GPU Impairment Accounting: How to Handle GPU Value Depreciation on Balance Sheets

GPU impairment accounting has become one of the most consequential financial reporting issues for AI companies as hardware depreciation curves steepen and market prices fluctuate. Under ASC 350-360 (US GAAP) and IAS 36 (IFRS), long-lived assets must be tested for impairment when events or changes in circumstances indicate that the carrying amount may not be recoverable. For GPU hardware, the typical impairment trigger events include a significant drop in market prices for identical GPUs, the announcement of a next-generation architecture that renders current GPUs less competitive for training workloads, or a material decline in the company's own GPU utilization rates.

The H100 price decline of 40-50 percent between Q2 2025 and Q2 2026 triggered impairment charges across AI companies and GPU providers that carried Hopper-generation hardware at cost on their balance sheets. Publicly disclosed impairment charges in the AI infrastructure sector totaled approximately $800 million in 2025 and an estimated $1.8 billion in the first half of 2026, according to SEC filings analysis. Companies that had not stress-tested their GPU carrying values against market price declines found themselves with outsize impairment charges that surprised investors and, in some cases, triggered debt covenant violations.

The distinction between depreciation and impairment is critical. Depreciation systematically allocates cost over useful life and is predictable. Impairment is a sudden write-down triggered by an external event that changes the asset's recoverable value in a way not captured by the depreciation schedule. A GPU being depreciated over 5 years with $0 salvage value carries a book value of 60 percent of cost after 2 years. If the same GPU can be purchased new for 50 percent of its original cost after 2 years, the asset is impaired and must be written down, even if the depreciation schedule appears reasonable.

Under US GAAP (ASC 360-10), impairment testing follows a two-step approach. Step 1 compares the undiscounted future cash flows expected from the GPU asset group to its carrying amount. If the carrying amount exceeds the undiscounted cash flows (i.e., the asset group cannot generate enough cash to cover its carrying value), Step 2 measures the impairment loss as the difference between carrying amount and fair value. The key nuance is the use of undiscounted cash flows, which creates a higher threshold for impairment recognition compared to IFRS. An AI company with GPUs used for revenue-generating inference can often pass Step 1 even if GPU market prices have dropped sharply, because the undiscounted revenue projections still exceed the carrying value.

IFRS (IAS 36) uses a single-step approach that compares carrying amount to the recoverable amount, defined as the higher of fair value less costs to sell and value in use (discounted cash flows). The discounting of future cash flows means IFRS impairment is more sensitive to GPU price declines and higher discount rates than US GAAP. An IFRS-reporting AI company with $20 million of GPUs on its books, generating $5 million of annual cash flow with a 15 percent discount rate, would recognize impairment when carrying value exceeds approximately $28 million in present-value terms, a lower threshold than the undiscounted comparison under US GAAP.

Practical difference: an AI startup with $10 million carrying value of H100 GPUs generating $1.5 million annual cash flow would likely pass Step 1 US GAAP impairment testing (undiscounted: $1.5M x 4 years remaining = $6M, which is below the $10M carrying value, so actually this would fail). Let me be precise: the undiscounted future cash flows must exceed carrying amount. If remaining useful life is 3 years and annual cash flow is $1.5M, undiscounted cash flows total $4.5M. Since $4.5M < $10M carrying value, impairment exists under both GAAP and IFRS. However, if the GPUs could be repurposed to higher-value workloads generating $3M annually, undiscounted cash flows of $9M might still fail Step 1. The lesson is that impairment testing requires management to assert the specific cash-generating use case for the GPUs, and this assertion must be supportable.

Fair value measurement under ASC 820 establishes a three-level hierarchy for GPU valuation. Level 1 inputs are quoted prices in active markets for identical GPUs, which exist for common configurations like H100 SXM 80GB PCIe, where secondary-market clearing prices are observable through public GPU marketplaces. Level 2 inputs include quoted prices for similar GPUs or observable market data such as NVIDIA's own pricing adjustments, lease rate implicit in GPU lease contracts, and broker quotes for bulk GPU transactions. Level 3 inputs are unobservable and rely on management estimates, including projected residual values, assumed useful lives, and discount rates applied to future lease income.

For impairment testing purposes, most AI companies should use Level 2 fair value measurements for their GPUs, combining observable secondary market data with adjustment factors for configuration differences, warranty status, and remaining economic life. A used H100 with 18 months of remaining life and an active transferable warranty trades at a 15-25 percent premium to the same GPU without warranty. Bulk GPU transactions (100+ units) typically trade at 5-10 percent discounts to single-unit prices. Companies that use Level 3 estimates without market corroboration risk auditor rejection and potential restatement if the assumptions are not supportable.

The frequency of fair value reassessment matters. GPU market prices can move 10-20 percent within a quarter based on supply announcements (e.g., TSMC capacity updates, NVIDIA architecture launches) or demand shifts (e.g., DeepSeek's R2 launch driving spot H100 prices up 30 percent in February 2026). Companies with material GPU balances (over $5 million or 10 percent of total assets) should perform quarterly fair value assessments even when no obvious impairment trigger exists. The cost of quarterly assessment is low relative to the risk of a large surprise impairment charge.

The choice of depreciation method for GPU assets significantly affects reported earnings and balance sheet carrying amounts. Straight-line depreciation over 4-5 years is the most common method in practice, adopted by approximately 70 percent of public AI companies according to 2025-2026 financial disclosures. Straight-line is simple, predictable, and matches the pattern of ratable revenue generation from inference workloads. However, it systematically overstates asset values on the balance sheet during the early years when market prices are dropping fastest, creating latent impairment risk.

Double-declining balance (DDB) depreciation over 3-4 years better matches the economic reality of GPU value decline. Under DDB with a 4-year life and $0 salvage, a $30,000 H100 would carry book values of $15,000 after Year 1, $7,500 after Year 2, and $3,750 after Year 3. This matches the observed market price trajectory more closely than straight-line, reducing impairment risk. The tradeoff is that DDB depresses reported earnings in early years. At a $10 million GPU deployment, the difference between straight-line and DDB depreciation in Year 1 is approximately $2.5 million lower net income under DDB.

Component depreciation is an emerging best practice for large GPU fleets. Under this approach, the GPU board, HBM memory, NVLink bridge, and cooling system are depreciated over different useful lives. The GPU core depreciates over 4 years, the HBM memory over 3 years (due to faster technology cycles in memory), and the cooling system over 6-8 years. Component depreciation increases accounting complexity but produces a more accurate carrying value and reduces the risk of large one-time impairment charges because each component's book value more closely tracks its economic value.

SEC registrants must disclose impairment charges in MD&A and financial statement footnotes, including the nature of the impairment event, the amount of the charge, the fair value measurement method, and the line item where the impairment is recorded. For GPU impairments specifically, the SEC has increased scrutiny on the assumed useful lives and residual values used in depreciation policies. Several AI companies were subject to SEC comment letters in 2025-2026 questioning why their 5-year useful life assumption for H100 GPUs was supportable given the 3-year technology cadence NVIDIA had publicly established.

Best practice disclosure includes: a sensitivity analysis showing how impairment would change under different market price assumptions, a reconciliation of carrying amounts to market values for the GPU fleet, and the specific trigger events monitored quarterly. Companies that disclose these elements report lower stock price volatility around impairment announcements because investors have been pre-educated on the range of possible outcomes. A typical disclosure might state: "A 10 percent decline in GPU market prices would result in an impairment charge of approximately $1.2 million based on the Company's GPU carrying value of $18.5 million as of December 31, 2026."

Private AI companies should maintain the same impairment testing rigor even though they face less public scrutiny. Venture debt agreements typically include covenants that reference GAAP financials, and an unrecognized impairment that is later discovered during an audit can trigger retroactive covenant breaches. Companies planning for IPO or acquisition should establish robust GPU impairment testing processes at least 12-18 months before the transaction to avoid the discovery of impairment-related accounting issues during due diligence.