A model and a machine,
installed in the building
where the data already is.
Finding a GPU in New York is not hard. Finding one you are allowed to put the corpus on is. We fine-tune on data that stays in the building, build the server it runs on, and install it on site. A week usually gets you credentials for a test run. A machine takes longer.
The blocker is
rarely the model.
A law firm's privileged documents, a fund's positions, a hospital's patient records. Each one has a reason it cannot go into an API: a partner, an insurer, a client contract, a regulator. The model that would help most is the one nobody is allowed to send the data to.
We work from Manhattan. The machines are built and burned in at our facility upstate, then installed on your site. That distance is the point. An air-gapped install means somebody physically carries the machine in, and the machine has never been on the public internet. We do the driving ourselves.
Four pieces.
You own all of them.
Fine-tuning on a corpus that stays put
LoRA, QLoRA or full-weight training on documents you already have, run on hardware in your building or ours. The weights come back to you. We keep nothing.
The machine it runs on
Three reference designs, from a 16 GB node under a desk to a 192 GB dual-GPU rig. Sized to your model, built, burned in, and documented for an audit.
A deployment you can hand to an auditor
On your own network, or air-gapped with the machine never connected to the internet at all. The build is documented as it goes, so “where does this data live” has a document for an answer.
Somewhere to run it while you decide
Want to test the idea on real weights before committing to a build? Rent GPU time on the same hardware by the hour, or have us run a bounded pilot. Ask first and we will say whether a pilot settles the question or just adds a step.
What actually makes
a city install hard.
Floor space and power. A 192 GB rig draws around 1,150 W under load and wants a dedicated circuit. The PLATINUM-96 runs about 940 W and fits a standard 20 A circuit. In a Manhattan office that is usually the binding constraint, so we check it before we quote.
Heat and noise. A 600 W card puts that heat into the room. The PALLADIUM-16 idles around 85 W and sits under a desk quietly. The bigger machines belong in a server room or a colo cage, and we will say so plainly if your closet is the wrong place for one.
Who carries it in. Air-gapped means a physical handover. We build and burn in upstate, then install on site, and the machine arrives having never been on the public internet. Somebody on our side carries it in.
Manhattan, and
a build facility upstate.
NEW YORK CITY
Where we work from. First conversations, scoping, and the on-site install itself. If something is faster to settle in person than by email, say so.
BUILD FACILITY
Where the machines are built and burned in. The same hardware runs our own research, so a design we hand over has been under load for months before it reaches you.
On-site work runs nationwide. The city is home, and it is where most of these conversations start.
The questions
we get asked.
Can you work with data we are not allowed to export?
Yes. That is most of the work. The corpus stays in your building and gets trained on hardware installed there, or the machine is handed over having never been connected to the internet.
Do we have to buy the hardware to start?
No. Rent time on the same hardware by the hour and test the idea on real weights first. If the arithmetic later says own, we will say that too. If it says keep renting, or that a smaller model on a smaller machine would do the job for half, you will hear that.
What does a build cost?
Part costs move week to week, and the right spec depends on what you run. So we quote to the job. The quote is itemised, with parts at cost and the build fee on its own line.
How long does an engagement take?
Usually a week from first conversation to credentials for a compute job. A build takes longer. Parts have lead times, and we do not shorten the burn-in.
Tell us what you are working on.
What the model has to do, and what is in the way. We will tell you whether we can help and what it would cost.
Start a project