How Long AI Hardware Stays Useful
The most common objection to buying an AI appliance is a worry about AI hardware lifespan: won't a machine costing tens of thousands of dollars be obsolete in eighteen months? It is a fair question from anyone who has watched this field move, and it deserves an honest answer rather than a reassuring one.
The honest answer is that the box stays useful for as long as it does the job you bought it for, and the evidence says that is years. The models it runs keep improving while the silicon stands still, which is the opposite of how obsolescence usually works.
Where the Fear Comes From
The eighteen-month fear comes from watching frontier-model headlines, where each release makes the last one look dated within a quarter. That pace is real, but it describes the race at the top of the field, not the tool in your office.
An office appliance is not competing in that race. It is judged by whether the briefing arrives, the contract gets found, and the draft gets written -- whether the work happens, not whether something faster exists somewhere.
The Job Defines Useful
A machine is obsolete when it can no longer do its job, not when a better machine exists. Every other asset in your office already lives by this standard; nobody retires the phone system because a newer model shipped.
The question to carry through the rest of this article is therefore not "will something better appear?" Something better will always appear. The question is whether this box keeps doing this job.
Office Work Has Plateaued
The work an office system does -- retrieval over your own documents, drafting, briefings, routine workflows -- has settled into good-enough territory for models that fit comfortably on today's hardware. Pulling the right clause out of last year's contracts does not need a frontier model; it needs a competent one with access to the contracts.
The tasks that genuinely demand frontier ability are a different category of work. Deciding which tasks those are is a routing decision, not a reason to avoid owning the hardware that handles everything else.
Software Extends the Hardware
Model efficiency keeps improving fast enough that the same silicon runs meaningfully better models each year. Epoch AI's analysis of algorithmic progress found the compute needed to reach a given level of language-model performance has halved roughly every eight months.1
The effect shows up concretely in small models overtaking old giants. Stanford's AI Index reports that a 3.8-billion-parameter model in 2024 matched a benchmark threshold that had required a 540-billion-parameter model two years earlier, and that the cost of GPT-3.5-level inference fell more than 280-fold in about eighteen months.2
For an owned box, that trend is obsolescence in reverse: the hardware you bought keeps inheriting better software. FactoryOS is built around exactly this -- models stay pinned until you choose to upgrade, and capability improvements arrive as software on the same machine, on your schedule rather than a vendor's.
Refresh Norms Run Five Years
Enterprises already plan on server lives your CFO would recognize: Uptime Institute's global survey found most businesses refresh their servers every three to five years, and refresh cycles have been lengthening, not shortening.3 Depreciation schedules assume the same, spreading the purchase over a five-year useful life.
That norm holds for hardware that does not improve with age. A box whose models get better every year clears the bar more easily, and the cost comparison breaks even well inside that window.
Capacity Retires Hardware First
When owned AI hardware finally does get replaced, the trigger is almost always capacity -- more users, heavier workloads, bigger ambitions -- not failure and not dated abilities. That is an upgrade decision made from success, and it is the kind of growth you can budget for rather than a loss you absorb.
The old box keeps a role after the upgrade: overnight batch work, a second workload, a test system. It gets demoted, not discarded, which is what residual value looks like in practice.
The Cloud Ages Too
The alternative does not escape aging; it hides it. In the cloud you never own anything that grows old, but you never stop paying, and the models your workflows depend on get retired on the vendor's schedule, not yours.
A deprecation notice is obsolescence you do not control, arriving as a migration project with a deadline. On hardware you own, the model you validated keeps running until you decide otherwise.
Eighteen months from now, an owned box will be running better models than it does today. The real question is whether your cloud vendor can promise your workflows the same stability.