The technology
The Sarv Grid is a trust and orchestration layer for AI compute. It verifies every node before it earns work, isolates every workload while it runs, wipes every trace when it finishes, and accounts for every job it settles. That's what turns a million strangers' machines into infrastructure.
The pipeline
Every job that enters the Grid walks the same path. No shortcuts, no exceptions for convenience. This is the engineering target the 100-Vault pilot is built to harden and measure.
The supply side
The Grid's capacity comes from machines people already own — starting with the Sarv Vault. Household life always has priority; only genuinely unused, attested capacity is ever offered.
Before any work flows, the Vault attests: genuine hardware, signed software, untampered state. Owners choose how much capacity to offer: Private Balanced Full
Jobs are split, encrypted, and placed only on nodes whose attestation satisfies the job's policy — geography, hardware class, privacy tier.
Each job executes in an ephemeral enclave — designed to have no access to the owner's data, keys or life — and is zeroised on completion. See the wall →
Every job is metered and logged. Owners see exactly what their hardware did and earned; clients can audit where their workload ran.
First workloads
A distributed network of home devices will never beat a hyperscale data centre at latency-critical work. So we don't chase it. The Grid starts where it genuinely wins: delay-tolerant, batch, privacy-hungry compute.
Turning document libraries into searchable knowledge — overnight, in bulk.
Large queues of model calls that can run over hours, not milliseconds.
Transcoding, transcribing and analysing audio and video libraries.
Automated work that runs while the world sleeps — reports, monitoring, research sweeps.
Live chat at scale, gaming, trading — latency-sensitive work stays in data centres.
Training the largest models needs dense interconnect the Grid doesn't have.
No anonymous code, ever. Every workload class is reviewed, signed and policy-bound before it touches a Vault.
Sovereignty
Know where your data runs, who it runs for, and under what conditions. Policies aren't promises in a PDF — they're enforced by the scheduler itself. A job that doesn't fit your rules never reaches your Vault.
And for organisations with obligations — public sector, healthcare, defence-adjacent research — sovereign compute isn't a preference. It's a requirement the centralised cloud answers with fine print. We answer it with geography and hardware.
| Policy control | What it means |
|---|---|
| UK only | Workloads and data never leave the country. |
| Verified hardware | Jobs run only on attested, genuine Vaults. |
| Private workload | Your job is never co-resident with others on a device. |
| Renewable preferred | Schedule to Vaults with clean local power first. |
| Customer keys | Results encrypted to keys only you hold. |
Why it can win
Existing decentralised networks — Akash, io.net, Render and others — already sell raw spare capacity. What none of them answers is the question that unlocks the valuable work: how do you safely run sensitive AI inference on a computer owned by a stranger? Attested hardware, sealed workloads, zeroisation and auditable settlement — that combination is the moat we're building. See how we test it →
The economics, in the right order
The chain runs one way: private AI makes the Vault worth owning; the Grid makes idle capacity verifiably safe to use; utilisation makes the network valuable; only then do owner economics matter. When that works, a Vault owner shares in what their hardware earns — illustrative scenarios put it at £0–60/month*, measured honestly in the pilot, including the possibility it's £0.
If the model only works as "earn money from your box", it doesn't work. That's crypto-mining economics without the crypto. The pilot's retention-at-£0 test exists precisely to prove the value stands without the income.
Join the network