AI Hardware for Small Business: Do You Need a New Computer?

No. For nearly every small business an ordinary laptop and a browser are enough, because ChatGPT, Claude, Gemini and Copilot do the heavy computing on the provider's servers. Special hardware, meaning a graphics card or a mini PC, matters only when you choose to run a model on your own machine.

That choice is rarer than the marketing suggests. People buy hardware to solve a privacy worry, then discover that a business plan already answers it, or they buy it to save subscription fees and find that the machine, the setup time and weaker answers cost more than $20 a month ever did. The test below tells you which side of that line your business sits on.

Follow me on Instagram@sagnikteaches

What your current laptop already does

When you type into a chat assistant, your computer sends the text over the internet, waits, and shows the reply. The effort sits elsewhere. A five-year-old office laptop that opens a dozen browser tabs without freezing is already a full AI workstation for drafting, summarising, translating, spreadsheet help and meeting notes.

Connect on LinkedInSagnik Bhattacharya

The same is true of the AI built into software you use: Copilot in Microsoft 365, Gemini in Google Workspace, the assistants inside accounting and booking tools. They run in the cloud and appear in the app. The hardware demand is the app's, which was already being met.

Subscribe on YouTube@sagnikteaches

Three things do make an old machine feel slow, and none of them is AI. Too many browser tabs, an almost-full disk and a browser that has not been restarted for weeks. Fix those first. If a laptop is genuinely too old to run the browser or the office suite you pay for, replace it because of that, not because of AI.

A 10-minute test for whether your computer is the bottleneck

1. Restart the machine and open only the browser and the one assistant you use.
2. Paste a long document (say ten pages) and ask for a summary.
3. Note whether the delay is a few seconds of waiting for the reply, which is the provider thinking, or the page itself stuttering when you scroll or type, which is your machine.

Waiting for the reply is normal and no hardware fixes it. Stuttering suggests the machine is short of memory or full. A music school's front-desk PC I'll use as an illustration: it froze when a booking system, a spreadsheet, email and the assistant were open together. Closing two of the four apps solved it. The cost of the fix was nothing.

Why anyone runs a model on their own machine

A local model is a downloadable set of weights, a large file of numbers, that a program such as Ollama or LM Studio runs on your hardware. Ollama installs on macOS, Windows and Linux; LM Studio runs on Apple Silicon Macs and on Windows and Linux machines. After the download, it works without sending text to anyone.

There are four honest reasons to want that.

• A rule or contract that forbids sending the data out. A client agreement, a funder's condition or a professional body's guidance may say that certain records may not go to a third-party service.
• No connection. A workshop, a van, a site office or a rural clinic with unreliable internet.
• High volume of one small job. Tidying thousands of product descriptions or classifying thousands of messages, where a per-use charge adds up and a modest model is good enough.
• Curiosity and learning. A fair reason, as long as you call it that and do not call it a business case.

Notice what is not on the list: "it is cheaper than a subscription" and "it is smarter". Neither holds for most small firms. The models you can run on a single office machine are smaller than the flagship models behind the paid plans, and they make more mistakes on long or subtle tasks. Treat them as a different tool, not a free version of the same one.

The tutorial on running AI on your own computers walks through the installation itself. The question here is narrower: whether any of the four reasons justifies spending money on equipment.

Memory is the number that decides what fits

Local models are limited by memory, not by clock speed. The model has to fit in the graphics card's memory (called VRAM) or, on a Mac, in the shared memory. A rough rule: a model compressed to 4 bits per weight takes about half a gigabyte per billion parameters, plus working space for the text you give it.

These are my own rule-of-thumb figures, not a vendor's: actual files differ by model and compression, and long documents add to the memory needed. Check the size shown on the model's download page before you commit. LM Studio's documentation recommends 16 GB or more of RAM on Mac and Windows, and says smaller models may still run on 8 GB Macs. Its Windows guidance suggests at least 4 GB of dedicated graphics memory for GPU use.

A quick sum for an optician who wants to tidy dictated notes. An 8-billion-parameter model needs about 8 × 0.5 = 4 GB, plus roughly 1 to 2 GB of working space, so a machine with 16 GB of memory can run it with room left for the rest of the practice software. That is a normal modern laptop, and an extra purchase may not be needed at all.

The sneaky case: "local" models that are not local

Some tools blur the line. In the Ollama app and command line, models tagged ":cloud" run in Ollama's cloud: they are never downloaded, and Ollama processes those prompts on its own side, not on your machine. If your reason for running locally is that nothing leaves the building, check that the model you chose does not carry that suffix, and read what the tool says about where processing happens before you paste anything sensitive.

What the options cost

Here are four set-ups with USD figures. Cloud plan prices come from the vendors' pages; hardware prices are list prices at the time of checking and move often, so confirm them before you plan.

The graphics card is the visible part of the bill. The full cost includes the computer around it, a power supply that can feed it, setup time, and someone to update and fix it. Those hours are real even when no invoice arrives.

Worked comparison: four cloud seats against one desk machine

An illustration for a care agency with four office staff who write visit summaries, replies to families and rota notes, and who would like nothing identifiable to leave the building. Compare three years of each route.

On paper the local route is $1,260 cheaper. Read the table properly, though. The cloud plan gives the staff a flagship model; the Mac gives them a smaller one, with weaker summaries and more mistakes to catch. If the weaker model forces an extra 10 minutes of checking per day per person, four people lose roughly 40 minutes a day. At $25 an hour that is about $17 a day, or over $4,000 across 250 working days a year. That wipes out the saving in the first year.

The local route wins only when privacy is a hard requirement. If the agency's concern is that staff might paste identifiable details into a free personal account, the cheaper fix is a business plan, a rule about names (strip them before pasting) and a written list of what is allowed. The guide to keeping customer data private when staff use AI sets out those rules.

Privacy: what local does and does not buy you

Running a model locally removes one recipient: the provider. Several risks remain, and they are the ones most often forgotten.

• The machine itself. A lost or stolen laptop with documents and chat history on it leaks the same material. Turn on full-disk encryption and a screen lock.
• Backups. If the folder holding model outputs is copied to a cloud backup, the text has left the building anyway.
• Shared access. A single machine used by four people with one login gives you no record of who saw what.
• Browser extensions and "helpful" add-ons that quietly send text elsewhere.

Cloud, in turn, is not the careless option it is often painted. Business plans of ChatGPT, Claude, Microsoft 365 Copilot and Gemini in Workspace do not train on your business content by default, and offer admin controls and logs a lone machine lacks. For many firms that is stronger protection than a mini PC under a desk, because it comes with someone else maintaining the security. Whether it satisfies a particular client contract is a question for the contract, or for your adviser. The comparison of cloud and on-device AI for business data sets out which risks each route removes and which it leaves.

Four small businesses, four answers

A veterinary practice

Staff want help writing reminders, discharge notes and replies to owners. The text contains animal names and owners' contact details. Answer: a business plan, with a rule to leave out phone numbers and addresses. A local model adds cost without solving anything the rule does not. Buy nothing.

An optician

The practice wants tidier dictated notes and recall messages. Answer: try the 8-billion-parameter model on an existing 16 GB laptop for a month. If it manages the notes, you have local privacy for free. If not, fall back to a business plan.

A music school

Teachers write lesson summaries for parents, and the office handles invoices. Answer: a team plan for the teachers who use it; no hardware. The front-desk PC needs fewer open apps, not a new graphics card.

A care agency

A funder's contract says client records must not be processed by a third-party service. Answer: this is the real case for a local model. Ask whether the contract wording covers a business plan with a data-processing agreement; if not, use a machine with at least 16 GB of memory, encrypt it, and keep the model to narrow jobs such as rewording non-identifying text.

Buying the wrong thing: a gaming PC for email drafts

One failure to avoid is the proud purchase. An owner reads that a graphics card runs AI "for free", buys a $2,500 desktop with the biggest card on offer, spends a weekend installing a model, and then uses it twice a week to rewrite emails. A $20 plan would have done the same job better, with no maintenance.

Warning signs that you are heading there: the purchase comes before a written list of jobs; the benchmark you are quoting is speed, not accuracy on your own documents; the word "future-proof" appears in your reasoning; and nobody in the business has the time or inclination to update the software.

The cure is a short, dull checklist.

1. List the three jobs you want AI to do and name the documents involved.
2. For each, write down whether any rule or contract stops it going to a cloud provider.
3. Run all three on a business plan for four weeks, with personal details stripped out.
4. Buy hardware only for a job that failed step 2 or step 3, and size the machine to that job.

A decision rule you can apply today

Notice the row about thousands of small jobs. Pay-per-use pricing is quoted per million tokens (small pieces of words). Anthropic lists its smallest current model, Haiku, at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens; check the current model and price on its pricing page before you rely on it. A quick illustrative sum for a catalogue of 2,000 product descriptions, assuming about 300 tokens in and 200 tokens out per item: 600,000 input tokens cost about $0.06 and 400,000 output tokens about $0.20, so the whole run costs well under a dollar. Even a run ten times larger stays under $3. A hardware purchase of $900 or more needs an enormous volume to pay back against that, so price the run before buying anything. How to price it is covered in the free versus paid tools comparison, which shows where free tiers stop.

If you want a second pair of eyes on what you actually need, that is part of what my AI implementation consultation covers; see the contact page.

If you do buy: what to look at

Suppose the test says a local model is justified. Choose with these points in mind.

• Memory first. On a Mac, pick the memory option for the model size you need; it is normally fixed at purchase. On a PC, graphics memory sets the ceiling: 16 GB cards handle the 13 to 14 billion class, whereas 8 GB cards limit you to the smaller ones.
• Software support. Ollama supports recent NVIDIA cards (it states a minimum compute capability and driver version), AMD cards through its ROCm drivers, and Apple chips via Metal. Check your exact card against its documentation before buying.
• One owner. Name the person who updates the software, applies security patches and answers "it's slow" messages. If no one volunteers, that is your answer.
• A fallback. Keep a cloud plan for the jobs the local model fumbles. Most firms that run a local model end up with both.

Whichever you choose, test with your own documents, not a vendor's demo. Give the model ten real examples of the job, mark the output against what you would have written, and count the corrections. The count tells you more than any benchmark.

If you are still deciding which AI jobs are worth doing in the first place, it can help to score your own skills first with the five-skill AI fluency self-check, because a good habit of checking output matters more with a smaller local model than with a flagship one.

Hardware questions owners ask before buying

Do I need an AI PC or a Copilot+ laptop?

No tool named in this tutorial needs one. Cloud assistants run on the provider's servers and your machine only displays the result, so any computer that runs a current browser smoothly will do. A special chip may speed up a few on-device features, but it is a nice extra, not a requirement for business use.

Is a local model more private than ChatGPT or Claude?

Your text stays on your machine, which removes the provider as a recipient. It does not make the machine itself safe: a lost laptop, an unencrypted backup or a shared login can leak the same documents. Business plans from the big providers also do not train on your content by default.

Can I run a model on an old laptop?

You can try a small model, and it will probably be slow and give weaker answers. Treat it as a free experiment, not a plan. If it handles the one narrow job you have in mind, such as tidying notes, you have your answer without spending anything.

How much memory does a local model need?

As a rule of thumb, a model compressed to 4 bits needs about half a gigabyte per billion parameters, plus working space. An 8-billion-parameter model therefore wants roughly 5 to 6 GB free; a 14-billion one about 9 to 10 GB. More memory allows longer documents.

Should I buy hardware or pay for a business plan?

Start with the plan. Four seats on a business plan cost about $80 to $100 a month, and you can cancel. Hardware is a commitment, so buy it only after a trial shows a job the cloud cannot take, because of privacy rules, a contract or a poor connection.

Further reads

• The AI Tool Stack a One-Person Consultancy Actually Needs — A stack that a one-person firm can run on a laptop.
• Annual or Monthly Billing for AI Tools: Which Saves You More? — Work out which billing option saves more on cloud plans.
• Automation in Clothing Manufacturing: What a Small Workshop Can Automate — A workshop where automation, not AI hardware, is the question.
• Automated Managed Service Provider: Hardware Refresh Plans From Asset Lists — Plan a hardware refresh from an asset list.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: NVIDIA RTX 5060 family and RTX 5090 pages (starting prices and memory), Apple Mac mini buy and specs pages, Ollama download, GPU and cloud documentation, LM Studio system requirements, ChatGPT Business pricing page, Claude pricing page (plans and API prices), Google AI subscriptions page, Microsoft and Google Workspace privacy pages.