Built for shared infrastructure.

01

Multi-tenant inference.

Several tenants serve models from the same pool of GPUs. Today most teams pin a fraction of each GPU per tenant for the day, end up at 30–40% utilization across the cluster, and pay full price for capacity that's idle most of the time. The unsafe alternative is time-sharing without enforcement — and one tenant's hung kernel can starve everyone else.

What Tenura does:

• Each inference request acquires a short lease (seconds to minutes) for the GPU slice it actually needs.
• If a request's program crashes or hangs, the lease's TTL elapses and the resource returns to the pool — automatically and deterministically.
• Capacity-seconds metering bills each tenant for what they used, not for a static reservation.

grafos deploy run llama-serve.wasm \
  --tenant acme --gpu 1 --mem 80G \
  --ttl 30s --priority guaranteed

Idle cycles between requests go back to the scheduler, which can admit Standard- or Scavenger-tier work to fill the gaps.

02

Burst training.

A research team needs eight GPUs for a six-hour training run, then nothing for three days. Static reservation ties up the hardware. Manual provisioning forgets to release. Spot interruption mid-run is too disruptive.

What Tenura does:

• The job acquires a lease for the cells and resources it needs, with a TTL that covers the run plus a buffer.
• Lease renewal is poll-driven — if the job dies, renewal stops and the lease expires on schedule.
• Guaranteed-priority leases are not preempted; Standard and Scavenger leases on the same hardware are.

grafos deploy run train.wasm \
  --tenant lab --gpu 8 --mem 1T \
  --ttl 8h --priority guaranteed \
  --renewable

When the run finishes — successfully or not — teardown runs and the GPUs return to the pool. Other tenants can immediately admit work onto them. No leftover NCCL state, no stuck QPs.

03

Shared research clusters.

A research lab gives every grad student "their" GPU — except half are idle most of the time and the other half are blocked behind whoever's running ablations. The lab buys more hardware to make the queue feel less full. The hardware spends most of its life idle.

What Tenura does:

• Replace static "GPU per student" with leases that expire when a notebook idles. Capacity returns to the pool in seconds.
• Set per-student or per-project quotas; Scavenger-tier leases let students opportunistically use idle capacity at a lower priority.
• Audit-chain records show who used what for how long — useful for grant reporting and lab budgeting.

# Default lease: 30 minutes idle and the resource releases.
grafos deploy run notebook.wasm \
  --tenant research/student-id --gpu 1 \
  --idle-timeout 30m

04

Disaggregated memory and storage.

Modern fabrics let memory and storage live anywhere on the wire — CXL pools, RDMA-attached DRAM, NVMe-oF namespaces. The hard part isn't the protocol; it's the cleanup. Allocations that survive crashes turn into orphaned mappings, fenced QPs, dangling block exports.

What Tenura does:

• Memory, queue pairs, and block namespaces are leased with the same primitive as GPUs. Same TTL, same teardown contract.
• Failed teardown puts the resource in FENCED — no new lease lands on dirty hardware until an operator clears it.
• Cross-resource leases are atomic: a workload that needs both a GPU and 80 GiB of CXL gets both or none.

grafos deploy run pipeline.wasm \
  --tenant acme --gpu 1 --cxl-mem 256G \
  --rdma-qp 4 --block 4T --ttl 1h