Andro Request access

Now onboarding teams

The AI cloud for inference and agents

Serverless inference, microVMs for agents and confidential computing on CPUs and GPUs. NVMe underneath, 800 Gbit/s in between.

Request access Take the tour

acme/production
All systems normal

Inference

qwen3-32b

Request

POST /v1/chat/completions
{
  "model": "qwen3-32b",
  "stream": true,
  "messages": [{
    "role": "user",
    "content": "What is a microVM?"
  }]
}

Response

A microVM is a virtual machine stripped to the essentials: its own kernel, isolation enforced by hardware, and a boot time short enough to give every task a fresh one.

First token
0.21 s
Tokens
34
Tokens per second
58
Replicas

A preview of the Andro console, running in your browser. Names and figures are illustrative.

Requests per second
0
Replicas
0

01 / 05

Idle costs nothing

Press and hold anywhere to send traffic. Let go, and the replicas fall back to zero.

02 / 05

One agent, one machine

Every dot is a microVM with its own kernel. Click one and let it go rogue.

03 / 05

Even we cannot read it

This is your memory as the host sees it. Your cursor is the key.

04 / 05

State that stays

Machines stop, start and get replaced. The volume, and everything on it, does not move.

05 / 05

800 Gbit/s

One link moves 100 GB every second.

That is 0 GB since you got here.

Five layers, one cloud

What you get, from the API down to the network.

Serverless inference

Call a model, get tokens, pay for tokens. No instances to size and no idle GPUs on the bill: capacity follows your traffic up, and falls back to zero when it stops.

MicroVMs for agents

Agents run code that nobody has reviewed. Each one gets its own virtual machine with its own kernel, so whatever goes wrong inside stays inside.

36 machines up. 36 kernels, none shared.

Confidential computing on CPUs and GPUs

Memory stays encrypted while it is in use, on the CPU and on the GPU. The host cannot read your prompts, weights or files, and neither can we.

Every machine signs a report of what it is running, and you verify it before you send anything. Try it: the check on the right is real.

sha-384

Checking the signature

A sample report signed with a demo key. Your browser checks the ECDSA P-384 signature; nothing leaves this page. Change any value, say debug_access to on.

Persistent NVMe storage

Volumes sit on NVMe and stay put when machines stop. An agent can shut down tonight and find its files tomorrow; weights wait on flash between runs.

Private networking at 800 Gbit/s

Your machines talk over a private VPC on links of up to 800 Gbit/s. That is 100 GB every second, several times what a single NVMe drive can read.

800 Gbit/s1.4 s
100 Gbit/s11.2 s
10 Gbit/s1 min 52 s

Move 140 GB in real time. Line rate: size × 8 ÷ link speed.

In the lab

Seven things we are building on top of the platform. None of them is generally available yet.

In the lab01 / 07

Confidential microVMs

An agent's whole machine inside an encrypted enclave. You check a signed report of what booted before you hand it data or keys.

Press anywhere to boot the fleet again.

In the lab02 / 07

Inference receipts

A signed receipt with every response: which weights answered, in which enclave, how many tokens. Each one lands in a public append-only log.

Press anywhere to send a request.

In the lab03 / 07

Flight recorder

Every command, file write and network call an agent makes, chained into a signed log. Replay a run to see what it did, and prove it later.

Move across the tape to rewind the run.

In the lab04 / 07

Session snapshots

Park the context of a long conversation on NVMe and resume it days later without paying to read it all again. Priced like storage, not like tokens.

Press anywhere to park it, and again to resume.

In the lab05 / 07

Proof of residency

Cryptographic proof of where a request was served, for teams whose data must stay inside a border.

Drag the machine across the border.

In the lab06 / 07

A latency guarantee

A written limit on time to first token for ordinary pay-as-you-go traffic. When we miss it, you get credits.

Press anywhere to cause a bad minute.

In the lab07 / 07

Sealed weights

Bring your own model, encrypted. The key is released only to an attested GPU enclave, so nobody operating the cloud can copy the weights. That includes us.

Try dragging the weights out.

Request access

Andro is opening to teams in stages. Tell us what you want to run and how much of it.

Email hello@andro.ae