounting choice that decides the income statement. When a company buys a server, it does not record the whole cost as an expense in the year of purchase. It spreads the cost over the equipment's estimated useful life, a process called depreciation. The longer the assumed life, the smaller the annual charge and the higher the reported profit. For a business whose largest asset is a fleet of chips, that single estimate carries more weight than almost any other number in the accounts.

CoreWeave's S-1 registration statement disclosed that it depreciates its technology equipment, including graphics processing units, or GPUs, on a straight-line basis over six years, and that it extended that estimate from five years to six at the start of 2023. Six years is in line with the lives the largest cloud providers used for general servers at the time. The question is whether an accelerator bought for training frontier models behaves like a general server.

Two public data points frame the debate. First, Amazon disclosed in its annual report that, effective January 1, 2025, it shortened the useful life of a subset of its servers and networking equipment from six years to five, citing the faster pace of technology development, particularly in artificial intelligence and machine learning. Amazon had lengthened server lives in earlier years; this was a reversal, made by an operator with more experience of running servers at scale than anyone. Second, Nvidia said at Computex in June 2024 that it had moved to a one-year rhythm for its data center products. Hopper was followed by Blackwell, and Nvidia has published a roadmap to Rubin and beyond. A chip bought today will face several newer generations within its accounting life.

Why does a newer chip hurt an older one? Because customers pay for useful work, and a newer generation typically does more work per dollar of power and floor space. When the new chip arrives, the market rate for renting the old one falls. The old chip still functions, and it may find profitable use for years running smaller models or serving answers to users rather than training. But its earning power declines faster than a straight line would suggest.

Here is the mechanism that matters for investors. If a chip's real economic life is shorter than its accounting life, reported profits are overstated in the early years and the shortfall appears later, either as a write-down or as a stream of revenue that no longer covers the remaining depreciation. The cash does not change. The timing of reported profit does.

Now connect depreciation to debt. The loans described in the previous lesson are generally sized to amortise over the term of the customer contract, which is often shorter than six years. That is a sensible design: the lender is repaid while the chip is still under contract. It means the first contract carries the financing. Whatever the chip earns after the first contract ends is the equity holder's upside, and that upside depends entirely on what an older generation can command in the market.

So the right question is not simply whether six years is correct. It is what happens in the second contract. A chip that finishes a four-year commitment and can be re-leased at a reasonable rate for two more years supports the six-year assumption. A chip that finishes its commitment and can only be rented cheaply in a crowded spot market does not. The first contract is visible in the backlog. The second contract is a forecast.

Operators argue, with some justification, that older chips have stayed in demand longer than sceptics expected, because the amount of inference work has grown so quickly. Sceptics argue that one strong cycle is not a long track record. Both positions are reasonable, and the evidence will arrive year by year as early fleets roll off their first contracts. That is the data to watch: re-contracting rates on older hardware, disclosed in filings and earnings calls.

Faraday's read is that depreciation is where a neocloud's optimism is recorded. Compare each operator's useful-life assumption with its contract terms and with the pace of Nvidia's roadmap, and treat the difference as a measure of how much the reported profit depends on the future being kind.