About 90% of hospital servers sit idle while the hospital buys the same compute back from the cloud at a multiple of the price. Clique joins the machines you already own into one computer that runs AI in clinical priority, inside hardware enclaves. Patient data never leaves the building.
01 · The problem
Imaging GPUs on four, a research cluster on two, workstations idle overnight on every floor. Each was bought for its own peak, so the daily work runs in the cloud at a markup.
02 · The join
Clique inventories every CPU, GPU and volume in hours, then joins them into one fabric with one queue. Hardware you already own starts working as one machine.
clique n. a set of nodes in which every one is connected to every other.
16 nodes · 120 connections · one queue. It's the graph in our name.
03 · The night shift
A pathologist uploads whole-slide images at five and goes home. Clique finds twelve idle GPUs across four departments and runs sealed inference through the night.
04 · The scheduler
When a suspected stroke hits the ER at 4 a.m., the emergency lane preempts every batch job in the estate, then resumes them exactly where they left off.
05 · The head node
It sits in the basement server room: one install that discovers the hardware, schedules the work, and proves what happened.
06 · Security
Every workload is sealed in its own vault, hardware encrypted so even Clique can't see inside. Integrity is proven every sixty seconds; every event lands on an audit chain kept for seven years.
Models travel. Data doesn't. Vendor models deploy into the enclave, never the other way around.
07 · The health system
Most systems are several hospitals with several server rooms and no shared queue — the flagship buys more GPUs while the community sites sit idle. Clique federates them into one pool. A model improved on pathology at the main campus sharpens the read at the community hospital, because the weights travel and the studies never do.
Each campus keeps its own enclave, its own attestation and its own audit chain. Nothing about joining a system weakens what any one building already proved.
Site-to-site links carry model weights and scheduling signals only. A study acquired at one campus is scored at that campus.
With expertise from
The product
For IT, CIOs, compliance
Every GPU, workload, enclave and audit event, live in one view. Auditors get their export in one click.
Explore the Control PlaneFor clinicians, researchers
Drop a labeled dataset and pick a task. Clique trains in a sealed container and hands IT a versioned artifact to review.
Explore the StudioThe math
Reference model: 500-bed hospital, 800K imaging studies a year.
These are modeled numbers, and yours will differ; a two-week pilot on your own hardware measures your actual idle capacity and costs.
Questions
Clinical use always wins. The scheduler checkpoints the batch job, moves it to another node, and hands the machine back in seconds. Nobody in a reading room ever waits on a research workload.
Nobody, including us. Workloads run inside hardware-encrypted enclaves such as Intel TDX, AMD SEV-SNP, and NVIDIA H100 confidential computing, so even Clique can't see what's being processed. Data never leaves the hospital network.
The machines are yours and keep doing their normal jobs. Queued batch work pauses and resumes when the head node comes back. Nothing clinical depends on Clique being up.
One agent install per machine, and the inventory finishes in hours. The system stays read-only until your team approves the scheduling policies. Clinicians keep their existing tools; results land in PACS.
The servers and workstations you already own. Sealed workloads need TEE-capable parts: Intel TDX or AMD SEV-SNP on the CPU side, NVIDIA H100 confidential computing on the GPU.
Two weeks, on your own hardware. It starts with a read-only inventory, so nothing runs until you approve it. At the end you have your measured idle-capacity and cost numbers, and you decide.
The pilot
Tell us about your environment and we'll scope a two-week pilot on the hardware you already own. We reply within 24 hours.