Cloud for the agentic era

Compute

Nevia’s default compute is a Firecracker microVM. Each microVM runs on a bare-metal host and is ready in milliseconds. Compute scales to zero when idle and scales up with demand.

Virtual machines for larger, long-running workloads are on the roadmap, with startup times measured in seconds.

MicroVMs

Each workload runs in its own microVM with its own kernel, and workloads are isolated by hardware virtualisation rather than by shared-kernel containers. MicroVMs run directly on bare-metal hosts without an extra hypervisor layer. Your application runs inside the microVM and can be deployed from an OCI image.

Compute is separate from storage. A machine's data lives on a replicated volume in the region, not on its compute host. This lets the machine sleep, wake and branch without moving data.

Benefits

• Starts in milliseconds: Nevia allocates compute from pre-warmed pools, with startup times in the low hundreds of milliseconds.
• Scales to zero: Sleeping machines do not consume vCPU or RAM.
• Isolated by default: Each workload gets its own kernel and its own virtual hardware, with nothing to configure.

Purpose-built orchestration

Nevia's purpose-built orchestrator manages each microVM's lifecycle and enforces isolation between workloads. Every machine is isolated from its neighbours through hardware virtualisation, control traffic stays off the guest's network interface, and ingress and egress are governed by declarative rules.

Virtual machines

Some workloads are not bursty, like always-on applications with steady demand or services you want sized once and left running. For these, we are building a virtual machine offering with a stable footprint. It runs on the same engine as microVMs, starts in seconds and supports the same branching capabilities.

Storage

Nevia offers two storage primitives: Block storage for data that needs speed and branching, together with S3-compatible object storage.

Block storage

Every VM gets a data disk that behaves like a local NVMe drive. Underneath, it is a volume backed by a pool of SSDs with built-in replication, connected to the VM over a high-speed private network. Compute and storage always sit in the same region to keep read and write latency low.

Copy-on-write is built into every volume. Snapshots and clones are metadata operations and a clone shares every block with its parent until one side writes.

Benefits

• Sub-second branching: Create a branch from a snapshot in under a second, whether it's a 1 GB workspace or a 1 TB database. No data is copied.
• Pay only for the delta: A branch shares its parent's data and only uses extra storage for new or changed blocks.

Resilient out of the box

Data is replicated across independent drives and spread across many storage nodes by default, with nothing to configure. After a failure, the system automatically rebuilds the affected replicas. Storage is separate from compute, so your data survives if a compute host fails or is replaced.

Object storage (roadmap)

Use object storage for files, build artifacts, exports and large blobs that you write once and read many times. It is S3-compatible, so existing tools and SDKs work out of the box.

Networking

Every Nevia Computer and Database gets its own name, URI and ready-to-use endpoint secured with TLS. Private networking and egress controls are on the roadmap.

Public endpoints

Each machine is named in DNS and reached through an ingress proxy on its host, which routes traffic to the destination VM. Computers get HTTPS URLs for the ports they publish, with TLS handled for you. Databases are reached via a standard Postgres TCP connection, with TLS passed straight through to the database instance.

Benefits

• Ready to connect: Create a machine and it is already addressable. No load balancers, certificates or DNS records to set up.
• Wakes on connection: A sleeping machine resumes when a client connects.
• Control traffic stays off your network: Shell, console and file transfer use a separate channel, so they never touch the machine's network or its rules.

Private networking and egress control

We are building private connections between your workspace Computers and Databases, with default-deny egress with per-machine allowlists. You decide what each machine can reach.

Agent platforms

Every agent your product starts can provision its computers and databases through API, CLI or MCP. Most of them sit idle most of the time, so they sleep until the agent needs them again.

Coding agents

Run coding agents on full Linux machines in the cloud, and let them install the tools and dependencies they need. Copy configured environments to reuse them across sessions, and save checkpoints to restore your setup and continue where you left off.

Reinforcement learning

Set up a computer in the state you want, checkpoint it, and restore that checkpoint for every rollout. Each run starts from the same memory and files.

Sovereign AI

Computers, databases and their branches stay in the region you create them in. Pick one in Europe and the data stays in Europe.

Background agents

Agents that run for hours or wake once a day keep their memory and running programs while they sleep. They pick up where they stopped, with no restart.

Preview environments

Branch a computer and its database for each pull request and review the change against a copy of real data. Delete both when the pull request merges.

Data analysis

Agent-written code runs on a computer next to the Postgres database that holds the data. Point it at a branch and the original stays untouched.

Multi-tenant platforms

Give each customer their own computer and database, each in its own microVM. Idle ones sleep, so customers who rarely log in don't keep machines running.