Capabilities

Private AI infrastructure

Compute rental for running AI models, with the hardware, model and runtime selected for your workload. Exact configurations and availability are discussed privately.

Hardware categories · configuration confirmed per rental

Memory has to fit the workload.

NVIDIA RTX 5090 GPUs
GPU VRAM is specified separately from the host’s system memory.
Unified-memory systems
The offering includes terabytes of unified memory. Capacity per host and capacity usable by one workload are confirmed separately.
One workload
The proposal identifies its actual memory allocation and software support. Separate machines are not one interchangeable memory pool.

These are hardware categories, not a live inventory or a benchmark. Unified memory and system RAM must not be read as RTX GPU VRAM.

Start with the workload

A model can load successfully and still run out of memory when prompts grow or requests overlap. Model revision, precision, context length, input and output lengths, concurrency and runtime all affect fit.

Describe the task and expected usage if you have not chosen a model. If you have, include its revision and runtime. Model licences still apply; a rental does not change their terms.

Compute access, managed deployment and on-site installation are different scopes. The proposal identifies who runs the model, updates the runtime, administers the host and supports the application.

Access, storage and connections

We map where the model runs, who can administer it, what gets logged and how data can leave.

This example shows the access paths around a rented server.

Client device or application

Requests, prompts and returned outputs

Prompts and outputs use an authenticated link. Encryption ends where agreed.

Rented server boundary

Location, tenancy and hardware

Access endpoint

Authenticates and routes requests to runtime

Compute and runtime

Model revision, precision and runtime version

Attached storage

Model weights and application data

Administration into host

Client, Graywater and provider admins can change software, read storage or view processes.

Agreed external paths

Optional downloads, updates, tools, logs and backups. Destinations and retention are set.

Solid lines show access. Dashed lines show possible external connections.

Private access is a defined boundary

“Local” can mean a workstation, customer premises or a rented server. Model behaviour, physical location and privacy are separate properties. Host administrators may be able to inspect storage or running processes; an API key alone does not prevent that access.

The deployment scope identifies who controls the model, prompts, application settings, runtime and network. Logging, backups and external connections need the same attention as the inference endpoint.

Follow a request through those boundaries

What the rental proposal covers

Allocation and term
Exact host configuration, capacity available to one workload, tenancy, location, rental basis and availability.
Runtime
Model revision, licence, precision, runtime and dependency versions; who installs and updates them.
Storage
Volumes, backups, export and end-of-rental deletion.
Network
Client access, exposed services, permitted outbound connections and any transfer limits.
Administration
Client, Graywater and provider access; maintenance responsibilities.
Logs
Content and metadata, locations, readers and retention.
Support
Included maintenance, response arrangements and exclusions.

A performance figure is useful only with its hardware, model revision, precision, runtime, input and output lengths, concurrency, measurement method and date. Time to first output and generation throughput answer different questions.

Describe the task and access you need