Private ChatGPT on a dedicated DGX Spark
ChatGPT for your team. On a machine only you rent.
Run an open, ChatGPT-class assistant on your own dedicated DGX Spark. Same workflow for your team (chat UI and API), but every prompt stays on a dedicated machine assigned to you alone.
$0.79/hour, billed per minute · pay $10, get $20 in credit on your first top-up
Shadow AI
Blocking doesn't work. Replacing does.
If you're searching for how to block ChatGPT on the firewall, you already know the problem.
The leak is invisible
Contracts, source code and customer records get pasted into a US-hosted consumer tool, one convenient prompt at a time. Samsung banned ChatGPT in 2023 after engineers pasted internal source code into it. Most companies never find out. What the research actually shows.
The block backfires
Filter it on the corporate network and the same prompts move to personal phones: the same leak, now with zero visibility. Bans don't remove the need; they remove your ability to see it.
The replacement wins
Give the team a sanctioned assistant that's just as convenient. Adoption is voluntary because the tool is good, and every prompt lands on your dedicated EU machine instead of someone else's cloud.
What you get
The assistant your team already knows.
A familiar chat interface
Open WebUI on your instance: conversations, folders, file upload and shared team workspaces. Your team won't need training to switch; it works the way they already work.
An OpenAI-compatible API
The same endpoint shape your tools already speak. Point internal apps, editors, and scripts at your own instance by changing one base URL.
Open models, your choice
Qwen, Llama, DeepSeek: swap models as the field moves. Up to 128 GB of unified memory holds large models; for a responsive team chat, a mixture-of-experts model such as Qwen3 Coder 30B (~81 tok/s single-stream, measured on our fleet) is the practical choice.
Your logs stay on your machine
Logs, prompts and documents stay on the node's local disk, inside your own container. GPUwerk does not access them except at your request or where the law requires it (see the DPA).
The cost math
Per-seat pricing punishes growth.
A dedicated machine doesn't care how many people use it.
Per node, metered
Measured GPU runtime, not seats.
A month, always on
730 node-hours, however many people use it. How many users one node serves depends on the model; see benchmarks.
Unified memory
Room for large open models and long contexts on one machine.
| Compared on | Private ChatGPT on GPUwerk | ChatGPT Enterprise |
|---|---|---|
| Pricing model | $0.79/hour per node, metered | Per seat, quoted on request |
| What you pay for | Node-hours while it runs | Seats, whether used or not |
| Data location | Your dedicated machine, EU-Central | Provider cloud, US company |
| Who can read prompts | You; GPUwerk only under the DPA | Provider, under its terms |
| Model quality | Best open models: excellent, a step behind the frontier | Frontier models |
We quote no seat price for ChatGPT Enterprise because OpenAI publishes none: it is sold on a negotiated quote. Compare that quote with measured GPU runtime and the number of nodes your workload needs. Read the last row carefully too: frontier models are still ahead of open ones, and if your use case needs the absolute frontier this isn't your page. For everyday drafting, coding, and internal knowledge work, current open models clear the bar.
Beyond the machine
Help with adoption, if you want it.
Team onboarding
We can run a practical team training session: prompt patterns, safety guardrails, and quality checks.
Implementation support
We help connect the model to your apps, including what your tools should do when the instance is stopped or busy.
Governance pack
A lightweight rollout plan: who can use it, who reviews prompts, and what the pilot has to show before wider use.
Questions.
Is a private, self-hosted assistant as good as ChatGPT?
For drafting, coding, internal Q&A and most day-to-day work, current open models (Qwen, Llama, DeepSeek at 70B-class and up) are close enough that most teams don't notice the difference. Frontier proprietary models still lead on the hardest reasoning and coding tasks. If your use case genuinely needs the absolute frontier, this isn't the right fit. For everything else, it is.
How is this different from just blocking ChatGPT?
Blocking removes visibility, not the need: prompts move to personal phones instead. A private assistant is a sanctioned replacement people actually want to use, so the sensitive data stays on your own instance instead of going around the block.
How long does setup actually take?
A working chat interface with a model loaded is achievable in an afternoon: deploy a Spark, pull a model, start Open WebUI. GPUwerk can also set it up for you: see the "Have me set it up" option below.
Can our developers use it as an API too, not just chat?
Yes. The same instance exposes an OpenAI-compatible endpoint, so editors, internal tools and scripts point at it by changing one base URL. No separate deployment is needed.
Where does the data actually live?
On your dedicated node in EU-Central. Nothing is shared or virtualized: no other customer's workload ever runs on the same machine. As a matter of policy set out in our data processing agreement, GPUwerk does not access what runs inside your container except at your request for support or where a legal obligation requires it; infrastructure administrator access to the underlying machine exists, as on any hosted service. See the full compliance and data-residency details.
Private ChatGPT this week.
Deploy a Spark, load a model and invite the team. First top-up: pay $10, get $20 in credit, about 25 hours of compute for the setup and trial.