AI hardware refresh cycle: how often should enterprises upgrade GPU servers?
 |
Simon Harrsen - EVP North America
Published November 2, 2026 · 6 min read
|
NVIDIA now releases new data center GPUs about every year. Most finance teams still depreciate a server over five or six years. That gap is the central problem in any AI hardware refresh cycle.
Refresh too early, and you write off equipment that still works. Refresh too late, and you run AI workloads on hardware that costs more per result than a newer system would. This article explains how fast AI hardware ages, how long it actually lasts, and how enterprises can plan refresh timing around economics instead of the calendar.
- An AI hardware refresh cycle is the planned interval for replacing GPU servers, and it runs shorter than the traditional five to six year server cycle.
- NVIDIA launches a new architecture about every two years, with an upgraded version in between, so a new product arrives roughly every year.
- Physical GPU lifespan often runs six years or more, but economic life, the period when hardware is still the cheapest way to do the work, is usually shorter.
- CHG-MERIDIAN customers are choosing shorter terms for AI hardware, and most return and replace the equipment at the end of the term.
- Refresh decisions work best when tied to triggers such as cost per result, power limits and resale value, not a fixed date.
What is an AI hardware refresh cycle?
An AI hardware refresh cycle is the planned interval at which an organization replaces its GPU servers and related AI equipment. It sets when hardware leaves service, when the next generation arrives and how the change is funded.
Traditional server refresh has moved slowly, and mostly in one direction. Amazon extended the useful life of its servers from four years to five in 2022, and then to six years in 2024. A five to six year hardware refresh cycle became the norm for general purpose servers.
AI hardware breaks that pattern. In 2025, Amazon shortened the useful life of a subset of its servers from six years back to five. The company pointed to the faster pace of technology development, particularly in artificial intelligence and machine learning, according to Amazon's 2024 annual report.
When one of the largest cloud providers in the world shortens its schedule, the signal is clear. AI equipment ages faster than the servers that came before it.
The difference matters for budgets. A refresh cycle decides when capital is spent, how much value is recovered from old equipment and how long a workload runs on hardware that is falling behind.
How fast does NVIDIA release new AI hardware?
NVIDIA releases a new major GPU architecture about every two years. Between those launches, it ships an upgraded version of the current architecture. The result is a new data center product roughly every year.
The recent roadmap shows the pace. Hopper arrived in 2022. Blackwell followed in 2024, and Blackwell Ultra in 2025. In May 2026, NVIDIA announced it was ramping Vera Rubin into full production, with shipments set to begin in the fall.
| Year | Generation | Type of release |
| 2022 | Hopper (H100) | New architecture |
| 2024 | Blackwell (B200) | New architecture |
| 2025 | Blackwell Ultra (B300) | Upgrade |
| 2026 | Vera Rubin | New architecture, shipping from fall |
| 2027 | Rubin Ultra | Planned upgrade |
| 2028 | Feynman | Planned architecture |
Years reflect launches and NVIDIA's published roadmap. Volume shipments often trail announcements by several months.
Each step brings large gains. NVIDIA says Vera Rubin delivers 10 times the agent throughput of the previous Grace Blackwell platform. Results on real enterprise workloads will vary, but the direction is consistent. Each generation lowers the cost of producing the same AI output.
For a buyer, this creates a moving target. A GPU server bought this year will see at least two newer generations launch before a five year schedule ends.
How long do data center GPUs last?
A data center GPU has three different lifespans. They rarely match, and mixing them up leads to poor refresh decisions.
Physical life is how long the hardware keeps working. With good cooling, it is often six years or more. Amin Vahdat, who leads AI and infrastructure at Google, has said that Google's seven and eight year old TPUs, its own AI chips, are still at full utilization. AI hardware can keep running long after it stops being the newest option.
Accounting life is the period over which a company depreciates the asset. Depreciation spreads the purchase price over that period as an expense. For servers, the common range is five to six years. Those GPU depreciation schedules can drift far from what the hardware is actually worth.
Economic life is how long the hardware stays the cheapest way to do the work. It is usually the shortest of the three. Once a newer generation produces the same output at a much lower cost per result, older hardware becomes expensive to keep, even if it works perfectly.
For most enterprises, economic life is the GPU lifespan that should drive a refresh plan.
Why are enterprises choosing shorter terms for AI hardware?
Enterprises are choosing shorter terms for AI hardware because its value falls faster than the value of earlier server generations. CHG-MERIDIAN sees this in its own customer base. Customers are picking shorter lease terms for GPU servers.
What happens at the end of the term is just as telling. Most customers return the hardware and replace it with a current generation. Some extend the lease for a period instead.
A five to six year useful life is still realistic for a GPU server. The hardware does not wear out at year three. What changes is its value. Estimates from CHG-MERIDIAN's international asset risk management team show a GPU server losing well over half of its value within two years. That is far faster than a six year book schedule assumes. These are estimates in a young market, and they will change as more resale history builds up.
A shorter term lets the customer hand the equipment back while it still has meaningful resale value. The next generation then arrives on a planned date, not after an emergency budget request.
How should you plan an AI hardware refresh cycle?
A strong AI hardware refresh cycle plans for replacement before the equipment is even bought. Four practices make that easier.
Match the term to the workload
Training large AI models pushes hardware to its limits, so training clusters feel the pressure of each new generation first. Inference, which runs trained models to answer requests, is usually less demanding. Inference hardware can often stay in service longer. Setting a different refresh interval for each workload avoids replacing everything on one schedule.
Cascade older hardware to lighter work
Retired training servers do not have to leave the data center on day one. Older servers can move to inference, testing or internal tools. This stretches the value of the original investment while the newest hardware handles the heaviest jobs.
Refresh on triggers, not dates
A calendar date is a weak reason to replace hardware. Stronger triggers include a newer system that cuts the cost per result, power or cooling limits in the data center, the end of firmware and software support, and a closing window for resale value. A sound refresh strategy reviews these triggers at least once a year, and again whenever a new generation ships.
Budget for the next generation now
Every refresh plan needs a price for the replacement. AI servers are expensive, and prices move with supply and demand. Total deployed AI server costs include both the purchase price and end of life costs. Building that figure into the plan early keeps the next server refresh from becoming a surprise capital request.
How does leasing fit an AI hardware refresh cycle?
Leasing fits an AI hardware refresh cycle because it ties the term to economic life instead of physical life. The lease ends when the hardware is likely to stop being the best value, not when it stops working.
Leasing also moves the resale risk. In a fair market value lease, the lessor sets payments based on its own estimate of what the equipment will be worth at the end of the term. If GPU resale prices fall faster than expected, the lessor carries that risk instead of the enterprise.
This structure suits hardware whose future market value is hard to predict. The enterprise pays for the years of use it actually needs, not for the full physical life of the equipment.
At the end of the term, the customer has three options. The customer can return the equipment, extend the lease or upgrade to a current generation. Returned equipment is wiped, tested and resold.
For enterprises running AI on premises, server and data center leasing turns each refresh into a planned event. To compare leasing with buying and cloud, see our overview of AI infrastructure financing.
Talk to us and discover how our expertise can benefit your business.
Frequently asked questions
Most enterprises plan to refresh AI hardware more often than general purpose servers. A GPU server can stay useful for five to six years, but newer generations arrive about every year. Many organizations now choose shorter lease terms and replace hardware when a newer system lowers the cost per result, rather than on a fixed date.
An enterprise data center GPU can physically run for six years or more with good cooling. Its economic life is usually shorter, because each new generation produces the same AI output at a lower cost. Most companies depreciate GPU servers over five to six years.
A GPU server typically has a useful life of five to six years, which matches common depreciation schedules for servers. Its resale value falls much faster, though. That is why many enterprises return or replace GPU servers well before the hardware stops working.
General purpose data center servers are commonly kept for five to six years. Amazon, for example, moved its server useful life to six years in 2024, then cut a subset back to five years in 2025, citing the faster pace of artificial intelligence and machine learning development.
Roughly, yes. NVIDIA launches a new major architecture about every two years and an upgraded version in between. Hopper arrived in 2022, Blackwell in 2024 and Blackwell Ultra in 2025. Vera Rubin begins shipping in fall 2026, with Rubin Ultra and Feynman on the roadmap for 2027 and 2028.
Skipping a generation can make sense when current hardware still meets workload needs and fits power limits. The decision should rest on cost per result, not on the release calendar. A newer system that cuts the cost of each output enough can justify the switch, while a small gain usually cannot.
Yes. Inference, which runs trained models to answer requests, is usually less demanding than training. Older GPU servers retired from training can often move to inference, testing or internal tools. This cascade stretches the value of the original investment.
Common triggers include a newer system that lowers the cost per result, power or cooling limits in the data center, the end of firmware and software support, and a closing window for resale value. Reviewing these triggers every year, and at each new GPU launch, keeps refresh decisions grounded in economics.
After a refresh, GPU servers are usually moved to lighter workloads, resold or recycled. Before resale, the hardware is wiped and tested. Under a lease, the equipment goes back to the lessor, which handles remarketing and carries the resale risk.
Leasing is often the better fit when refresh cycles are short. A lease ties the term to the hardware's economic life and shifts resale risk to the lessor. Buying can still work for organizations that plan to run hardware for its full physical life. The right choice depends on workload, budget and accounting goals.