Every “local AI setup” article ends with a shopping list. This one starts with the opposite question: what is the least you can spend to do the work you actually have?
I costed four realistic options against the jobs people ask about — chat, summarising, transcription, image generation and small-scale training.
The short answer
Chat, summaries, transcription: the laptop you own, plus a $40 RAM stick if possible. Faster CPU inference and always-on: a used 16 GB mini PC, about $150. High quality, bursty use: a cloud API at $5–20/month. Images and training: free cloud GPU, not local hardware.
The four options
| Option | Upfront | Monthly | CPU inference | GPU work | Best for |
|---|---|---|---|---|---|
| Existing 8 GB laptop | $0 | $0 | 1.7B–4B | No | Privacy, offline, small tasks |
| Laptop + 16 GB RAM | $35–60 | $0 | 3B–7B, fragile | No | Removing the paging cliff |
| Used 16 GB mini PC | $120–220 | ~$2 power | 7B at 2–6 tok/s | No | Always-on, faster CPU |
| Cloud API | $0 | $5–20 | n/a (better quality) | Limited | Bursty, high-quality work |
| Free cloud GPU | $0 | $0 | n/a | Yes, with limits | Images, training |
Prices are U.S. street prices in November 2026. Power cost assumes 15–35 W average and $0.15/kWh.
Option 1: keep the laptop
The correct starting point. It costs nothing and it is enough for more than people expect:
- 1.7B–4B language models at 4–18 tokens per second
- Whisper base transcription at ~0.7× real time
- Python, web development, writing
- Free cloud GPU for anything GPU-shaped
The limits are memory and thermals, and they are real. If you find yourself closing the browser to run a model, this option has reached its ceiling.
Option 2: add RAM
The cheapest meaningful upgrade. About $40 turns a fragile 7B setup into a comfortable one, and dual-channel memory adds 5–15% to generation speed.
It does not make the CPU faster, and it cannot fix soldered RAM. The full upgrade guide covers compatibility and installation.
Verdict: the first thing to buy, if your laptop allows it.
Option 3: a used mini PC
This is the value pick that most guides ignore. Off-lease business mini PCs — Lenovo Tiny, Dell Micro, HP EliteDesk Mini — are widely available with:
- 8th–10th gen Intel or Ryzen CPUs
- 16 GB RAM (often upgradable to 32 or 64)
- NVMe storage
- a desktop power limit, not 15 W
A representative used unit costs $120–220. It will not match a modern laptop on single-core speed, but it sustains load without throttling and can run 24/7.
Measured expectation: roughly 1.5–2× the tokens per second of this laptop on the same 7B model, and a much better experience over long sessions. Still no GPU, so no Stable Diffusion and no training.
What to check before buying used
- CPU generation: 8th gen Intel or newer; older chips lack instruction sets and efficiency.
- RAM slots: at least one free, ideally two.
- Storage: NVMe slot, not just eMMC.
- Power supply included. Proprietary adapters are annoying to replace.
- Return policy. Off-lease stock varies; buy from a seller that accepts returns.
Option 4: pay for cloud
A $5–20 monthly API budget buys more raw capability than any $200 local machine. For chat and summarization, the quality difference between a 3B local model and a frontier cloud model is large — and the cloud model answers in seconds.
The trade-offs:
- Privacy: your text leaves the machine.
- Connectivity: no offline use.
- Cost predictability: usage-based pricing can surprise you.
- Latency: usually fine, occasionally not.
The honest framing: local AI is not cheaper than cloud AI at low volume. It is cheaper at high volume, it works offline, and it keeps data on your machine. If none of those matter, the cloud is the better tool.
Cost per realistic job
| Job | Cheapest path | Cost |
|---|---|---|
| 20 chat questions a day | Local 1.7B | $0 |
| Summarise a 10-page PDF weekly | Local 3B | $0 |
| Transcribe 5 hours of interviews/month | Local Whisper base | $0 (about 3.5 h of machine time) |
| Generate 100 images | Free Colab/Kaggle | $0, with queues |
| Fine-tune a small model | Kaggle free GPU | $0, within quota |
| Reliable 7B chat all day | Used mini PC | ~$150 once |
| Production-quality writing | Cloud API | $5–20/month |
The pattern is consistent: local for repetitive, private, small-model work; cloud for bursty, GPU-shaped or quality-critical work.
What I would buy, by situation
Student, 8 GB laptop, tight budget. Buy nothing. Run 1.7B locally, use free Colab for GPU work, use a free cloud chat tier for hard questions. Add RAM only if a slot exists and you keep hitting the swap cliff.
Developer with soldered 8 GB. Skip the RAM. A used 16 GB mini PC at ~$150 is the better spend, and it doubles as a home server. Keep the laptop for portable work.
Writer or researcher. RAM first, then a good microphone. Transcription and local drafting benefit more from memory and audio quality than from CPU speed.
Anyone who needs image generation. Do not buy hardware. Use free cloud GPU tiers until they block a project you care about, then rent GPU time by the hour.
Anyone who needs reliability. Pay for cloud. Free tiers disconnect; local hardware throttles. A $10–20/month API is cheaper than the time lost to workarounds.
The mistake to avoid
The expensive mistake is buying a new laptop “for AI” when the actual bottleneck is 8 GB of RAM and a browser with 30 tabs. The second most expensive mistake is buying a GPU for a workload that runs twice a month.
Measure the job first. The benchmark table and the field reports exist to make that measurement cheaper than the purchase.
FAQ
Is a mini PC better than a laptop for local AI?
For sustained inference and always-on use, yes. For portability and battery life, no. Many people end up with both: a laptop for work, a mini PC as a local server.
Can I add a GPU to a mini PC?
Some models support a low-profile GPU or an eGPU enclosure, but the power supply and chassis often limit it. At that point, a used desktop is a better value.
What about Apple Silicon?
An M-series Mac with unified memory is excellent for local AI, but the entry price is far above $200 and the memory is not upgradeable. It is a different budget tier.
Is a Raspberry Pi enough?
For tiny models and transcription, maybe. For anything interactive, no. The CPU is far slower than this laptop.
Related reports
- 8 GB vs 16 GB: the upgrade guide
- Free cloud GPU limits
- Best small LLMs for 8 GB
- Gear that earns its place