How long does AI hardware stay useful? An answer in three parts, because the accounting life, the software life and the model life differ.
Short answer. Treat three to six years as a starting point and judge it on three separate clocks. The accounting clock: India’s Companies Act, 2013 (Schedule II, Part C) lists 6 years for servers and networks and 3 years for end-user devices such as desktops and laptops. The software clock: Ollama’s support list still includes NVIDIA GPUs from compute capability 5.0. The model clock: Qwen says its 32B base model matches its earlier 72B one, so a given amount of memory keeps getting more capable, while context windows and model sizes also grow. Memory headroom matters more than raw speed. This is general information, not accounting advice.
01
Clock one: the accounting life
The Companies Act, 2013 says the useful life of an asset shall not ordinarily differ from the life in Part C of Schedule II; a company that uses a different life must disclose it and justify it with technical advice. It applies to companies under that Act, and other kinds of organisation follow other rules. Whether an AI appliance counts as a server or an end-user device is for your auditor to decide. This is a summary of the Act, general information and not accounting or tax advice.
Asset class in Schedule II, Part C
Useful life
Computers and data processing units: servers and networks
6 years
Computers and data processing units: end-user devices such as desktops and laptops
3 years
Office equipment
5 years
02
Clock two: software support
Runtimes keep older hardware working for a long time. Ollama’s documentation supports NVIDIA GPUs from compute capability 5.0 upward, a list that still names GTX 750-class cards. New model designs, however, need new runtime versions: Qwen3.8-27B mixes linear-attention and ordinary attention layers, and vLLM’s README names hybrid designs of that kind among the models it supports. Check that the runtime you intend to use supports your exact chip, and note that support for older chips is a statement about running, not about running quickly.
03
Clock three: models
In these examples, models became more capable per gigabyte and cheaper to run at long context. Qwen reports that its Qwen3 dense base models perform as well as larger Qwen2.5 ones, for example the 32B matching the previous 72B. The table shows the design effect: from our arithmetic on published configurations, the newest design needs a fraction of the cache of an older one. The counter-trend is length: Qwen3.8-27B offers a 262,144-token context and lists 1,000,000 as possible, which asks for memory if you use it.
Model
KV cache per token
Cache for 32,768 tokens (estimate)
Llama 3.x 70B
320 KiB
10.7 GB
Qwen3-32B
256 KiB
8.6 GB
Qwen3.8-27B
64 KiB (plus a small fixed state)
2.1 GB
04
How to buy for a longer life
Buy memory capacity first. It decides which models fit; speed is replaced by newer chips, and on compact machines capacity is usually what you cannot add later.
Prefer open interfaces. An OpenAI-compatible endpoint lets you move applications to a new box without rewriting them.
Check the runtime’s hardware list for the exact chip, and the vendor’s warranty and support term in writing.
Keep documents, indexes and logs portable and backed up, so replacing the box is a copy, not a rebuild.
On the software side, ANVE’s published position (see the security page) is that every order includes a perpetual licence for the software delivered with the unit, there is no remote kill switch, and the box keeps working if ANVEAI shuts down. ANVEAI installs updates on site today; signed update packages applied offline by USB are not on ANVE yet; ANVE does not update itself over the internet.
ANVE has not published a warranty or support term, because no sealed unit exists yet. Ask any vendor, ANVE included, for it in writing before you commit.
05
Questions people ask
Does an AI box need replacing every three years?
Not necessarily. Three years is the Companies Act figure for end-user devices, and six years is the figure for servers. Which applies, and how long you use it, is your decision with your auditor.
Will a box bought now run models released in 2030?
Nobody can promise that. Memory capacity is the first thing to check, because new models either fit in it or do not; software support for the chip is the second.
Does a GPU wear out from running AI all day?
This page found no published wear figures for these parts, so it gives none. Ask the vendor about continuous-use warranty terms and the cooling the design assumes.
06
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
