Infrastructure
We run the metal, or you do.
Velone is infrastructure, not a wrapper on someone else's API. We operate Firecracker hosts across eight regions, and the same build runs inside your own account when that is what compliance requires.
8
regions
Six live, two landing this quarter.
<200ms
warm cold start
p50, request to usable shell.
99.95%
control plane SLA
On the annual contract.
0
shared kernels
One guest kernel per session.
01 / the substrate
Hardware isolation, warm pools, real capacity planning.
- Hypervisor
- Firecracker microVMs on bare metal. One guest kernel per session, jailer-confined, with a minimal device surface and seccomp filters on the host side.
- Hosts
- AMD EPYC and Graviton classes, NVMe local storage, 25Gbps networking. No oversubscription of vCPU on production pools.
- Pools
- Warm pools per region and per runtime keep booted VMs waiting. Pool depth autoscales on demand and can be pinned if your latency budget is strict.
- Snapshots
- Full memory and disk snapshots to local NVMe, then to object storage. Restore of a suspended session is typically under a second.
- GPU
- L4 and A10G pools on request for agents running local inference. Same isolation model, scheduled separately.
- Control plane
- Regional, stateless in front of a replicated store. Policy is cached at the host so a control-plane partition fails closed rather than open.
- Networking
- Per-session virtual interfaces with no route to each other. All egress traverses the gateway; there is no direct path out of a sandbox.
- Storage
- Workspace volumes encrypted at rest with per-session keys. Disks are zeroed on teardown and never reused across tenants.
02 / regions
| Region | Location | Sandboxes | Gateway | Status |
|---|---|---|---|---|
| us-east | Ashburn, VA | Yes | Yes | live |
| us-west | Hillsboro, OR | Yes | Yes | live |
| eu-central | Frankfurt | Yes | Yes | live |
| eu-west | Dublin | Yes | Yes | live |
| ap-south | Mumbai | Yes | Yes | live |
| ap-northeast | Tokyo | Yes | Yes | live |
| ap-southeast | Singapore | Yes | Yes | Q4 2026 |
| sa-east | São Paulo | Yes | Yes | Q4 2026 |
Data residency is a region property, not a best effort. A session pinned to eu-central never schedules, snapshots, or logs outside the EU.
03 / deployment
Three ways to run it.
The same build ships in all three. There is no cut-down self-hosted edition with the interesting features removed.
managed
Velone Cloud
Our hosts, our regions, our on-call. The fastest way to a governed agent, and where most teams start.
byoc
Your cloud account
Data plane in your AWS, GCP, or Azure account. We operate the control plane; workload data and egress never leave your perimeter.
air-gapped
Your metal
Fully disconnected install for regulated estates. Signed release bundles, offline licence, no phone-home.
04 / how we measure
The numbers, and what they mean.
Latency claims are worthless without a definition, so here is ours.
cold start (warm pool) p50 148ms p95 212ms p99 340ms
measured from POST /sessions to a shell accepting input
cold start (cold host) p50 870ms p95 1.2s p99 1.9s
includes host selection and VM boot, no pool available
policy decision p50 1.1ms p95 3.4ms p99 6.2ms
in-process at the host, cached policy, excludes escalation
snapshot / restore p50 610ms p95 940ms
4GB working set to local NVMe
evidence append p50 0.8ms p95 2.1ms
local durable write, async replication to your store05 / questions
Are you reselling E2B or Modal?
No. We operate our own Firecracker fleet on bare metal, because the binding between the VM, the identity, and the policy has to happen below the API for the guarantees to mean anything. We will still run on top of a sandbox vendor if you already have a contract and want us to.
What happens if the control plane goes down?
Running sessions continue with cached policy. New privileged actions that require an escalation fail closed. Sessions that need a policy refresh stop rather than proceed on stale rules.
Can we bring our own hardware?
Yes, for BYOC and air-gapped installs. The requirement is nested-virtualisation-capable hosts, or bare metal, plus an object store and a Postgres-compatible database.
Do you support Kubernetes?
The control plane runs on Kubernetes happily. The data plane deliberately does not run agent workloads in pods, because a shared kernel undermines the isolation claim. Hosts join the fleet as VM hypervisors instead.
keep reading
Ring 0
Tell us where it has to run.
Most conversations start with a compliance constraint, not a feature request. Bring yours and we will tell you honestly whether we fit.