Choose local AI hardware
from the work backwards.

Start with five business questions

Hardware shopping goes wrong when a buyer starts with processor names. Start with the work instead:

Job

Is this short drafting, long-document review, cited search, transcription, image work, software development or a mix?

People

How many staff may use it in the busiest hour, not merely how many have accounts?

Material

How large are the documents and recordings, and how quickly will the approved collection grow?

Boundary

Must core work continue offline, and which web, mail, storage or business-system connections are allowed?

Service

Who will patch, back up, restore, support and eventually replace the machine?

Turn the answers into three examples: an ordinary job, the heaviest realistic job and an awkward edge case. Those examples become the sizing test and later the acceptance test. A benchmark from somebody else’s machine cannot answer whether your scanned report, long meeting or shared workload is usable.

Memory decides what fits

For local language models, memory capacity is usually the first constraint. The machine needs space for the model files being used, the working context, the operating system and the services around the model. Document search, transcription, vision and several active users can add pressure. A model that technically loads while leaving no working room is not a sensible business specification.

Memory bandwidth affects how quickly many local models produce text. Model format, compression and runtime matter too, so two machines with the same capacity can feel different. That is why a quote should name the exact model, format, engine and tested context rather than promise performance from “128 GB” alone.

Leave headroom for the agreed workload and routine system activity. Do not buy speculative capacity for a model nobody has tested, but do not size to the last available gigabyte either. Memory on many compact systems cannot be expanded later.

Storage holds more than documents

Storage must cover the operating system, model library, document index, uploaded files, recordings, exports, logs and safe working space for updates. Backups normally belong on a separate target; a second copy on the same internal drive is not protection from drive failure or loss of the machine.

Ask how much data exists now, how quickly it grows and which source remains authoritative. If the appliance indexes files from a document system, decide whether it stores copies, an index or both, and what deletion from the source should do. Retention is a business and legal decision, not a number the hardware vendor should silently invent.

Three practical platform paths

LocallyAI currently quotes three main compact-hardware paths. They run the same Harness, but operating system and model-engine support differ:

Product ranges and availability change. The current machines page records LocallyAI’s supported paths and links to official Apple, HP, Minisforum and NVIDIA specifications. The platform comparison covers the brand-level decision; this guide stays focused on sizing the business requirement.

A desktop appliance or a bigger server?

Many small offices do not need a rack server. A compact appliance can be quieter, easier to place and simpler to repair through a manufacturer’s normal channel. Bigger workstations or multiple appliances make sense when the tested job needs discrete GPU memory, more simultaneous work, separation between departments or a supported CUDA workflow.

More hardware also brings more heat, noise, power, network and support work. Before buying a high-powered workstation, confirm:

• the office has a suitable circuit, outlet and power-protection plan;
• the location has ventilation and is not accessible to visitors or casual staff;
• the network path is wired and sized for the staff and files using it;
• the manufacturer warranty and local repair route suit the business;
• there is a plan for work while the machine is unavailable.

A UPS can help with short interruptions and safe shutdown, but it is not a backup or a guarantee of service continuity. Have the business’s IT provider or electrician check site-specific power and network requirements.

Capacity is about the busiest hour

Ten registered users do not necessarily require ten times the hardware. What matters is how many heavy jobs overlap. A receptionist transcribing a meeting, two staff searching long files and another person using a larger reasoning model can be more demanding than twenty people making occasional short requests.

Ask the supplier to show what happens under the expected peak: Does work queue? Can a smaller “Fast” model handle everyday jobs while a larger model handles selected work? Do speech and document indexing compete with chat? When would a second appliance be simpler than one larger machine? The answers should shape the quote and the handover instructions.

Local, hosted or a deliberate mix

A hosted product puts more infrastructure responsibility with its provider, but the buyer still owns account configuration, suitable use and data decisions. A local system removes the external model service from the normal path, but the buyer now needs a machine, backup, physical security and support plan. The Australian Cyber Security Centre’s cloud shared-responsibility guidance for small and medium businesses is a useful prompt when comparing what a hosted provider carries and what remains with the customer.

Price the whole ownership period

Compare the same useful life and workload on both sides. For local hardware include the installed purchase, GST treatment, electricity, backup target, warranty or hardware cover, support, expected update work, replacement planning and paid connected services. For hosted software include every required seat, plan tier, usage allowance, separate transcription or search product, expected headcount and exit costs.

A local model does not create a per-message model bill, and LocallyAI does not charge a Harness fee for every login. Capacity still has a ceiling. A growing team may need a larger or second appliance, and custom integrations remain separately scoped. Use the worked comparison method rather than accepting a universal break-even claim.

Buying for a regional or interstate office

The same workload test applies nationally. Outside LocallyAI’s listed New South Wales on-site regions, the normal route is a machine configured and tested before shipping, followed by remote setup with a staff member or IT provider at the office. The local person handles power, sign-in and the network point; configuration and acceptance can be completed over a scheduled screen share.

For a remote delivery, put freight, insurance, manufacturer repair, spare-device expectations and support hours in writing. No supplier can make distance disappear, but the job can be designed so a hardware fault follows a known path rather than starting a search for the receipt.

Hardware questions for the quote

Exact configuration

Manufacturer, model, processor, memory, storage, operating system and warranty.

Exact AI stack

Model names and versions, file formats, runtimes, context limits and licences.

Evidence

Results from the ordinary, heavy and awkward acceptance jobs on this configuration.

Peak use

Expected simultaneous jobs, queuing behaviour and the point where another appliance is needed.

Site needs

Power, wired network, physical placement, ventilation and staff-device requirements.

Recovery

Backup target, restore owner, keys, test schedule and the plan during hardware repair.

Change

What can be upgraded, what is fixed at purchase and what gets re-tested after change.

Exit

Data export, configuration handover, secure disposal and replacement route.

Next: read how an Australian installation should run and what support and maintenance should cover.