Northstar One keeps the work that cannot leave, in-house.
A small fleet of machines we own, running models we host ourselves. It exists for the work where sending data to a frontier provider is the wrong answer: sensitive material, long-running jobs, and the experiments that would be too expensive to run per token.
Not every task belongs on someone else's computer.
Some material cannot be sent to an external provider at all. Some jobs run continuously and would be priced per token into absurdity. And some questions are only worth asking if you can afford to ask them a thousand times.
The solution
Owned machines, owned weights, a deliberate boundary.
Work is placed on purpose: on-device, on our own hardware, or on a frontier model — and the choice is recorded rather than assumed. The fleet is reached over a private network, never exposed to the public internet.
Audience
Work that carries a data boundary or runs without stopping.
Sensitive dataContinuous jobsClassification at volumeEmbeddings and retrievalModel evaluation
Business model
Infrastructure behind the products, and behind client engagements.
TodayInternal workloads
AccessPrivate network only
Client workWithin engagements
The learning loop
Knowledge develops outside the live product.
A dedicated server environment supports product research, model development and recurring evaluations. Tested improvements move into product components — after review, not before.
01
Collect
Product data, examples and references, gathered from work that has actually shipped.
02
Develop
Domain models and knowledge rules, refined against those examples.
03
Evaluate
Fixed test sets and cross-review, so an improvement has to prove itself.
04
Apply
Tested versions move into the product; knowledge enters the shared base after review.
Finding → targeted correction → recheck
Work that stays put
Material with a hard data boundary is processed on machines we own, reachable over a private network only.
Questions worth repeating
A job priced per token is asked once; a job on our own hardware can be asked a thousand times.
Adapted from the studio's architecture overview, September 2026. The environment is described there as an available lab with scheduled research and evaluation cycles; the diagram claims no measured speed gains.
Stage & milestones
Where Northstar One is today
Updated Q3 2026
Built · Done
The fleet
Owned hosts reachable over a private network, serving self-hosted models with an explicit placement decision per job.
Now
Serving internal work
Classification, embeddings, evaluation and long-running jobs for the products in this portfolio.
Next
Client-facing deployments
Offer the same boundary to organisations that cannot send their material outside.
Later
Measured placement
Publish how we decide what runs where, with the cost and quality evidence behind it.
Get involved
Some data should not leave your building either.
If your work carries a hard data boundary, talk to Studio Northstar about what a private deployment would look like.
