Cloud AI charges for every request
Cloud and enterprise AI services are easy to start with, but they charge per use, so costs climb with every request. Automated research work makes them expensive quickly.
Cleanroom runs AI models in the GPU memory your research jobs leave unused. Your researchers get a private AI service, with no new hardware to buy, no per-use cloud bills, and no data leaving campus.
A job needs more memory, so Cleanroom hands it back.
Try it: click or tap a GPU to give its research job more memory.
Research increasingly depends on AI, especially for automated work and for data that can't leave campus. Today's options all fall short.
Cloud and enterprise AI services are easy to start with, but they charge per use, so costs climb with every request. Automated research work makes them expensive quickly.
Using an outside provider means sending prompts and documents off campus, which rules it out for data that has to stay within the institution.
Running your own AI service usually means buying new AI hardware and hiring dedicated staff, which most universities can't afford.
On research clusters, most GPU-hours leave most of the GPU's memory empty, even while its compute is busy. Cleanroom runs AI in that empty memory.
See the evidence56%
of 62,827 profiled GPU-hours on a U.S. research university's shared GPU cluster peaked at a quarter or less of GPU memory.
~50%
of GPU-enabled jobs used up to 25% of GPU memory. Li et al., ISC High Performance 2023
28.6%
average GPU memory use across 118,276 jobs. Sencan et al., PEARC '25
Both are peer-reviewed studies of NERSC's Perlmutter supercomputer.
Your cluster's jobs use its GPUs just as they do today, and many leave most of each GPU's memory empty.
Cleanroom places AI models in that unused memory, turning your cluster into a private, campus-wide AI service.
Whenever a research job needs its memory back, Cleanroom releases it immediately.
Some universities have built their own AI services to stay in control. Cleanroom gives you the same control and flexibility on hardware you already own, without building anything yourself.
Choose which models to run, including open models whose weights your institution keeps.
Because you keep the model weights, results stay repeatable even if a cloud model changes or disappears.
One shared service for all your researchers, without per-request bills that grow with every use.
Researchers sign in with their campus credentials, and usage can be billed to grants or departments.
Set rules for each type of data, so sensitive data stays on hardware your institution controls.
AI runs on the GPUs your institution already owns.
We're starting with institutions that already run shared GPU clusters and have researchers who want AI.
For institutionsResearch computing centers at mid-sized U.S. research universities that run shared GPU clusters (about 100–500 GPUs) and have researchers who want AI, but lack the staff to build their own service.
We're looking for research computing teams to work with. In a first conversation, we'll compare notes on your cluster and what your researchers need from AI.