The Two Clocks: Depreciation and Obsolescence

Depreciation is where GPU finance either holds together or comes apart. In a spreadsheet, depreciation is a tidy line — spread the purchase price over a useful life and expense a slice each year. In the real world of accelerators, depreciation is a fight over one number: how long is the useful life? Get it wrong and every downstream figure — margins, collateral coverage, return on capital — is wrong with it.

Start with the distinction that trips people up. There is accounting depreciation, the schedule a company puts in its books, and there is economic depreciation, the actual decline in what the asset can earn. They are not the same. A firm can depreciate a chip over six years for accounting while the market repriced its rental rate down forty percent in eighteen months. The books say the asset is worth most of its cost; the rental market says otherwise. That gap is where losses hide.

What drives economic depreciation for accelerators? Two clocks, running together.

The obsolescence clock. Each new generation resets the frontier. When faster silicon ships, tenants doing frontier training migrate to it, because training cost scales with time-to-result. The prior generation does not stop working — it slides down-market to inference, fine-tuning, and price-sensitive workloads that tolerate slower hardware for a lower rate. Its rental price steps down at each transition. The schedule of those steps is set by the vendor's release cadence, which has compressed toward roughly annual. Faster cadence means faster demotion.

The wear clock. Accelerators run hot, near full utilization, in dense racks. Thermal cycling, power stress, and cooling wear degrade units and raise failure rates over time. High utilization is good for revenue and bad for lifespan. The very thing that makes a cluster profitable — keeping it pinned at capacity — shortens the physical life of the metal inside it.

The two clocks interact in a way operators would rather not advertise. To beat obsolescence, you run the chips hard while they are still frontier-grade, extracting return before the next generation lands. But running them hard accelerates wear. So the window where a chip is both fast enough to command a premium and healthy enough to run reliably is narrow — often shorter than the accounting life on the books.

Why does the accounting life get stretched anyway? Incentives. A longer assumed life means a smaller annual depreciation charge, which means higher reported margins and profit. It also means a higher carried book value, which supports the collateral coverage on the debt. Both the operator and its lenders have a reason to prefer the optimistic number. The auditor may sign off because there is no hard rule, only judgment. That is precisely why a stretched depreciation assumption is one of the loudest warning signs in the sector: it flatters earnings and collateral at the same time, and it does so quietly.

Residual value is the second landmine. Every model assumes the chip is worth something at the end — a resale price into the secondary market. But the secondary market for accelerators is young, thin, and reflexive. When demand is hot, used units hold value and the residual assumption looks conservative. When demand cools — a training pause, a cheaper new generation, a credit squeeze — everyone tries to sell used chips at once, into a market with few buyers. Residuals gap down exactly when they are needed to cover a loan. The residual assumption is only stress-tested in the scenario that breaks it.

Put the pieces together and you get the obsolescence risk that defines this asset class. The chip is a depreciating, wearing asset whose decline is scheduled by a third party — the chipmaker's roadmap — and whose salvage value is set by a fragile secondary market. Debt written against it assumes a decay curve. If the real curve is steeper than the assumed one, the collateral falls below the loan balance, and the cushion the lender counted on is gone.

For an analyst, three questions cut to the core. What useful life is assumed, and how does it compare to the vendor's release cadence? What residual value is baked in, and is there a real secondary market to support it? And does the debt amortize faster than the asset economically decays? Answer those and you know whether a GPU-finance structure is conservative or is simply borrowing against a future that the technology roadmap will not deliver.