A private AI server, sized to your work — or a private cloud
BrainOutput runs where your confidentiality policy says it should: on a machine under your desk, on a system in your server room, or on dedicated GPUs in a private cloud. We supply and provision the hardware, install the platform and the models that fit it, and stay responsible for the whole stack.
Three machine classes
We size the machine to the models you need and the number of people using them. Specifications below are the manufacturers' own figures, verified on 2026-10-03; prices are dated third-party market references, not quotes.
NVIDIA DGX Spark — desktop unit
A private AI server that fits on a desk: the entry point for a team that wants its sensitive work answered at home.
- 128 GB LPDDR5x coherent unified memory (a 64 GB OEM-only variant is listed as coming soon)
- Up to 1 PFLOP of FP4 compute (GB10 Grace Blackwell)
- Sold through Amazon, Micro Center, PNY and OEM partners (Dell, HP, Lenovo, ASUS…)
Market reference: from about US$6,950 for the 128 GB model (ServeTheHome, 2026-10-03).
RTX workstation — RTX PRO 6000 Blackwell or RTX 5090 class
A workstation GPU for a heavier private model, or several lighter ones serving a whole team.
- RTX PRO 6000 Blackwell Workstation Edition: 96 GB GDDR7 ECC, 1,792 GB/s memory bandwidth, 4,000 AI TOPS (FP4, sparse), 600 W
- GeForce RTX 5090: 32 GB GDDR7, 512-bit bus, 3,352 AI TOPS
NVIDIA DGX Station — large system
The largest desktop-class systems we provision: room for the biggest open-weight models and many simultaneous users.
- 748 GB coherent memory: 252 GB HBM3e at 7.1 TB/s + 496 GB LPDDR5X (GB300 Grace Blackwell Ultra)
- Up to 20 PFLOPS of FP4 compute
- Ordered through OEM partners (ASUS, Dell, Exxact, Gigabyte, HP, MSI, Supermicro); lead times of 2 to 16 weeks
Market reference: from about US$100,000 / €90,000 excl. VAT (reseller pi3g, October 2026).
Rather not buy the machine? Private-cloud inference
The same platform runs on dedicated GPUs rented in a private cloud: no hardware to buy, no capacity shared with other tenants' workloads. Marketplaces such as NVIDIA DGX Cloud Lepton now aggregate GPU capacity from more than 20 partner clouds (CoreWeave, Lambda, Nebius…) through a single interface (NVIDIA, accessed 2026-10-03).
The routing policy is identical in both cases: what is sensitive stays on infrastructure you control — your machine or your private cloud — and nothing leaves that boundary unless your policy allows it.
What we deliver with the machine
Provisioned, not shipped in a box
We install the BrainOutput platform, the models sized to the hardware, and the routing policy we agree with you — you receive a working private AI, not a pallet of parts.
Sized during the diagnostic
The AI diagnostic tells you which class of machine — or which private-cloud capacity — your workloads actually need, before you commit to anything.
Operated with you
We stay reachable after delivery: model upgrades, new machines, and moving workloads between your hardware and a private cloud are normal operations, not new projects.
We publish no prices of our own: every configuration is quoted per engagement. The market references above are dated third-party figures, shown so you can budget honestly.
Frequently asked questions
Can BrainOutput run entirely on our own hardware?
Yes. That is the default shape: the platform, the models and your data all stay on your machines. A cloud model is only ever used for non-sensitive work, and only when your policy explicitly allows it.
How much does a private AI server cost?
As a dated market reference: a DGX Spark-class desktop unit with 128 GB of unified memory sells from about US$6,950 (ServeTheHome, 2026-10-03); a DGX Station with 748 GB of coherent memory is listed from about US$100,000 / €90,000 excl. VAT (pi3g, October 2026). We publish no prices of our own — hardware, provisioning and support are quoted per engagement.
Which model fits on which machine?
It depends on the model's size and how many people use it at once — that is exactly what the AI diagnostic measures against your workloads. We never publish a model-to-machine claim we have not run ourselves.
On-premise or private cloud — how do we choose?
On-premise when data residency or policy demands it, or when usage is steady enough to own the capacity. Private cloud when you want dedicated GPUs without buying hardware, or to absorb a peak. The platform treats both identically, and you can mix them.
Do you resell NVIDIA hardware?
We supply and provision machines through the manufacturers' retail and OEM channels, and we are your single contact for the whole stack — hardware, platform, models and support. Warranty and lead times are those of the manufacturer or OEM; DGX Station lead times currently run 2 to 16 weeks (accessed 2026-10-03).
See it running before you size anything
Book a demo and we will show you the platform running on our own machines. Or start with an AI diagnostic: we assess your processes and tell you what hardware — if any — your AI actually needs.
Sources
- NVIDIA DGX Spark — product page — accessed 2026-10-03
- ServeTheHome — DGX Spark pricing (128 GB and 64 GB) — accessed 2026-10-03
- NVIDIA RTX PRO 6000 Blackwell Workstation Edition — product page — accessed 2026-10-03
- NVIDIA GeForce RTX 5090 — product page — accessed 2026-10-03
- NVIDIA DGX Station — product page — accessed 2026-10-03
- pi3g — DGX Station price and availability — accessed 2026-10-03
- NVIDIA Newsroom — DGX Cloud Lepton announcement — accessed 2026-10-03