GPU Depreciation Schedule and Monthly Cost at 3, 5, and 6 Years

A GPU depreciation schedule is a useful-life assumption that moves monthly hardware cost by a factor of two between three and six years. Operators now disclose schedules across that full range, which makes the assumption a decision rather than a convention. This page lays out the documented range with sources, works the straight-line math at each candidate life, names the assumptions that hide inside any schedule, and gives the workload rule that connects the schedule to the own-versus-managed decision.

What the Industry Now Assumes: The Documented Range

The documented range now spans three to six years, and the drivers behind it are nameable. At the short end, a three-year schedule reflects the historical accelerator cadence: a generation meaningfully ahead of its predecessor arrived often enough that three years captured economic reality. In the middle, one major cloud operator moved server depreciation from three years to four, and reports five-to-six-year assumptions elsewhere in its fleet. At the long end, a dedicated GPU cloud operator assigns six years to technology equipment, and operators extending from three-to-four years out to six are estimated to save roughly eighteen billion dollars annually in depreciation expense. The assumption driving the stretch is stability: current-generation GPUs serving inference or continued training remain economically productive longer than early accelerator generations did. The drivers that pull schedules back are workload obsolescence (model requirements outgrowing hardware) and maintenance tails on dense nodes, which is why independent analysis frames useful life as a range shaped by workload rather than a single defensible point.

Company schedules are public disclosures as of October 2026 and shift with reporting periods; the range is market context, not a recommendation.

The Math and the Assumptions Inside It

Straight-line depreciation is monthly cost equals capital amount minus salvage value, divided by useful life in months, and the reader should rerun it at each candidate life before choosing. Work the example with stated assumptions: an eight-GPU training node capitalized at 288,000 dollars (a 36,000 dollar per-GPU system cost is a realistic planning magnitude, not a quote) depreciates to a zero salvage at 4,000 dollars per month over six years, 4,800 dollars at five years, and 8,000 dollars at three years; the three-to-six-year choice alone moves the monthly figure by a factor of two. Three assumptions hide inside any schedule and deserve explicit values. Salvage value: assume near zero for accelerator hardware, because generation turnover historically collapses resale value rather than tapering it. Utilization: depreciation charges calendar time while value comes from productive hours, so a fleet at low utilization pays the same schedule for less output. Maintenance tail: power, cooling, and part costs rise as nodes age, so later years cost more per productive hour than the straight line shows. Tax treatment sits outside this math and belongs to the reader's finance function, including class-life rules for computers in their jurisdiction.

Rerun the example with the reader's own numbers: capital amount, installation, financing cost, and measured utilization all change the monthly figure, and the three-to-six-year spread stays a factor of two at any capitalization.

Choose by Workload Half-Life, Not Accounting Convenience

Set the schedule from workload half-life: the time until model or framework requirements make the hardware the constraint rather than the asset. Three conditions select the route. Own on a longer schedule when workload half-life clearly exceeds five years: stable inference fleets on current-generation GPUs, where a six-year schedule is honest and salvage risk is tolerable. Own on a shorter schedule when the team refreshes competitively every two to three years: accept the higher monthly figure, budget salvage near zero, and let the schedule tell the truth about refresh cost. Move to managed flat-rate capacity when workload half-life is short or genuinely uncertain: research-driven teams, fast-moving product AI, or mixed training-inference fleets where the next generation materially changes economics; predictable flat-rate monthly pricing under managed infrastructure converts depreciation and refresh risk into operating expense that scales with commitment, and the provider absorbs the hardware lifecycle. The disqualifying combination to avoid: a six-year schedule chosen to shrink monthly expense while the same team plans a competitive refresh in year three, which understates cost now and surprises the budget later. Managed private infrastructure delivered by providers like OneSource Cloud exists for exactly that mismatch, with single-tenant environments and zero egress fees keeping the operating-expense route clean.

The schedule decision and the ownership decision are the same decision made twice; teams that answer them inconsistently pay for the inconsistency at refresh time.

FAQ

Do GPUs really last six years in production?

Six years is a documented reporting assumption, not a universal physical reality. Operators serving stable inference on current generations report six-year lives, while honest life depends on workload half-life, utilization, and maintenance tails. The schedule should follow the workload, with salvage assumed near zero if the plan includes competitive refreshes.

How does OneSource Cloud pricing remove GPU depreciation risk?

OneSource Cloud delivers managed single-tenant GPU environments on predictable flat-rate monthly pricing with zero egress fees, so the customer books a stable operating expense while the provider owns the hardware lifecycle, including refresh, maintenance tails, and salvage risk. Capacity commitments scale with the team's actual horizon instead of a depreciation convention.

Summary: Let the Workload Set the Number

The documented range gives permission for anything from three to six years; the workload half-life gives the answer. Run the straight-line math at each candidate life, price the assumption traps explicitly, and treat a six-year schedule with a three-year refresh plan as the contradiction it is. For teams whose half-life argues against ownership, OneSource Cloud flat-rate managed capacity is the structural alternative.

To move GPU capacity from capital schedule to predictable monthly expense, see OneSource Cloud managed AI infrastructure.