The bottleneck shouldn’t be your laptop.
A 96 GB Blackwell node and a 128 GB DGX Spark, ten minutes from campus. Free for RPI undergraduates, cheap for grad researchers, hourly for everyone else. Bring the problem that keeps running out of memory.
Start wherever you fit.
Students, clubs & researchers
We hold capacity back for RPI coursework, club projects, thesis work and grant-funded research. Tell us what you’re building and we’ll sort out an allocation. Undergraduates don’t pay for it.
Apply for RPI access Everyone elseStartups, labs & engineering teams
Training and inference capacity, a machine of your own, or a model fine-tuned on data that can’t leave your building. We’ll tell you straight whether we’re a good fit for it.
Get in touchFind your club. See what it unlocks.
Different clubs hit the compute wall in different places. Pick yours — or the one closest to what you do — and we’ll be specific about what changes.
RPai — AI & open source
Student-led AI club: research, projects, lectures and hackathons year-round.
What you can run here
- Fine-tune a 70B model with QLoRA instead of renting an A100 by the hour
- Train from scratch on a real dataset without a Colab timeout
- Serve a model to the whole club over one endpoint during a build night
Why this rig
96 GB of VRAM is the difference between quantising a model until it stops being interesting and running the thing you actually wanted to run. No queue, no preemption, no 12-hour session cap.
Claim time for your club →Don’t see yours? Every RPI club can have an allocation — just ask.
Compute,
build, deploy.
Three things, on hardware we own and look after ourselves.
GPU compute
Two Blackwell nodes with SSH, Jupyter, Docker and CUDA already set up. Your own container and your own storage — nobody else is sharing the GPU while you’re on it.
Server builds
We spec, source, build and burn in machines you keep. Each one comes with the full parts list, the thermal numbers under load, and a build guide detailed enough to follow yourself.
Fine-tuning on your own data
For corpora that can’t go to a cloud API. We fine-tune on your hardware or ours, and nothing leaves the machine — no telemetry, no third-party API in the loop.
Working with research groups
Thesis compute, benchmark runs, evaluation harnesses — including the unglamorous part, which is making a result somebody else can reproduce.
Some data can’t
go to an API.
Not “shouldn’t” — can’t. Export-controlled drawings. Fab process logs under NDA. Patient records. Privileged documents where third-party disclosure waives the privilege. Wafer yield data that is the entire competitive position of the company that owns it.
For that work the question is not which model is best on a leaderboard. It is whether the weights, the corpus and the inference all stay inside a building you control. That is what we build.
How sovereign deployment works →Your corpus never moves
Training happens on your hardware or in our rack under an agreement you have read. No telemetry, no third-party API in the loop, and nothing logged to a vendor you did not choose.
You keep the weights
You get the checkpoint, the inference stack, the eval harness and the training pipeline — enough to re-run the whole thing after we are gone. No hosted endpoint you have to keep paying for.
Air-gapped if it has to be
We will spec, build and burn in a machine that never touches a network, document it for an audit, and hand it over with the model already on it.
The kind of work these machines are for.
If your job looks like one of these, the hardware is already set up for it. If it looks like none of them, say so anyway — odd workloads are usually the interesting ones.
Fine-tuning
QLoRA on a 70B base fits either node with room to spare, and TROY-1 will go past 100B at 4-bit. Full-weight, all parameters trainable, is a different budget — about 8B. Long runs are fine; checkpoints survive between sessions.
Inference & serving
vLLM and TensorRT-LLM with continuous batching. 96 GB holds a 70B model at 4-bit with a long context, which is usually the awkward case.
Simulation
GROMACS, OpenMM and similar molecular dynamics work. CFD, N-body and FEA too. Multi-day runs are normal here and the node stays yours for the duration.
Agents & long context
SPARK-1’s 128 GB unified pool suits multi-step agent workflows, RAG pipelines, and fine-tuning runs whose working set is awkwardly bigger than a normal card’s VRAM.
Diffusion & vision
Training and sampling for SD3, FLUX.1 and similar. Large batches without the memory juggling smaller cards force on you.
Coursework
Class projects that need a real GPU for a few evenings. Small, but it is the reason a lot of people first get in touch, and it is free for undergraduates.
Two nodes,
in a rack, in Troy.
Real machines we specced, bought, built and keep running. Here is exactly what is in them, down to the part numbers.
Full specifications →The parts that are actually in our machines. Not partnerships, and not endorsements — just what we bought.
Software we built on these machines.
We build our own things too, which is the main reason we can tell you whether a machine suits a job — we have usually run something like it.
ToxScreen
Predicts hERG and cytochrome P450 toxicity from a SMILES string in seconds, and gives you a calibrated confidence score with it rather than a bare label.
Lithic
Describe a part in plain English and get CAD geometry, toolpaths, G-code and a manufacturing report back, with cycle time and material use estimated.
Naxi fine-tuning
Our Qwen3-14B run on 62,000 examples scored 82.3 NTQS, 41 points above GPT-4 on the same eval set. Same method, applied to whatever corpus you have.
What a project here looks like.
On proprietary data
Fab process copilot
A semiconductor fab has twenty years of tool logs, recipe revisions and excursion reports that cannot leave the site. Fine-tune on-premise so a process engineer can ask “what changed before the last yield dip on this tool” and get an answer with citations into the logs.
Export-controlled drawings
An aerospace supplier holds ITAR-controlled technical data. A hosted API is not an option at any price. The model is trained and served on a machine we build, that sits in their facility, on their network, with the access list under their control.
Clinical note summarisation
A research hospital wants structured summaries from free-text notes. De-identification and inference both run inside the hospital network, so no protected health information crosses a boundary and the IRB submission has a straightforward data-flow section.
Part library search
A manufacturer’s PLM holds forty thousand parts with inconsistent naming across three acquisitions. Fine-tune on the revision history so an engineer can describe a part in words and find the one that already exists instead of drawing it again.
Domain translation
A documentation team needs dense internal engineering shorthand turned into manuals, without sending the corpus to an API. Same method as the 82.3 NTQS run: curate, fine-tune, evaluate, deploy on-premise.
Buying instead of renting
A lab works out it would keep a node busy most of the month. Renting stops making sense inside a year, so we spec, source, build and burn in a PLATINUM-96, hand over the parts list and the thermals, and they run it themselves.
On the engineering side
CFD surrogate model
A rocketry team has run hundreds of nozzle simulations and each one takes hours. Train a neural surrogate on their own solver output, then sweep the design space in seconds and only run the real solver on the handful of candidates worth it.
Sim-to-real policy
A robotics subteam trains a perception and control policy across thousands of randomised simulated environments overnight, then puts it on the robot in the morning. The bottleneck stops being compute and goes back to being the robot.
Compound screening
A chemistry group screening a big compound library for hERG liability moves off a queue-limited shared cluster onto TROY-1, runs ToxScreen over the whole library locally, and gets calibrated uncertainty on every prediction instead of a yes/no.
Fixture turnaround
A machine shop needs one-off fixtures for a new line. Instead of outsourcing every CAD/CAM iteration, their engineer describes the fixture in Lithic and has STEP geometry, toolpaths and G-code the same afternoon.
Observatory data
An astronomy club stacks a full night from the campus observatory between meetings rather than between semesters, and runs the N-body integrations their laptops refuse outright.
Quantum simulation
State-vector simulation is memory-bound, so every extra gigabyte buys another fraction of a qubit. 128 GB of unified memory on SPARK-1 goes meaningfully further than a workstation card for Qiskit Aer work.
Tell us what you’re working on.
A couple of sentences is plenty. We will tell you whether we can help, what it would take, and what it would cost.