OpenModelsHQ

How long will a GPU keep doing useful work?

A GPU can keep working after it has been written down, and stop being economical before it fails. Its useful life depends on the workload, software support and the cost of producing acceptable results.

Compare launch datesCalculate an ownership scenarioReport your experience

Editorial illustration of accelerators and server racks
Editorial illustration · conceptual hardware

Keep the different clocks separate

Physical service
Whether a particular unit still functions. Active units have not yet reached an observed end of life.
Software support
Which drivers, toolkits and frameworks still support the hardware and your workload.
Economic usefulness
Whether useful output earns or saves enough to cover the full operating and ownership costs.
Accounting and lending
Book depreciation and loan terms serve different purposes. Neither gives a GPU an expiry date.

Launch and current evidence

A dated manufacturer announcement is labelled separately from the catalogue's launch month or year. Age below describes the product generation at 14 Nov 2026, not the service age or remaining life of a unit you buy.

HardwareAnnouncement or recorded launchGeneration ageModels fitting in memoryEvidence
A100 80GB SXMGPU16 Nov 2020 (announced)5 years409
Arc Pro B60GPU2025 (catalogue)0 to 1 years360No duration evidence recorded
H100 SXMGPU22 Mar 2022 (announced)4 years409
H200 SXMGPU13 Nov 2023 (announced)3 years420No duration evidence recorded
L4GPU2023 (catalogue)2 to 3 years360No duration evidence recorded
L40SGPU2023 (catalogue)2 to 3 years396No duration evidence recorded
Radeon AI PRO R9700GPUJul 2025 (catalogue)1 year391No duration evidence recorded
GeForce RTX 3090GPUSep 2020 (catalogue)6 years360No duration evidence recorded
GeForce RTX 4090GPUOct 2022 (catalogue)4 years360No duration evidence recorded
GeForce RTX 50 Super seriesGPUCES 2027 (reported) (rumored)Not establishedNot calculatedNo duration evidence recorded
GeForce RTX 5060 Ti 16GBGPUApr 2025 (catalogue)1 year341No duration evidence recorded
GeForce RTX 5070 TiGPUFeb 2025 (catalogue)1 year341No duration evidence recorded
GeForce RTX 5080GPUJan 2025 (catalogue)1 year341No duration evidence recorded
GeForce RTX 5090GPUJan 2025 (catalogue)1 year391No duration evidence recorded
RTX PRO 6000 BlackwellGPU2025 (catalogue)0 to 1 years417No duration evidence recorded
RubinGPUThe standalone GPU catalogue milestone is unresolved. Early-access Rubin NVL72 deployments are recorded separately.GPU-only launch not establishedNot established451
Radeon RX 7900 XTXGPUDec 2022 (catalogue)3 years360No duration evidence recorded
Radeon RX 9070 XTGPUMar 2025 (catalogue)1 year341No duration evidence recorded
Tesla V100 familyGPU family10 May 2017 (announced)9 yearsNot calculated
GB300 NVL72Rack-scale system18 Mar 2025 (announced)1 yearNot calculated
Tesla T4GPU12 Sep 2018 (announced)8 yearsNot calculatedNo duration evidence recorded
B200GPU18 Mar 2024 (announced)2 yearsNot calculatedNo duration evidence recorded
Instinct MI300XGPU13 Jun 2023 (announced)3 yearsNot calculatedNo duration evidence recorded
Vera Rubin NVL72Rack-scale systemNVIDIA reports early-access availability and production customer workloads in September 2026. This is a rack, not a standalone GPU SKU.2026 (catalogue)Under 1 yearNot calculated

Memory fit uses our existing Q4 estimate at the configured context. It is not a throughput or quality benchmark. No row assigns a predicted number of years remaining.

What the published evidence says

These are selected disclosures and claims, with their scope retained. Vendor-reported deployments and vendor-cited forecasts are not independent fleet survival studies.

Reported deployment · published 1 Oct 2026

A100 generation

NVIDIA reports continued commercial A100 service and CoreWeave bookings extending through 2029.

Generation-level vendor report, not a unit commissioning cohort. It does not establish future reliability or profit.

Read the source

Forecast · published 1 Oct 2026

Eight-GPU H100 system

5 to 6 years estimated

NVIDIA cites Barkr estimating useful life for an eight-GPU H100 system.

A system-level estimate cited by the vendor. Not a guarantee for one H100 card or a measured failure distribution.

Read the source

Forecast · published 1 Oct 2026

GB300 NVL72 system

9 to 10 years estimated

NVIDIA cites Barkr estimating useful life for a GB300 NVL72 system.

A rack-scale forecast, not a completed service history. No individual GPU remaining-life estimate is inferred.

Read the source

Reported deployment · published 30 Sep 2026

CoreWeave V100 deployments

NVIDIA reports that CoreWeave V100 GPUs continue to run customer workloads.

A vendor statement about surviving deployments, without unit-level commissioning dates or failed-unit counts.

Read the source

Software support · checked 4 Oct 2026

Volta architecture, including V100

CUDA Toolkit 13 removed offline compilation and library support for Maxwell, Pascal and Volta.

A toolchain support change. Existing compatible binaries and older toolkits are a separate question; this is not a physical failure date.

Read the source

Accounting life · report period 31 Dec 2025

Alphabet servers and network equipment

6 years accounting assumption

Alphabet describes depreciation of servers and network equipment generally over six years.

Accounting estimate in the cited disclosure, not a GPU-specific survival or resale forecast.

Read the source

Accounting life · effective 1 Jan 2025

A subset of Amazon servers and networking equipment

5 years accounting assumption

Amazon changed the estimated useful life of a subset from six years to five, effective January 2025.

The change applies to a subset and is historical disclosure, not a universal cloud-server lifespan.

Read the source

Financing · published 31 Mar 2026

CoreWeave DDTL 4.0 facility

CoreWeave disclosed a facility secured by infrastructure and an associated customer contract, with maturity in March 2032.

Financing maturity and contract credit are not evidence that an individual GPU remains productive through that date.

Read the source

Reported deployment · published 30 Sep 2026

CoreWeave Vera Rubin NVL72 systems

NVIDIA reports production racks received earlier in September and Cognition running production workloads, with early-access availability.

Reported deployment of a specific rack-scale system. It does not establish fleet survival or the release status of every Rubin-based GPU product.

Read the source

A longer depreciation schedule lowers the annual expense for the same cost and residual value. It changes earnings timing, not the hardware’s performance or the cash paid for it.

The network also affects the workloads a GPU can serve economically. Compare NVLink domains and cloud networking before treating the same GPU count as the same system.

Is keeping or buying it worth the cost?

Price one consistent unit: a GPU with its share of host costs, or a complete server. Enter your own assumptions. This calculates cost; it does not predict how long the hardware will survive.

Add measured workload throughput

Enter the costs and useful-life assumption to calculate a scenario.

Undiscounted cost is the starting capital value less assumed future sale value, plus operating costs. Buying uses the quoted purchase price; keeping uses today’s net sale value. Financing, taxes and unentered costs are excluded. Output-token cost uses the measured output rate, not a blended input/output API price. INR uses ₹94.29 per dollar, fetched 14 Nov 2026. These are user scenarios, not market prices or investment advice.

Help build better operating evidence

Report hardware that is still productive as well as units that failed or were retired. Owner reports stay separate from published research and are reviewed before any summary is released.

Take the GPU operating-life surveyTrack the listed companies behind the infrastructure