The Atom is the AI factory, delivered as one finished block.
Not space you rent. A factory you receive. The Atom arrives factory-built, is set in place, commissioned as one unit, and runs behind your own boundary for a single sovereign tenant.
Updated
What an Atom is
An Atom is a complete AI factory block. Compute, liquid cooling distribution, electrical distribution and the operating layer are engineered together, built off site, and delivered as one commissioned unit. The customer receives a running factory, not a shell and a parts list.
The block is sized for one tenant. Nothing inside the boundary is shared, and the tenant's data, models and workloads stay inside it.
- 32 racks of high-density, liquid-cooled compute running NVIDIA's latest GPUs at time of delivery.
- 8.8 MW of IT load under a 13.2 MW facility envelope, a design PUE of 1.5.
- Direct-to-chip liquid cooling on a closed loop, with no routine draw from community water.
- Factory-built and crane-set, so site work is foundations, utilities and commissioning.
The Atom Edge
The Atom Edge is one quarter of an Atom: eight racks, 2.2 MW of IT capability under a 3.3 MW envelope. It is the same block at a standard smaller increment, which means a partner can start with one Edge and grow into a full Atom without changing the design.
The Edge is the unit behind SAVRN's partner models. A Boxcar or a Sidecar is an Atom Edge placed on a partner's asset.
How it runs
Every Atom runs on the SAVRN operating layer: a private model registry, governed AI agents with human approval on anything that ships, signs or spends, and a metered bill by the token. The tenant reads a utility bill. SAVRN's infrastructure team runs the plant.
At scale, Atoms line up into a campus. The digital twin shows the same block repeated until the factory is the size of the demand, with power and cooling growing in the same increments.
- Metered by the token, billed like a utility.
- Private model registry and agent oversight inside the tenant boundary.
- Every action logged to a tamper-evident audit trail.
Key details
- Delivered and commissioned as one unit, not assembled on site from parts.
- One sovereign tenant per Atom; no co-location inside the boundary.
- Powered on site by the Electron family: Proton fuel cells and Ion storage.
- Cooled on a sealed loop; heat is captured for reuse on the campus.
- Operated on the SAVRN platform with human gates and full audit trails.
- Grows in standard increments: Atom Edge, Atom, campus.
An Atom does not arrive on the community grid or drink from the community water system. It stands on its own power and its own sealed cooling loop, and the campus it anchors ships with a school.
How it is delivered
Site and load
SAVRN Grid screens the site for power, fiber and constraints, and the block count is set against the demand.
Power first
The Electron plant is specified and permitted so the block is never waiting on an interconnection queue.
Factory build
The Atom is built and tested off site while foundations and utilities go in.
Set and commission
The block is crane-set, connected, commissioned and handed over as one running unit.
Operate and meter
The SAVRN platform runs the plant; the tenant reads a token-metered bill.
Questions people ask
Which GPUs does an Atom run?
NVIDIA's latest GPUs at time of delivery. SAVRN does not pin a block to a GPU generation, because a block delivered next year should ship with next year's hardware.
Does SAVRN build the physical facility?
Yes. SAVRN designs, builds and delivers physical AI factories, working through manufacturing and construction partners the way every infrastructure developer does.
Can a customer start smaller than a full Atom?
Yes. The Atom Edge is a quarter block, eight racks at 2.2 MW of IT, and four of them make an Atom. It is the unit behind the Boxcar and Sidecar partner models.
Is the Atom shared between customers?
No. Each Atom serves one sovereign tenant. Data, models and workloads stay inside that tenant's boundary.
See The SAVRN Atom in the whole system.
Every piece here is rendered and narrated on the System page, and the platform that runs it has a guided walkthrough.