Skip to content
Olive Pines

Tell us the workload.We will tell you the machine.

Olive Pines specifies, builds and operates self-hosted AI — open-weight models on hardware you own, inside your own network.

Talk to an engineerForty-five minutes. An engineer, not a salesperson.
Indicative sizingv1 · from builds we have run
Tokens per day
Context length
Nodes
2
GPUs
5 × L40S
Draw at load
9.4 kW
Cost per million tokens
$0.21
Brief to production
14 days

Indicative only, and deliberately conservative. The real spec comes from measuring your traffic for a week.

The engagement

Four stages, one team

From the first sizing conversation to a cluster your own engineers operate.

  1. Plan

    workload sizing, written spec

  2. Design

    cluster and stack architecture

  3. Implement

    hardware built, models deployed

  4. Deploy

    handover, runbook, training

What we cover

Both sides of the same system

A consultancy stops at the architecture. A reseller stops at the box. We do the part in between, and the parts either side of it.

Capability matrix: software and hardware, one teamTwo symmetrical columns joined by a central spine. The software column covers model deployment, retrieval, automation, integration and MLOps. The hardware column covers workload sizing, GPU specification, build and rack, network and power, and ongoing operation. Every cell on both sides taps the same spine, which carries one label at its midpoint: one team.

Software

Model deployment
Open-weight models served on your own metal.
Retrieval and data
Your documents, indexed and queryable in-house.
Automation
The long jobs nobody wants to staff.
Integration
Identity, storage and logging you already run.

Hardware

Workload sizing
Parameter count and concurrency decide the machine.
GPU specification
Specified against the workload, not the catalogue.
Build and rack
Assembled, racked, networked, powered.
Ongoing operation
Monitoring, updates, a named engineer. Optional.

The metal

Run before it ships

Every configuration we sell is one we have had the models running on first.

Front elevation of an Olive Pines compute rackA 24U rack drawn front-on. A 1U top-of-rack switch sits at the top, below it three 4U GPU compute nodes separated by blank filler panels, then a 2U storage shelf, a 2U UPS and a 2U power distribution unit at the base. U positions are numbered down the left margin and the cabinet is 600 millimetres wide. The middle compute node is picked out and labelled: four GPUs, 192 gigabytes of video memory, 2.4 kilowatts.4× GPU192 GB2.4 kW
The rackSized to the workload, with headroom recorded at specification time.
The self-hosted AI stack Olive Pines builds and runsFive layers stacked bottom to top: hardware (GPU, NVMe and network), drivers and runtime (CUDA and containers), model serving (vLLM, Ollama and a request queue), retrieval and data (a vector index over an object store), and applications (assistants and agents). Arrows run upward between the layers. A bracket down the left side spans all five and is labelled "we own all of it".WE OWN ALL OF IT
The stackHardware to application, owned end to end.

The four questions that decide it

Not the ones that flatter us. These are the four that have actually stalled a deal, answered the way we answer them on a call — including the part where the answer costs you something.

Ask us something else

  • The spec prices the upgrade path before you buy: which chassis take more cards, what the power budget allows, what a second node costs. Capacity gets added to the build you already own.

    Costs youHeadroom you paid for and might never use. We size for eighteen months, not five years, and that is a judgement call we can get wrong.

  • Badly, below roughly two million tokens a day. An API is cheaper at that volume and you should use one. Above it the owned build wins, and we work the crossover out with your real numbers before you commit.

    Costs youCapital up front instead of a monthly bill, and the depreciation sits on your balance sheet rather than ours.

  • Your choice, made at the start rather than at the end. We operate it monthly with a named engineer, or we hand over runbooks, root access and a training session.

    Costs youIf you take the handover, somebody on your team owns a pager. That is a real hire or a real Friday evening.

  • You keep the machines, because you bought them. Open weights, open-source serving, configuration in files you hold, documented in runbooks you hold. Any competent infrastructure contractor can pick it up — we design for that on purpose, and it is the honest answer to a two-year-old company asking for a hardware budget.

    Costs youFinding that contractor, at a moment you did not plan for. We keep the stack boring specifically to make that search short.

Bring us a workload.

We will measure it, size the machines and tell you what it costs — before anything is bought, and before you owe us anything.

45 min · no deckTalk to an engineer