Run AI on your own hardware. Maybe hardware you already own.

Private models for businesses that can't send their data out, or are tired of paying per question. We size it, build it, tune it, and stay with you while it runs.

49 → 62.5 tokens a second, same card

17% → 83% recall after one prompt rule

5-7 → 29-48 tokens a second, 125B model

Why bother.

The data can't leave. Client files, health records, anything a privacy officer worries about.

The bill keeps growing. Every question to a cloud AI costs money, every day, forever.

They want to own it. No terms of service that change under them.

And one honest reason not to: the best cloud models are still better at hard reasoning than anything you can run locally. We'll tell you which parts of your work are fine locally and which aren't. Often the answer is both.

Three ways in.

Give old hardware a second life.

That workstation in the comms room, the server that got mothballed when you moved to the cloud. A lot of it can run a useful model with the right card and a bit of work.

Our own AI server is a 2018 Dell Precision workstation. It wouldn't see its new 24 GB graphics card at all: the BIOS gave the card too small a window of memory, and Dell offers no setting to change it. We found the hidden firmware value that fixes it, about ten minutes at the machine, and published the fix so others can use it.

Then the tuning. Picking the right quantised model for the card moved one model from 49 to 62.5 tokens a second on the same hardware. On Intel Arc we wrote our own model server to get multiple users served at once: about 60 tokens a second for one person, about 117 across three. (Our own box, October 2026.)

49 → 62.5

Local AI on a 2018 workstation: the firmware fix, the 24 GB card, and the measured recall and speed changes

17% → 83%

A fresh build, sized to your real work.

Not to a benchmark. We capture your actual requests, measure how much context they really need, and test the exact card class on rented hardware for about a dollar a run before you buy anything.

That's how our own evaluation went. It cancelled a planned graphics card purchase when the model on it went silent on exactly the cases that mattered, a quarter of the time. And it found that the real problem wasn't hardware at all: one rule the code needed was missing from the prompt. Fixed, recall went from 17% to 83%.

Hosted in Australia, without the hardware.

When the data has to stay onshore but you don't want a box in the building, we run your models on sovereign AI infrastructure in Australia, through ResetData.

And when a job needs a lot of GPU for a short time, like converting a model or processing a backlog, we rent it by the hour, do the work in one batch, and shut it down. The cost of rented GPUs is rarely the hourly rate. It's the machine left running after the job finished. Ours stop themselves.

What we actually do.

Size it: your workload, your context lengths, your power and slots, your budget.

Test models on your real cases, not leaderboards. Score what matters, like the incidents caught, not whether the output was valid JSON.

Choose and tune the model for the card: the right quantisation, the right serving engine, memory kept inside the safe line.

Fix what's in the way: drivers, firmware settings, the bug in someone else's code. One of ours is now part of llama.cpp, the engine much of local AI runs on.

Give it an address your tools already speak (the same API shape as OpenAI), so existing software just points at it.

Lock it down: private network by default, keys per person, sign-in for the chat and coding tools.

Stay with you. Models move fast. We keep yours current and tell you when something better fits.

What we run ourselves.

A 24 GB Intel Arc Pro card in a 2018 workstation, serving chat, coding and tool-using agents to a small group by sign-in.

A 125-billion-parameter coding model on that one card. The engine is someone else's open-source project; our part was getting it running on Intel Arc and taking it from 5 to 7 tokens a second to 29 to 48.

A private assistant on a local model.

Research pipelines that read hundreds of government PDF reports with a local model, at no cost per page.

Got hardware sitting idle, or a data problem that rules out the cloud?

Tell us what you've got. We'll tell you what it can run.

Talk to us about local AI