Confidential AI hosting · by the hour

Book your own RTX PRO 6000 Blackwell at CotrievalAI today.

A private NVIDIA GPU machine for uncensored and open-source models. Prepared for your booking alone, closed to the internet, and ended when you say so.

Book your RTX PRO 6000 See the machines

NVIDIA RTX PRO 6000 Blackwell Server Edition

96GB

GDDR7 memory with ECC

4PFLOPS

FP4 Tensor Core

1,597GB/s

Memory bandwidth

24,064

CUDA cores

Per GPU, from NVIDIA's specifications for the RTX PRO 6000 Blackwell Server Edition.

Why it takes a few minutes

Prepared for you, not pulled from a pool.

Most hosted models answer from servers that many customers share. CotrievalAI makes a new machine for each booking and loads the model into its memory for you alone. For GLM-5.3-Flash that takes 10 to 15 minutes. Nobody else's work has ever run on it, and it is deleted when your booking ends.

booking per machine

10–15 min

to create it and load the model

6 h

at most, then it is deleted

GLM-5.3-Flash, uncensored

1M-token context · London

4× H100 80GB

Preparing your machine 02:14

Machine created0:00
Startup script running0:53
Model weights in memory1:42
vLLM container startingnow
Model serving—

Ready in: 10 to 15 minutes
Ends: When you say, or after 6 hours

Machines

Two machines. Each one yours for the hour.

Each runs on its own NVIDIA GPUs in Google Cloud. You get the price per hour and the time limit before anything is created.

NVIDIA RTX PRO server with RTX PRO 6000 Blackwell GPUs

Machine / 01Created per booking

4× RTX PRO 6000

or 4× H100 80GB · GLM-5.3-Flash, uncensored

Price per hour quoted at booking

Model: An uncensored build of GLM-5.3-Flash for long coding, research and security work
Context: 1M tokens
Speed: 226–231 tokens a second for one user on 4× H100, about 145 on 4× RTX PRO 6000 (measured Oct 14)
Serving: vLLM, with an OpenAI-compatible API
Ready: 10 to 15 minutes after booking
Lifetime: Deleted when you end it, or after 6 hours
Region: Wherever GPUs are free: the US, Europe or Australia

Book this machine

Confidential VM

Machine / 02Memory encrypted

1× RTX PRO 6000

Images and video with sound · ComfyUI

Price per hour quoted at booking

Models: Qwen-Image 2.1 for pictures, MiniMax H3 for video
Speed: First 1024 px picture 15.6 s after a start, then 5.2 s each (measured Oct 15)
Memory: Encrypted while it runs (AMD SEV)
GPU: NVIDIA confidential-computing mode
Network: Sealed. No outbound traffic except the private tunnel.
Your files: Uploads and results are held only in encrypted memory
Lifetime: Stops by itself after 3 hours, and its memory is cleared

Book this machine

The price per hour and the time limit are shown before your machine is created.
Machines run on Google Cloud Spot capacity, which Google can take back at any time.
If one region has no free GPUs, your booking tries the next one.

How it works

Booked like a room. Used like your own machine.

There is no cloud account to open and nothing to install.

01
Choose a machine
The language model, or the machine for images and video with sound.
02
See the price first
The hourly price and the time limit, before anything is created.
03
It is prepared
A machine is made for your booking and the model loaded into its memory.
04
Use it, then end it
The machine and its disk are deleted. Unused time is refunded.

Privacy

Private by design. Here is exactly what that means.

Each line is true of the machine you book today. Where a protection is not on, it says so.

Privacy protections on each machine
Protection	4× GPU language model	1× RTX PRO 6000 images
A machine used only by your booking	Yes, created new	Yes
Only you can open it, over HTTPS	Yes	Yes
Memory encrypted while it runs	Not on this machine	Yes, AMD SEV
GPU in confidential-computing mode	Not on this machine	Yes
Outbound network sealed	Not on this machine	Yes
Gone when the booking ends	Yes, machine deleted	Yes, memory cleared

Google offers confidential GPUs only as single-GPU machines, and GLM-5.3-Flash needs four, so the language-model machine runs as a standard virtual machine.

FAQ

Questions

Anything else, write to hello@noahgeneralgroup.com.

Why does the language model take minutes to be ready?

Because the machine is new. Each booking gets its own machine, and the model's weights are loaded into its memory before it serves you. For GLM-5.3-Flash that takes 10 to 15 minutes. A model that answers instantly is usually running on a server shared with other customers.

Which GPU do I get?

The image and video machine always has one NVIDIA RTX PRO 6000 Blackwell. The language model runs on four RTX PRO 6000 Blackwell GPUs or four H100 80GB GPUs, whichever region has free capacity; your quote names it before anything is created.

Who else can use my machine?

No other customer. Each machine serves one booking, nothing on it is open to the internet, and you reach it through a private tunnel. The image machine also encrypts its memory in hardware while it runs.

Is the language-model machine a Confidential VM?

No. Google offers confidential GPUs only as single-GPU machines, and GLM-5.3-Flash needs four. It runs as a standard virtual machine that is created for your booking, closed to the internet, and deleted when the booking ends.

Can a booking be interrupted?

Yes. Machines run on Google Cloud Spot capacity, which Google can take back at short notice; if that happens, the unused time is refunded. When one zone has no free GPUs, the booking tries the next.

What happens when my booking ends?

The machine and its disk are deleted. On the Confidential GPU, what it held in memory, including your uploads and results, goes with it, so download anything you want to keep first.

How do I book?

Pick a machine and the hours in the booking portal. You see the price before anything is created. Your time starts when the machine is ready.