Nothing leaves the country
Inference, storage, logs and backups all sit in New Zealand. No cross-border replication, no offshore failover region quietly holding a copy of your prompt.
AI models on
New Zealand metal.
Nothing leaves the country.
AI that runs on New Zealand hardware — for the agencies and enterprises whose data isn't allowed to leave the country.
No offshore hop. No retention. Never trained on your data.
your prompt→
nz endpoint→
open model→
your answer
One hop. It never leaves.
Your request hits a New Zealand endpoint, a model running on New Zealand hardware answers it, and that's the whole path. There's no upstream provider behind us to forward it to.
Open weights, not a black box
Open-weight models you can inspect, pin to a version, and take with you. No vendor's roadmap gets to move underneath you.
Evidence, not assurances
Every request writes an append-only audit record — identity, time, model, tokens. Content isn't logged unless you turn it on. Export it straight into your own SIEM.
The same question, two journeys
If you've ever pasted something into ChatGPT, you've already done the thing this page is about. Here's the only part that changes.
Your words go overseas
- A computer in another country reads it.
- That country's laws apply to it while it's there.
- What happens next is a policy you can't inspect.
Your words stay here
- A computer in New Zealand reads it.
- New Zealand law is the only law that touches it.
- It's deleted the moment the answer comes back.
What "local" actually means
It's a slippery word, so here's the literal version: the physical machine that reads your question is in New Zealand, and your question isn't copied anywhere else on the way to it.
Not an overseas service with a New Zealand login page.
Not a New Zealand copy that quietly syncs to a bigger one offshore.
It's a machine you could drive to.
Drafting
Summarisation
Extraction
Classification
Transcription
Embeddings
Your fine-tune
The open models are good enough for the work that's actually queued
Summarising case files, transcribing interviews, drafting OIA responses, searching internal policy. None of it needs a frontier model. All of it needs the data to stay put.
Built for the risk assessment, not around it
Designed against NZISM and the Protective Security Requirements, and built to support your Privacy Act 2020 and Public Records Act obligations. We bring the architecture and the evidence; your assessor keeps the judgement.
Zero kilometres offshore
Every call to a public AI API is a cross-border disclosure you have to justify under IPP 12 — every time, for every dataset. Wooly removes the question rather than answering it.
Offshore API
~12,000 km
Wooly
0 km
✦ Classification-aware
Per-workspace handling rules decide which tenancy a workload is allowed to touch, so IN-CONFIDENCE and SENSITIVE material only ever routes where you've approved it.
Drop-in API
Standard endpoints for chat, embeddings and transcription. Change a base URL and a key — the rest of your stack doesn't move.
point at the NZ endpoint
swap the key
everything else unchanged
Zero retention
Prompts and completions live in memory for the life of the request, then they're gone. Turn logging on only where your own record-keeping needs it.
prompt — discarded
completion — discarded
metadata — audited
Bring your own weights
Run a model you've chosen, fine-tuned, or been given — including checkpoints you can't publish.
your-finetune-v3.safetensors
served on your tenancy
weights stay yours
Two shapes, one API
Start on shared, move to dedicated when the workload needs its own hardware — without rewriting anything.
Shared NZ cloud
- Multi-tenant New Zealand infrastructure
- Per-token pricing, no commitment
- Zero retention by default
- Intended for UNCLASSIFIED and IN-CONFIDENCE
Dedicated tenancy
- Your own GPUs in an NZ facility
- No other workload shares your accelerator
- Private link or IPsec into your network
- Intended for SENSITIVE material
Ask us what's certified, what's in progress, and what isn't. You'll get it in writing before you commit — we won't imply an accreditation we don't hold.
What Wooly won't do
Worth stating plainly, because it's usually the second question in the room.
No facial recognition or biometric identification. We host language, speech-to-text and embedding models. Identifying people from images isn't a product we offer.
No predictive policing or risk scoring of individuals. The models here summarise, extract, transcribe and search. They don't rank people.
No training on your data. Not for our models, not for anyone else's, not in aggregate, not de-identified.
No silent model swaps. The version you pinned is the version that serves you until you change it. Updates get announced, not rolled out under the same name.
No unattended decisions. Output is drafted for a human to check. We'll help you build the review step; we won't help you remove it.
SSO / SAML
SCIM
RBAC
SIEM export
IPsec
Private link
Fits the stack you already run
Your existing identity provider, your existing logging, your existing network boundary. The endpoint can be made unreachable from the public internet entirely.
1 hop
from your application to the model — there is no upstream provider behind us.
0 bytes
of your data used to train, fine-tune or evaluate any model. Contractually.
Open weights
inspectable and swappable — pin a version and it stays pinned until you change it.
Same day
to point an existing application at a New Zealand endpoint.
Keep it onshore.
The workload that failed the risk assessment is the one this was built for. Shared or dedicated — New Zealand either way.