The deepest questions in AI verification live below the OS, and no cloud will hand you the machine.
Physical, supervised access to frontier silicon: the root of the machine, not a virtualized slice of it.
Reconfigurable to your experiment. Need a different setup? We rebuild the rig to spec.
Free for qualifying safety & verification research, funded through the Verifiable Compute Foundation.
Frontier silicon, multi-vendor by design.
Ten accelerators across five SKUs, because real verification can't assume one vendor's chip. Bare-metal, with root access to every one.
Hardware available
10
Accelerators
5
SKUs
2
Vendors
Every machine is bare-metal, with BIOS/UEFI and BMC-level root access: NVIDIA B300, H200, H100, A100 and AMD Instinct MI355X.
Five layers down the machine.
We don't hand you a bag of instruments. We open the whole stack, from what a workload reveals at the top to the silicon at the bottom. Each layer is a research frontier a cloud keeps sealed.
Workload Forensics
Read what a GPU is doing from how it behaves. Side-channel analysis of workload behavior (distinguishing training from inference, and fingerprinting the model on the accelerator) from power and emission signatures.
Trusted-Boundary Probing
Test the walls of the enclave. Probe TEE and confidential-computing boundaries under real datacenter conditions, where the guarantees either hold or leak, and only physical access can tell you which.
Fabric & Interconnect
Watch and bend the network. Passive TAPs across the interconnect, NIC-level experiments, and controlled network manipulation: observe every packet, or inject conditions no production fabric would let you.
Firmware & Root of Trust
Get beneath the operating system. Firmware instrumentation, boot-chain inspection, and BIOS/UEFI and BMC-level control: the ground truth every higher-level attestation ultimately rests on.
Silicon In-Path
Put your own logic in the data path. Reconfigurable DPUs, FPGAs, and SmartNICs sit between the accelerator and the world, so you can build the observation (or enforcement) point your experiment needs, in silicon.
Need a different setup? We reconfigure the cluster to meet your research needs.
Get access.
Free access
Fair-shareQualifying safety and verification research groups receive compute at no cost, allocated through an open fair-share process and covered by research funders through the Verifiable Compute Foundation.
Dedicated capacity
On requestIf your needs exceed the free allocation, dedicated remote capacity on the cluster is also available: reserved time and custom instrumentation for larger or confidential programs.
Weekly office hours
Join us every Tuesday, 10–11 AM Pacific, to talk through research and deployment possibilities.