GPU Depreciation: Book Life, Economic Life and Loan Terms
Summary
- GPU depreciation is the decline in a GPU server's value over time, usually booked straight-line over a disclosed useful life. Most public companies set that life at five to six years for servers and network equipment.
- Tax depreciation can be immediate through 100% bonus depreciation or Section 179, and it runs separately from book life and economic life.
- In our view, debt should be sized to a life shorter than book life. USD.AI's underwriting models four years, and a loan that amortizes within that life keeps loan-to-value falling.
Introduction
If you own your cluster and carry debt on it, GPU depreciation runs on three clocks at once. The books spread the cost over a useful life, and the tax code may let you deduct it in year one. The market decides how long the GPUs actually earn. Only the last one repays the loan.
Picking the wrong clock has a cost. A useful life that is too long overstates margins, and a loan sized to it can outlive the servers' earning power. A depreciation schedule for GPUs that matches how they earn keeps the debt ahead of the decline in value.
This article covers the three lives and what public companies disclose about server useful lives. It works through the straight-line math on a B300 server and explains what bonus depreciation and Section 179 change. It ends with how lenders size the loan term.
Book Life, Tax Life and Economic Life Are Three Different Clocks
GPU depreciation is the decline in a GPU server's value as it ages, recorded against three separate lives. On the books, most public filers spread the cost straight-line over five to six years. For tax, a qualifying purchase can be deducted in the year it is placed in service. The economic life lasts as long as the servers earn more than they cost to run.
The three numbers rarely match, and each answers a different question. Book life shapes reported margins, tax life shapes cash taxes, and economic life decides whether the revenue outlasts the debt.
Book life: the useful life on the financial statements
Book depreciation spreads cost minus salvage value evenly over a useful life that management estimates, a method called straight-line. Salvage value is what the company expects to recover when it retires the asset. The estimate belongs to management, and it can change.
When a company revises a useful life, the change applies prospectively to the remaining book value. Past periods stay as reported. The revision moves reported earnings and leaves cash unchanged, which is why a longer life can lift margins without a dollar more coming in.
Tax life: when the deduction hits
Tax rules can let an operator deduct a server's cost far faster than the books do, often in year one through bonus depreciation. That changes when cash taxes fall due and nothing about how long the hardware earns. The rules and limits are covered in the tax section below.
Economic life: how long the GPUs earn more than they cost to run
Economic life is the clock lenders care about, because it is the one the revenue runs on. USD.AI's glossary defines it as "how long a GPU can earn more than it costs to run," and it is often shorter than physical life.
Rental rates drive economic life, because what customers pay per GPU-hour sets how much each server earns over its costs. New generations reprice old ones. NVIDIA announced the Rubin platform on Jan 5, 2026, and every launch like it resets what buyers expect from older servers.
Physical life is how long the hardware keeps running, and it is usually longer than both book life and economic life. A server that still powers on has not necessarily kept its earning power.
What Companies Disclose About Server Useful Lives
Disclosed server lives generally sit between five and six years, and recent filings show them moving in both directions. The table uses each company's own asset class wording, and only one filing names GPUs.
| Company | Filing and date | Asset class wording | Useful life | Notes |
|---|---|---|---|---|
| Amazon | FY2025 10-K | A subset of "servers and networking equipment" | Five years, cut from six effective Jan 1, 2025 | Cites the pace of AI and machine learning; +$1.4B depreciation and −$1.0B net income in 2025 |
| Alphabet | FY2025 10-K (Feb 2026) | Servers and network equipment | Six years | "We depreciate servers and network equipment generally over a period of six years." |
| Microsoft | FY2026 10-K (fiscal year ended Jun 30, 2026) | Servers and network equipment | Two to six years | One range for the class |
| Oracle | 10-Q, quarter ended Aug 31, 2026 (Note 3) | Equipment | Six years | "Comprised primarily of servers and networking equipment with estimated useful life of six years" |
| CoreWeave | FY2025 10-K | "Technology equipment" | Six years | Straight-line |
| Nebius | 6-K (Aug 12, 2026) | Server and network equipment | Five years, raised from four from Q1 2026 | Changed "to reflect usage patterns and current utilization commitments"; applied prospectively |
| IREN | 10-K (fiscal year ended Jun 30, 2026) | "HPC hardware" | Five years | Straight-line |
| Hut 8 | FY2025 10-K | "AI GPUs" | Five years | Straight-line |
| TeraWulf | FY2025 10-K (filed Feb 27, 2026) | Computer equipment (general policy) | Five years | "generally 5 years for computer equipment"; straight-line |
Most hyperscalers disclose five to six years, while Microsoft gives a two-to-six-year range. Amazon shortened a subset for AI, while Nebius lengthened to five years on contracted utilization. IREN and Hut 8 disclose five, and Hut 8 is the only filer here that names GPUs as the asset class.
In our view, debt should be sized to a shorter life than these filings use, and USD.AI's underwriting models four years. The stakes for reported earnings are large: in a Jan 2026 scenario, J.P. Morgan Asset Management estimated a three-year life would cut EPS by about 6–7% and operating margin by about 6–9%. That result covers four of the five hyperscalers it studied, under the bank's own scenario assumptions.
How to Calculate GPU Depreciation on a B300 Server
One NVIDIA HGX B300 server at about $600,000 depreciates $150,000 a year over four years, $120,000 over five and $100,000 over six. That price comes from USD.AI's pricing research for Q4 2026 and covers the server only, before tax.
The formula is the same at any life: annual depreciation equals cost minus salvage value, divided by useful life. Monthly depreciation is the annual figure divided by 12. The table assumes zero salvage value and is illustrative.
| Useful life | Annual depreciation | Monthly depreciation |
|---|---|---|
| 4 years | $150,000 | $12,500 |
| 5 years | $120,000 | $10,000 |
| 6 years | $100,000 | ~$8,333 |
A salvage assumption lowers each year's charge by the salvage value divided by the life. The server price is also only part of the asset base. Networking, owned racks, PDUs and cabling add to deployment cost, while installation, labor, shipping and taxes are soft costs that sit outside it.
What Bonus Depreciation and Section 179 Change for Cash Taxes
Tax rules can let you deduct a qualifying server purchase in the year it is placed in service. That changes your cash taxes and nothing about how long the GPUs earn.
The IRS issued IR-2026-06 and Notice 2026-11 on Jan 14, 2026. They confirm a permanent 100% additional first-year depreciation deduction for qualified property acquired and placed in service after Jan 19, 2025. The Notice is interim guidance, and acquisition is tested under written binding contract rules.
Section 179 is a second route to an immediate deduction, with a cap. For tax years beginning in 2026, Publication 946 sets the maximum Section 179 deduction at $2,560,000. That limit is reduced by the amount by which Section 179 property placed in service exceeds $4,090,000.
The rules also let you slow the deduction down. You can elect not to claim the special allowance for a class of property. Under Notice 2026-11, you can instead elect a 40% allowance in place of 100% for the first tax year ending after Jan 19, 2025.
Tax depreciation, book life and the economic life a lender underwrites are separate numbers. An immediate deduction leaves the other two where they were. Whether a specific server purchase qualifies depends on facts the guidance does not settle for you. Talk to your tax adviser.
Why the Loan Pays Down Faster Than the Servers Lose Value
USD.AI's underwriting models a conservative four-year useful life. Loans cover up to 80% of deployment cost over terms of 3–5 years, with straight-line monthly amortization. Straight-line amortization means equal monthly principal payments.
The worked example below shows why a term inside the modeled life keeps loan-to-value falling.
The underwriting assumption: a four-year modeled life
The assumption is set out in USD.AI's published underwriting standard: "USD.AI's underwriting models a conservative four-year useful life." That is one to two years shorter than the five-to-six-year lives most filers disclose.
The operator contributes at least 20% of deployment cost as equity. As of October 2026, the approved collateral list covers NVIDIA B200, B300, GB200, GB300 and Vera Rubin (VR200), plus AMD Instinct MI455X and MI430X.
Worked example: loan balance against server value, year by year
Take one $600,000 B300 server financed with a $480,000 loan at 80% loan-to-value and $120,000 of operator equity. Loan-to-value is the loan balance divided by the server's depreciated value. The figures are illustrative and assume zero salvage value and server-only, pre-tax Q4 2026 pricing from USD.AI's pricing research.
| End of year | Value, 4-year life | Value, 6-year life | Balance, 3-year loan | Balance, 4-year loan | LTV, 3-year loan vs 4-year life | LTV, 4-year loan vs 4-year life | LTV, 3-year loan vs 6-year life | LTV, 4-year loan vs 6-year life |
|---|---|---|---|---|---|---|---|---|
| 1 | $450,000 | $500,000 | $320,000 | $360,000 | 71% | 80% | 64% | 72% |
| 2 | $300,000 | $400,000 | $160,000 | $240,000 | 53% | 80% | 40% | 60% |
| 3 | $150,000 | $300,000 | $0 | $120,000 | 0% | 80% | 0% | 40% |
| 4 | $0 | $200,000 | – | $0 | – | – | – | 0% |
Over a four-year life, the three-year loan's LTV falls each year while the four-year loan's holds at 80%. Over a six-year life, both fall. Monthly principal is $13,333 on the three-year loan and $10,000 on the four-year loan.
The example does not show that a loan always pays down faster than value falls. A four-year loan matched to a four-year life only keeps pace. In our view, the term should sit inside the modeled life, and the life should be the conservative one. That is the core of how lenders underwrite GPU loans.
What this means for the operator's offtake contract
Revenue repays the loan, so contracted revenue should run at least as long as the amortization. On the three-year loan above, the offtake contract has to cover $13,333 of monthly principal per server for 36 months. Interest and operating costs come on top.
Ownership puts this math on your balance sheet. Rising rental costs are one reason more companies are choosing to own their compute.
Conclusion
Book lives of generally five to six years, immediate tax deductions and economic life are three separate numbers. Finance against the shortest one, because that is the clock your revenue runs on.
USD.AI's underwriting models a four-year useful life, with terms of 3–5 years, straight-line monthly amortization and loans of up to 80% of deployment cost. Compute IQ scores every deal on a 100-point scale across the operator, the customer and the contract. The score sets one of three pricing tiers, and deals below 60 aren't financeable. Compute IQ is built on more than $25 billion of GPU financing data points.
As of September 2026, USD.AI had reviewed over 700 GPU financing deals. If you are planning a deployment, GPU financing through USD.AI sizes the loan to your deployment and offtake contract. Talk to USD.AI.
FAQ
How is a GPU depreciated?
A GPU server is depreciated straight-line on the books, with cost minus salvage value spread evenly over its useful life. For tax, the deduction can be taken immediately through bonus depreciation or Section 179. Its economic life is set separately, by how long it earns more than it costs to run.
How long do GPUs last in data centers?
Most public companies disclose book lives of five to six years for servers. Economic life, how long the GPUs earn more than they cost to run, is often shorter than physical life. USD.AI models a four-year useful life.
Does 100% bonus depreciation apply to GPU servers?
The permanent 100% first-year deduction covers qualified property acquired and placed in service after Jan 19, 2025, and the IRS guidance does not single out GPU servers. Whether a given purchase qualifies depends on the property and its acquisition date under the written binding contract rules. Talk to your tax adviser.
