Hardware lifecycle, warranty, and retirement for AI data centers

In short

AI data center GPUs have a practical lifespan of two to three years for cutting-edge training workloads, and Meta’s Llama 3 405B training study implied an annualized failure rate of approximately 9 percent for high-utilization training clusters. Stanley Laman Group analysis Yet federal tax law treats them as 5 year property and many large operators report them on 5 to 6 year depreciation schedules for financial accounting, creating a gap between real replacement cost and reported earnings. 26 U.S.C. § 168, theCUBE Research, Tom’s Hardware, IRS Pub. 946 The One Big Beautiful Bill Act of 2025 restored 100 percent bonus depreciation for equipment acquired and placed in service after January 19, 2025. Tax analysis When GPUs are retired, an operator should follow NIST media sanitization guidelines, navigate federal and state electronic waste rules, comply with export controls if reselling overseas, and avoid the kind of decommissioning failure that cost Morgan Stanley $155 million in fines and settlements. NIST SP 800-88, EPA, BIS, OCC, SEC, Reuters, SEC, Reuters, State AG settlement

How long do AI GPUs actually last

A traditional CPU server might run for five to seven years before it is refreshed. AI GPU clusters are different. Their physical useful life under a heavy training workload is often only one to two years, with three years as a maximum. Tom’s Hardware, Princeton CITP analysis

The stress comes from constant high power draw. A single H100 can pull over 700 watts, and a cluster of 16,384 of those GPUs, the size Meta used to train its Llama 3 405B model, generates an immense thermal and electrical load. Over a 54-day training run, that cluster logged 419 unplanned disruptions. Of those, 148 were GPU failures, including NVLink errors, and 72 were high bandwidth memory failures. The annualized failure rate was roughly 9 percent, which works out to about a 27 percent cumulative failure rate over three years. Tom’s Hardware

A Google principal generative AI architect, speaking anonymously, estimated a service life of one to three years at 60 to 70 percent utilization, and said physical failure is likely within one to two years at maximum load. That claim has been widely repeated, though it remains unattributed to a named individual. The Meta failure data showed GPU related failures accounted for about 30 percent of disruptions during Llama 3 training, which Tom’s Hardware characterized as quite favorable and extrapolated to a 9 percent annualized failure rate. Tom’s Hardware

Even outside of catastrophic failure, the hardware refresh cycle for AI clusters has compressed to 18 to 36 months, compared with the traditional five to seven year server cycle. A new NVIDIA architecture arrives every 18 to 24 months, H100 followed A100 by two years, and the B200 followed two years later, and each generation can deliver performance leaps that create tension between rapid technology cycles and depreciation planning, though hardware may retain value through workload cascading. Introl, NVIDIA, NVIDIA Newsroom

How fast are AI GPUs depreciated

There are two depreciation answers for the same GPU, one for tax and one for financial reporting.

Tax depreciation

For federal income tax, computers and peripheral equipment fall into the 5-year property class under the Modified Accelerated Cost Recovery System, or MACRS. That means a server and the GPUs inside it are depreciated over five years using the 200 percent declining balance method with a half-year convention. IRS Publication 946

Two provisions can dramatically accelerate the write-off.

Bonus depreciation. The One Big Beautiful Bill Act, enacted in 2025, restored 100 percent bonus depreciation for qualified property acquired and placed in service after January 19, 2025. This allows a full immediate deduction in the year the equipment is placed in service. Equipment that was acquired on or before January 19, 2025 but placed in service during 2025 qualifies for 40 percent bonus depreciation. BDO

Section 179 expensing. An operator can elect to immediately expense up to $2,500,000 of qualifying property placed in service in 2025. Bonus depreciation has no annual dollar limit and can be used to create a net loss. Thomson Reuters

Both of these are federal rules. States do not automatically follow. California, Massachusetts, New York, and others have decoupled from federal bonus depreciation. State conformity varies. An operator doing business in those states must maintain a separate state depreciation schedule.

Financial accounting depreciation

For GAAP financial statements, a company picks a useful life that reflects how long it expects to use the equipment. In 2020, the four major U.S. cloud providers depreciated servers over three to four years. Then they all began to extend those lives, citing improvements in server durability and utilization.

• Microsoft extended server life to 6 years in 2022, which reduced fiscal year 2023 depreciation by an estimated $3.7 billion. The Register
• Alphabet moved to 6 years in 2023, saving about $3.4 billion in annual depreciation. DCD
• Amazon extended its servers to 6 years and then, in January 2025, shortened them back to 5 years, explicitly citing the shorter lifespan of AI equipment. The company estimated the change would reduce 2025 operating income by roughly $700 million. Data Center Dynamics
• Meta stretched its server life from 4 years to 5.5 years in 2025, which was expected to save about $2.9 billion in depreciation that year. Yahoo Finance

Collectively, these extensions are estimated to have reduced the hyperscalers’ annual depreciation expense by approximately $18 billion. Introl

A table summarizes the shifts.

These are GAAP useful lives, not tax depreciation. The tax class life remains 5 years regardless of what a company reports on its financial statements.

The depreciation mismatch and why it matters for deals

When a GPU fails or becomes obsolete in two years but is carried on the books with a five or six year depreciation schedule, the reported operating income does not reflect the true economic cost of replacement.

A Princeton CITP analysis illustrated the gap. It assumed Microsoft’s $80 billion annual AI infrastructure spend had half allocated to computing hardware with a true 3 year lifespan. Under a 6 year depreciation schedule, Microsoft would report roughly $6.5 billion in annual depreciation. If the real life were 3 years, actual replacement cost would be about $13 billion. The difference, about $6.5 billion a year, represents an earnings cushion that can mask the true capital intensity of an AI business. Princeton CITP

Barclays cut its earnings forecasts for AI firms by up to 10 percent for 2025 in part to account for more realistic depreciation assumptions that reflect shorter GPU useful lives. Princeton CITP

Alphabet’s own financials illustrate the effect. While its net property, plant, and equipment rose by $50 billion from 2020 to 2023, its annual depreciation expense actually fell by $1 billion because of the useful life extensions. MBI Deep Dives

For a deal, this mismatch matters in several ways.

• Valuation. If a buyer or investor uses GAAP depreciation to model free cash flow, they may overvalue a business that will need to replace hardware faster than the accounting suggests.
• Loan covenants. Financial covenants tied to EBITDA or operating income may look healthier than they would if the company used a shorter depreciation life. A lender that does not adjust for the real replacement cycle is taking more risk than it sees.
• Residual value underwriting. Equipment financiers often underwrite a residual value at lease end. A GPU that is worth 50 to 70 percent of its original price on paper may have little or no remaining useful life if it is already three years old and competing against a newer, faster generation.

What warranty covers AI GPUs

Not all GPU warranties are the same. The standard NVIDIA consumer warranty and the enterprise support for AI data center hardware are completely different products, and confusing the two can be costly.

Consumer warranty

NVIDIA provides a three-year limited warranty on GeForce branded graphics cards. It covers manufacturing defects and hardware component failures. But the warranty contains a critical exclusion. It is void if the product is used for datacenter use or GPU cluster commercial deployments. NVIDIA Manufacturer’s Warranty

In other words, a GPU that is installed in an AI data center has no coverage under the standard consumer warranty. That warranty is also non transferable.

Enterprise support

For NVIDIA DGX systems and other AI data center products sold through authorized channels, the operator buys an enterprise support contract separately. These are not warranties in the consumer sense. They are commercial support agreements with defined service levels.

NVIDIA’s enterprise support tiers, Business Standard and Business Critical, provide 24/7 portal access, defined response times, and advance replacement. NVIDIA Enterprise Support and Services User Guide

These support entitlements are tied to the original purchaser’s company. They are non-transferable and terminate when the product is transferred to another party. The support contract must be renewed annually.

OEM warranties

Many NVIDIA data center GPUs enter the market through OEMs. Dell, HPE, and Supermicro integrate the GPUs into their own systems and provide their own warranty and support terms. Exactly what is covered, and for how long, depends on the OEM’s contract, not on NVIDIA’s consumer or enterprise support pages. A buyer should review the OEM agreement carefully.

NVIDIA networking hardware warranty

NVIDIA-branded networking switches, adapters, cables, and transceivers carry a one year limited hardware warranty. It functions as a factory RMA. For Factory RMA under the limited hardware warranty, NVIDIA covers shipping of the defective unit to its factory and the customer pays return shipping. NVIDIA Enterprise Support and Services User Guide sec. 5.9

Rising warranty costs

Warranty claims are climbing fast. NVIDIA’s warranty payouts rose from $81 million in 2024 to over $894 million in 2025, an 11 fold increase. Its warranty reserve fund swelled from $2.59 billion at the end of 2024 to $8.22 billion at the end of 2025. NVIDIA’s warranty claim rate moved from 0.17 percent at the start of 2025 to 0.9 percent by the end of the year. AMD’s warranty payouts also doubled over the same period, from $110 million in 2024 to up to $238 million in 2025. PCMag

The takeaway for an AI data center operator is that warranty and support costs are becoming a material line item, and that those costs may rise further as hardware is pushed harder.

How to retire and dispose of AI hardware

Retiring a GPU is not as simple as pulling it from a rack and deleting a few files. Three regulatory frameworks apply, data destruction, electronic waste, and export controls. A single decommissioning project can touch all three.

Data destruction, NIST and SEC

Before a GPU with embedded memory, an NVMe drive, or a motherboard with firmware storage leaves your control, the data on it must be destroyed to a standard that holds up in court and under a regulatory order.

The federal standard is NIST Special Publication 800-88, Revision 2, Guidelines for Media Sanitization. It is a NIST guideline developed under FISMA that federal agencies use for media sanitization programs. NIST SP 800-88 Rev. 2

NIST defines three levels.

• Clear. Logical techniques, such as overwriting, that protect against non-specialized recovery attempts. Suitable for media that will be reused inside the organization.
• Purge. Logical or physical processes, including cryptographic erase or block erase, that protect against laboratory-level recovery. Purge is the preferred method under Rev. 2 whenever the technology supports it.
• Destroy. Renders the media unusable and recovery infeasible. It requires shredding, disintegration, or incineration. Rev. 2 no longer recognizes degaussing as a destruction method. For solid-state drives, IEEE 2883-2022 deprecates shredding as a destruct method and instead requires disintegration, incineration, or melting IEEE 2883-2022, NIST SP 800-88 Rev. 2

Separately, SEC Regulation S-P, Rules 30(a) and 30(b), requires broker-dealers, investment advisers, and certain other financial institutions to properly dispose of customer information. The rule applies to broker-dealers, registered investment advisers, investment companies, funding portals, and transfer agents, and the SEC has enforced it aggressively when decommissioning goes wrong. SEC press release, SEC press release

Electronic waste, RCRA and state law

Many electronic components in AI hardware contain materials that are hazardous waste under the Resource Conservation and Recovery Act, or RCRA. That federal law imposes a cradle to grave tracking obligation. The generator of hazardous waste must classify it, handle it, manifest it, and send it to a permitted treatment or disposal facility. The generator remains liable even after the waste leaves its site. EPA Hazardous Waste Basics

A useful safe harbor exists. Shredded circuit boards being recycled are excluded from the definition of solid waste, provided that they are stored in containers sufficient to prevent a release to the environment prior to recovery and are free of mercury switches, mercury relays, nickel cadmium batteries, and lithium batteries. 40 C.F.R. § 261.4(a)(14)

Batteries can be managed under the Universal Waste rules, which simplify labeling and accumulation requirements, if the batteries remain intact. Damaged, defective, or recalled lithium batteries require stricter packaging under Department of Transportation rules. 49 C.F.R. 173.185(f)

Civil penalties for RCRA violations can reach $124,426 per day per violation. 90 FR 1375

In addition to the federal framework, 25 states plus the District of Columbia have their own mandatory electronic waste laws. Many impose landfill bans on electronic devices. Colorado, for example, prohibits the disposal of computer e-waste in landfills. Colorado landfill ban on electronic devices An operator retiring hardware in those states must meet state specific recycling and documentation requirements.

Export controls on used GPUs

A GPU that is too slow for the latest training cluster still holds enormous compute capability and can be a target for restricted end users overseas. The Bureau of Industry and Security, or BIS, controls the export of advanced integrated circuits under the Export Administration Regulations.

NVIDIA’s H100, H200, B100, B200, and A100, and equivalent AMD GPUs, are classified under ECCN 3A090.a. That classification requires a license for export to all destinations, not just China, if the chip has a Total Processing Performance of 4800 or more, or a TPP of 1600 or more with a performance density of 5.92 or more. Computers and assemblies containing those chips fall under ECCN 4A090.a. EAR, 15 C.F.R. Part 774, Supp. No. 1 (CCL), 15 C.F.R. 742.6(a)(6)(iii)(A)

The policy toward China has shifted several times.

• NVIDIA developed the H20 chip as a China market variant with reduced interconnect bandwidth, hoping to avoid the tightest controls. BIS imposed new license requirements on H20 sales to China in April 2025, effectively halting them. Data Center Knowledge
• In August 2025, the US government announced a limited arrangement allowing NVIDIA and AMD to sell H20 and MI308 chips to China under a 15 percent revenue share license with the U.S. government, while still restricting H100 and A100 series exports. Data Center Knowledge
• In early 2026, BIS shifted from a presumption of denial to a case by case review for chips below advanced thresholds, explicitly naming the H200 and AMD MI325X as eligible for export to China with certifications. The H100, as a less advanced chip than the H200, also fell under the case-by-case review policy. Federal Register, Federal Register

These controls apply to used equipment just as they apply to new. An operator that sells a retired H100 to a buyer who then ships it to China can be liable for an unlicensed export. BIS civil penalties can reach approximately $370,000 per violation or twice the transaction value, whichever is greater. Criminal penalties can reach $1 million and imprisonment up to 20 years. Tax alert

The Morgan Stanley decommissioning failure and what it teaches

In 2016, Morgan Stanley decommissioned two data centers. It hired a moving company with no data destruction experience instead of its usual vendor, IBM, to save approximately $100,000. The moving company sold the decommissioned equipment through intermediaries. About 4,900 devices changed hands, including 53 RAID arrays containing roughly 1,000 hard drives and 8,000 backup tapes. Some of the drives appeared on an internet auction site with unencrypted personal information still stored on them. The SEC found that the firm’s failures over a five-year period affected the personal identifying information of approximately 15 million customers. Reuters, DCD

An IT consultant in Oklahoma who bought drives at auction discovered the data and alerted Morgan Stanley in 2017. Even after the initial breach, later audits found that 42 servers from a separate hardware refresh program were missing, and that the firm had failed to activate the encryption software on them for years. DCD

The regulatory and legal consequences were severe.

• The Office of the Comptroller of the Currency fined Morgan Stanley $60 million in 2020.
• The SEC imposed an additional $35 million penalty in 2022 for violating Regulation S-P’s Safeguards and Disposal Rules.
• A class action lawsuit settled for $60 million. SEC, OCC, Settlement

Total fines and settlements reached more than $150 million, over 1,500 times the $100,000 that was saved by cutting corners on decommissioning. DCD

The case illustrates the risks of treating IT asset disposition (ITAD) as a low-priority logistics task rather than a compliance function. The SEC’s order highlighted several failures, no written decommissioning policy, no vendor assessment for data security competence, no chain of custody tracking, and no post decommissioning audit to confirm that data was actually destroyed.

What the secondary market tells you about residual value

A working GPU that is two or three years old retains significant resale value, even if its useful life for training cutting-edge models is ending. The secondary market has matured quickly.

• Used H100 GPUs traded as high as $50,000 per unit during the 2024 supply crunch. By late 2025, after supply increased, used units median prices fell toward the mid $20,000s, and refurbished H100s at roughly two years old held 80 to 90 percent of contemporaneous new pricing. GPU market analysis
• Used A100 40GB GPUs trade at roughly $8,000 to $12,000 as of early 2026. The 80GB variant trades at roughly $12,000 to $18,000, representing roughly 48 to 72 percent of original new pricing. Introl
• CoreWeave reported that H100 GPUs coming off three-year contracts in 2025 were immediately rebooked at 95 percent of their original pricing. Its A100 GPUs from 2020 remain fully booked for inference workloads. Introl

The secondary market creates real residual value, but that value depends on a working, tested GPU with clean provenance, not a decommissioned piece of scrap. The resale also triggers export control screening, because a buyer may intend to ship the hardware to a restricted destination.

Practical steps for counsel, tax directors, and sponsors

For deal underwriting and financing. Model the real replacement cycle of one to three years, not just the GAAP depreciation schedule. Check whether loan covenants tied to EBITDA or asset coverage ratios are distorted by long depreciation lives. If the business will need to refresh hardware every two years, that cash outflow needs to be in the model.

For tax planning. Use 100 percent bonus depreciation for federal purposes where the acquisition and placed in service dates qualify. Confirm whether each state where the operator files follows or decouples from the federal bonus rule. Maintain separate state depreciation records where needed.

For lease and financing structures. The original use and acquisition date tests for bonus depreciation can be tricky with leased equipment. A lease that does not transfer the benefits and burdens of ownership may not qualify. Confirm with tax counsel before assuming a lease structure captures bonus depreciation.

For retirement and ITAD. Engage a certified ITAD vendor whose processes align with NIST 800-88 Rev. 2 and who can produce a documented chain of custody. Retain that documentation for at least three to seven years. Cloudaware Before any device leaves the site, verify that the data destruction method matches the sensitivity of the data, Destroy for regulated financial data, Purge or Clear for internal equipment being redeployed.

For e-waste compliance. Classify all retired hardware under RCRA and any applicable state e-waste law before it leaves the facility. Use the shredded circuit board exclusion for recycling, but confirm that batteries and mercury components are removed first. Treat damaged lithium batteries under the stricter packaging rules.

For export controls. Screen every prospective buyer and the ultimate destination against the BIS Entity List and the ECCN classification of the GPUs being sold. Assume that a used H100 requires the same license as a new one. If the buyer cannot provide an end-use certification that satisfies your compliance team, do not complete the sale.

Morgan Stanley in one number. Saving $100,000 on decommissioning led to over $155 million in fines and settlements. The most important step is retaining control of the hardware until the data is confirmed destroyed.

Key takeaways

• AI GPUs under a heavy training load physically last about one to three years, with a 9 percent annual failure rate and a cumulative failure rate near 27 percent over three years.
• For federal tax, GPUs are 5-year property and can qualify for immediate 100 percent bonus depreciation if placed in service after January 19, 2025, but state conformity varies.
• The large cloud operators report servers on 5-to-6-year GAAP useful lives, creating a substantial gap between reported depreciation and the actual replacement cost.
• Consumer NVIDIA warranties are void for AI data center use. Enterprise GPU coverage is a separate support contract that is non-transferable and must be renewed.
• Retirement of AI hardware requires compliance with NIST 800-88 data destruction standards, RCRA e-waste rules, state electronic waste laws, and BIS export controls, all at the same time.
• The Morgan Stanley ITAD failure teaches that cutting corners on decommissioning can cost over 1,500 times the amount saved.

Frequently asked questions

Q:What is the actual physical lifespan of an NVIDIA H100 GPU?

A:Under a heavy AI training workload, a datacenter GPU can begin to fail within one to two years, with a typical service life of one to three years at 60 to 70 percent utilization. A significant factor is thermal and electrical stress from sustained high power draw, with datacenter GPUs drawing and dissipating 700W or more. Tom’s Hardware

Q:Can I depreciate a GPU in one year for tax purposes?

A:Yes, if the GPU qualifies for 100 percent bonus depreciation. Under the OBBBA, equipment acquired and placed in service after January 19, 2025 can be fully deducted in the year it is placed in service. Equipment acquired earlier but placed in service in 2025 qualifies for 40 percent bonus depreciation. Section 179 also allows immediate expensing up to $2,500,000, subject to a phase-out and taxable income limit. IRS Form 4562 Instructions

Q:Does NVIDIA’s standard three-year warranty cover data center GPUs?

A:No. The three-year consumer warranty for GeForce cards explicitly excludes use in a commercial AI data center or GPU cluster. AI data center operators need an active enterprise support contract to receive support beyond the limited hardware warranty. Support may be included with the initial purchase or obtained through an authorized NPN reseller, OEM, or NVIDIA, and support entitlements are non-transferable. NVIDIA Enterprise Support and Services User Guide

Q:What is the difference between GAAP and tax depreciation for servers?

A:Tax depreciation uses the IRS-prescribed 5-year MACRS class life, regardless of how long the equipment actually lasts. GAAP useful life is management’s estimate of how long the server will be used. Large operators report server lives between 5 and 6 years for financial statements, and theCUBE Research argues AI GPUs have an economic life extending to 5 or 6 years by cascading from training to inference workloads, disputing claims that useful life is only 2 to 3 years. theCUBE Research

Q:What does NIST Clear, Purge, and Destroy mean?

A:Clear is a logical overwrite that protects against ordinary recovery tools. Purge uses cryptographic erase or block erase to protect against laboratory-level recovery. Destroy physically shreds or incinerates the media so that recovery is infeasible. Rev. 2 prefers Purge when the technology supports it, and it no longer recognizes degaussing as a destruction method. NIST SP 800-88 Rev. 2

Q:Do I need an export license to sell a used H100 to a buyer overseas?

A:Most likely yes. The H100 falls under ECCN 3A090.a, which requires a worldwide license for export. Selling to a buyer in China is categorically prohibited unless a specific license has been granted. BIS penalties can reach $374,474 per violation. BIS penalties, 15 C.F.R. 764.3

Q:How much did Morgan Stanley pay for its ITAD failure?

A:$155 million across OCC and SEC fines and a class action settlement. On multiple occasions, MSSB hired a moving and storage company with no experience or expertise in data destruction services to decommission hard drives and servers. SEC

Q:Can I simply throw old GPUs in the trash?

A:In many states, no. 23 states have landfill bans on electronic devices, and 25 states plus DC have mandatory e-waste laws. Colorado, for example, prohibits disposal of computer e-waste in the trash. Federal RCRA rules may also apply if the hardware exhibits a hazardous waste characteristic. EPA, 40 C.F.R. Parts 260-273

Q:What should I look for in an ITAD vendor?

A:The vendor should have a documented process aligned with NIST 800-88 Rev. 2, the ability to provide a serialized chain-of-custody report, and experience with the specific hardware you are retiring. Ask for evidence of R2 or e-Stewards certification, proof of liability insurance, and a clear description of how the vendor will handle batteries and other hazardous components.

Q:How does the secondary market for used GPUs affect residual value in a lease?

A:A working, tested GPU retains significant value. Refurbished H100s at two years old have held 80 to 90 percent of new pricing. But that value is only realized if the GPU is properly maintained, sanitized, and sold with a clear export control screening. A GPU sold as used rather than refurbished may be worth far less. Silicon Data

Subscribe to The Compute Law Brief

The Compute Law Brief is a free weekday newsletter on the law of AI infrastructure across tax, real estate, construction, power, and deals. The big US build markets and federal law. Three minutes a morning. No paywall, and no email gate to read the blog. Subscribe if you want it in your inbox.

Junde Liu, JD, LL.M. (Taxation) candidate at UF Law. Originally published on Compute Law Blog. This article is general information and does not constitute legal advice. Reading it does not create an attorney client relationship. The reader should not act on the basis of any content here without first consulting a licensed attorney in the relevant state. Last reviewed for accuracy May 23, 2026.