GPU Depreciation

GPU depreciation concerns the allocation of the cost of graphics processing units over their useful lives and the loss of their economic value. Accounting depreciation is not a running record of secondhand prices, and a functioning GPU can lose economic value before it physically fails.^[1][2][8]

For AI infrastructure, depreciation assumptions affect reported expenses and the carrying value of equipment. Rental revenue and modeled residual values provide other evidence about the equipment's economics, but neither is interchangeable with an accounting schedule.^[3][6]

What the measurements mean

The U.S. Bureau of Economic Analysis distinguishes historical-cost depreciation commonly used in business accounts from current-cost consumption of fixed capital in national accounts. Those national-accounting estimates also should not be treated as a price quotation for an individual GPU.^[1]

Accounting estimates and useful life

Under Australia's AASB 116, which incorporates IAS 16, depreciable amount is cost or its substituted amount less residual value. Useful life reflects expected utility to the entity and can be shorter than economic life. The standard requires annual review of residual value and useful life. Straight-line depreciation produces a constant charge if residual value is unchanged; diminishing-balance and production-based methods have different patterns.^[2]

AASB 116 also distinguishes current fair value from estimated disposal proceeds for an asset already at its expected end-of-use condition. Depreciation can continue when fair value exceeds carrying value, provided residual value remains below carrying value. An increase in today's price therefore does not, by itself, eliminate depreciation.^[2]

U.S. company filings provide examples under their own accounting policies, not applications of AASB 116:

These disclosures cover equipment categories, not a universal lifespan for every GPU model. Amazon linked its shortening to faster technological development, particularly AI and machine learning. Its 2025 filing reported $1.4 billion of additional depreciation and amortization from the useful-life reduction.^[4]

Impairment is a separate issue. CoreWeave states that it tests long-lived assets when circumstances indicate carrying values may be unrecoverable. Its policy compares an asset's or asset group's carrying value with expected undiscounted cash flows; if that test fails, the impairment charge measures the excess over fair value. It also warns that useful-life estimates and attempts to redeploy older equipment may prove unsuccessful.^[3]

The September 2026 residual-value comparison

NVIDIA's September 2026 investor presentation compared residual-value estimates with an accelerated depreciation curve. Its chart used eight-GPU HGX system costs expressed per GPU and showed these endpoint labels:^[5]

The slide attributed monthly residual-value averages to Silicon Data. Purchase-price bases were third-party estimates, explicitly not Silicon Data data, and mass-availability dates were assumed. The comparison curve reached 25% at year two and zero at year five. It was a comparison schedule, not a disclosure that every operator used that policy. The percentages do not establish realized sale proceeds or an owner's investment return.^[5]

What the underlying model estimates

Silicon Data describes its residual-value product as a discounted-cash-flow model. It uses rental forward rates for the first 36 months, followed by an assumed decline beyond that horizon. Utilization decay, interest rates, and an operating-cost ratio affect the estimate. Operating costs use a blended global ratio rather than the costs of a specific site.^[6]

The methodology's maximum physical-life parameter was approximately eight years and configurable. Silicon Data distinguishes this input from economic life: declining utilization and rental rates can push modeled value toward zero earlier. A benchmark estimate can therefore change with market inputs even when the physical machine is unchanged. The figure is conditional on the model and its assumptions, not a guaranteed liquidation price.^[6]

Productive life and obsolescence

In an October 1, 2026 article, NVIDIA argued that software improvements and a range of compatible workloads can keep older systems commercially useful. It cited continued A100 service and bookings extending into 2029. This is the supplier's argument about durability, rather than an independently established useful life for every installation. Its framework also separates earning capacity from demand: hardware capable of producing output does not necessarily sell all of that output.^[7]

Economic research distinguishes remaining productive capacity from remaining economic value. Karl Whelan's 2000 study of computer capital modeled technological obsolescence as retirement of equipment that could still produce output. It separated the value of the capital stock from its productive capacity. This work provides a conceptual explanation for retirement before physical failure; it does not estimate the lifespan of contemporary AI GPUs.^[8]

Depreciation in AI training-cost research

Depreciation assumptions also affect estimates of the cost of training a model. In The rising costs of training frontier AI models, Ben Cottier and colleagues allocated hardware expense according to estimated loss of value during a training run, rather than assigning the full hardware purchase cost to one model.^[9]

Their preferred calculation assumed depreciation of 0.14 orders of magnitude per year, based on historical improvement in GPU price-performance at 32-bit precision. The method approximated continuous progress even though releases occur in steps, and omitted hardware failure from the depreciation calculation. The authors tested sensitivity to the assumed depreciation rate. This is a research assumption for allocating training costs, not a prescribed financial-reporting method or a forecast of a particular GPU's resale price.^[9]

See also

• Cloud AI GPU Pricing Comparison
• Neocloud
• CME Compute Futures

References

1. ^1 ^2 ^3U.S. Bureau of Economic Analysis. "Depreciation." Glossary, updated April 26, 2018.
2. ^1 ^2 ^3 ^4Australian Accounting Standards Board. *AASB 116: Property, Plant and Equipment*. December 2022 compilation, paragraphs 6, 50-57, and 62.
3. ^1 ^2 ^3 ^4CoreWeave. 2025 Form 10-K. Filed March 2, 2026. Risk factors and Note 2.
4. ^1 ^2Amazon.com. 2025 Form 10-K. Sections on depreciation and amortization and property and equipment.
5. ^1 ^2NVIDIA. *Non-Deal Roadshow*. September 2026, page 22, residual-value comparison and footnotes.
6. ^1 ^2 ^3 ^4Silicon Data. "GPU Residual Value." Product methodology and FAQ, accessed October 2, 2026.
7. ^Shruti Koparkar. "Productive, Durable, Fungible: How NVIDIA AI Factories Maximize Return on Investment." NVIDIA Blog, October 1, 2026.
8. ^1 ^2Karl Whelan. *Computers, Obsolescence, and Productivity*. Finance and Economics Discussion Series, March 2000.
9. ^1 ^2Ben Cottier et al. *The rising costs of training frontier AI models*. arXiv:2405.21015, version 2, February 7, 2025, sections 2.2 and B.3.