About
We're building the lease layer for shared infrastructure.
GPUs, memory, storage, and fabric are too expensive to leave underutilized — and too sensitive to share carelessly. Tenura turns each into a capability-bound, time-limited lease so operators can drive utilization up without giving up isolation or control.
Why we're building this
The infrastructure we have wasn't designed to be shared.
Most multi-tenant systems retrofit isolation onto hardware that was designed assuming a single occupant. The compromises show up as silent waste: GPUs idle at 30–40% because static pinning is the only safe sharing model, memory pools allocated for peak so they can be partitioned, NVMe namespaces dedicated rather than scheduled. Operators end up buying around the problem.
We think the missing primitive is a lease: a cryptographically scoped grant of access that expires by default, can be revoked deterministically, and leaves the hardware in a known clean state when it ends. Once leases are first-class, all the other things operators want — accounting, quotas, preemption, audit — fall out of the same mechanism.
How we work
What we believe.
The spec is open. The implementation is in the open too.
The fabricBIOS wire protocol and security model are public. The Rust reference implementation — daemon, bare-metal firmware, runtime crates — is source-available and on a path to a permissive license once the API surface stabilizes.
Fail closed, not best-effort.
Cleanup that "usually happens" isn't cleanup. If teardown fails, the resource is fenced and an operator has to explicitly clear it. We'd rather block a new lease than risk leaking the last one's state.
Policy lives above the firmware boundary.
fabricBIOS exposes resources and enforces leases. It does not decide who deserves what, what to do under contention, or how to bill. Those questions belong to grafOS and the operator's policy stack.
Hardware is the slow path. Be honest about it.
Some of our most useful work is on the metal — Pi5 bare-metal bringup, GPU silicon validation, mlx5 path debugging. Pretending hardware ships in software timeframes is how teams get into trouble.
Where we are
Private beta.
Tenura is in private beta with a small set of design partners. Self-hosted is the default path; we run managed cells case-by-case for teams that want us to operate the control plane. We're focused on getting the lease primitive, capability tokens, and the cross-resource graph model right before broadening access.
If you operate or rely on shared GPU infrastructure and the problems above sound familiar, we'd like to talk.
Contact
How to reach us.
- General —
info@tenura.systems - Beta access —
beta@tenura.systems - Security disclosure —
security@tenura.systems(see the security page) - Press / partnerships —
info@tenura.systems