How Long Do AI GPUs Last? 6-Year Accounting vs 2-3 Year Claims

AI GPUs last longer than the harshest critics claim, but not as long as the most generous accounting implies. Data center GPUs typically keep working for five years or more, and they tend to hold a frontline role for two to four years, according to Data Center Knowledge. Older chips are still earning money in late 2026, yet part of that strength reflects a shortage rather than a permanent rule. For a business buying its own GPU servers, a four-year budget life with a three-year stress test is a sensible starting point.

How long do AI GPUs physically last?

Physical life and economic life are separate questions. Data Center Knowledge reported on September 8, 2026 that most data center GPUs will run for at least five years, and potentially much longer, if lifespan only means the hardware still functions (Data Center Knowledge). The same report says failures usually come from external or board-level causes, such as thermal stress, inadequate cooling, dust and humidity. It also says warranty and support windows, often three years, frequently become the practical ceiling even when the hardware still works. Economic life is shorter. Data Center Knowledge puts it on the order of two to four years if you track Nvidia's release cadence. Shorter estimates exist too: Tom's Hardware notes that an unnamed Google architect put data center GPU service life at one to three years back in 2024 (Tom's Hardware). Our guide to hardware refresh cycles covers the operational side of that decision.

What depreciation periods do Amazon, Meta, Microsoft and Google use?

The largest operators do not agree, which shows that useful life is a judgment rather than a physical constant. These policies generally cover servers and network equipment, so they are not a direct measurement of GPU life.

• Amazon cut the life of a subset of its servers and networking equipment from six years to five as of January 2025, after a study found "an increased pace of technology development" in AI and machine learning (The Register).
• Meta moved most servers and network assets to 5.5 years, effective January 1, 2025, according to Hudson Labs, which cites Meta's 10-K.
• Alphabet uses a six-year schedule for servers, according to theCUBE Research in November 2025.
• Microsoft's filing gives a range of two to six years for equipment, and theCUBE Research treats that as six (theCUBE Research). Microsoft's extension from 15 to 25 years, effective at the start of fiscal 2027, covers data centers and office buildings, not servers (Hudson Labs).
• Among GPU clouds, theCUBE Research listed CoreWeave at six years, Lambda at five and Nebius at four as of November 2025 (theCUBE Research).

What is Michael Burry's GPU depreciation argument?

Michael Burry, the investor made famous by The Big Short, argued in an X post on November 10, 2025 that hyperscalers understate depreciation by extending useful lives while buying Nvidia hardware on a two- to three-year product cycle (CNBC). Burry estimated the understatement at about $176 billion from 2026 through 2028, and said profits at Oracle and Meta could be overstated by roughly 27% and 21% by 2028. That figure is his own estimate, and CNBC said it could not independently confirm the practice. Burry had also disclosed put options against Nvidia and Palantir, so he was not a neutral observer.

The accounting rules leave room for both sides. An analysis in The National Law Review says a change in useful life is a change in estimate under ASC 250, must be supportable and auditable, and depends on how each company expects to use its assets. A longer life is therefore not improper in itself, but it remains a management estimate that can turn out wrong.

Are older GPUs like the A100 still in use?

They are, and the most recent evidence favors longer lives than Burry suggests.

• Nvidia CFO Colette Kress said on the company's November 2025 earnings call that A100 GPUs shipped six years earlier were still running at full utilization (call transcript).
• CoreWeave signed an A100 contract running into 2029, nine years after the chip debuted, and CEO Mike Intrator said in August 2026 that "pricing for prior generation SKUs is at or above where it was years ago" (Tom's Hardware).
• Intrator also told CNBC that a batch of H100s coming off an expired contract was rebooked at 95% of the original price, according to Tom's Hardware.
• Silicon Data's neocloud H100 rental index stood at $2.97 per GPU-hour on November 5, 2026 (Silicon Data), up from $2.57 on May 4, 2026 (Silicon Data).

Treat this as strong evidence rather than neutral evidence. Nvidia and CoreWeave both benefit when investors believe GPUs last a long time. An index reading is also a market rate for rented hours, not proof of what any single used server would earn.

Why do old GPUs hold their value, and will that last?

Two forces are doing most of the work. The first is scarcity: Nvidia said in August 2026 that it expects supply to remain a bottleneck at least through the end of its fiscal 2028 (call transcript). The second is facility fit. An air-cooled DGX A100 draws 6.5kW at maximum load and suits legacy halls designed for roughly 20kW per rack, while GB200 and GB300 NVL72 racks draw 120kW to 140kW and need direct-to-chip liquid cooling (Tom's Hardware). Neither force is permanent. Nvidia claims Vera Rubin NVL72 delivers one-tenth the cost per million tokens of GB200 NVL72, a performance claim it says is subject to change (NVIDIA). If gains near that size reach the market once supply catches up, older GPUs priced per token could lose value faster than today's rental rates suggest. That is our judgment, not a forecast from any source cited here.

Does a six-year life apply to a business buying its own servers?

Often it does not. Hyperscalers can move a GPU from training to inference to lighter work, a pattern theCUBE Research calls a value cascade, and it expects them to converge on five years. A business running one or two workloads has fewer places to cascade hardware. It also faces the three-year warranty window that Data Center Knowledge describes, unless it buys extended support. The opposite point also holds. An enterprise does not compete on cost per token, so a server stays useful for as long as it meets the latency, memory and security needs of its workload. A stable internal inference service can run well past the point where a cloud provider would retire the same GPU from frontline duty.

What useful life should you budget for GPU servers?

We suggest modeling three cases. This is a planning assumption rather than accounting or tax advice, so confirm the book and tax treatment with your accountant. The figures below are simple straight-line arithmetic per $100,000 of hardware.

• Stress case, three years: $33,333 a year. Use this if your workload changes quickly or your models are likely to outgrow GPU memory.
• Base case, four years: $25,000 a year. Use this for most production inference and internal AI platforms.
• Extended case, five or six years: $20,000 or $16,667 a year. Use this only if you have support coverage for those years and a workload that will tolerate older hardware.

Set residual value to zero in the base case and treat any resale proceeds as a bonus. If the project only works at six years, it is carrying more risk than the budget shows. If utilization is uncertain or low, renting may be the better answer, and our comparison of neoclouds and buying GPU servers walks through that choice. Current market prices are in our guide to H100 and H200 pricing, and the wider cost model is in our three-year TCO guide.

Nexus Compute supplies new GPU servers through authorized distribution channels with the full manufacturer warranty, so your service life starts with clean support coverage. We check compatibility against your platform before dispatch and quote within 48 business hours, with purchase orders and net terms for qualified organizations. Request a quote or email sales@nexus-compute.com with your workload and planned service life.