Skip to content

Early access by request

GPU rental in Uzbekistan

Describe your workload and we’ll reply with rental terms. No need to choose GPUs in advance.

Request early accessFind GPUs for your workload
  • GPUs billed per second
  • Data centre in Tashkent
  • Support around the clock, every day

Example

LLM inference, 70B parameters, FP8

≈ 87 GB of GPU memory

Fit in memory

  • 1 × RTX PRO 600096 GB
  • 1 × H200141 GB

An estimate from open formulas, not a benchmark or an invoice.

Open this example in the sizing tool

Workloads

What matters in each workload, and what its estimate rests on.

  • Training

    Training models from scratch, from language models to computer vision.

    Sized by GPU memory for the weights, gradients and optimiser state.

    Estimate Training

  • Fine-tuning

    Adapting open models to your data.

    Sized by GPU memory: LoRA and QLoRA need less of it than full fine-tuning.

    Estimate Fine-tuning

  • Inference

    Serving a model to users: assistants and answers grounded in your documents (RAG).

    Sized by GPU memory for the weights, with headroom for the KV cache and activations.

    Estimate Inference

  • Rendering and video

    3D rendering and video work, including generation.

    The scene must fit one RT-core GPU’s memory; encoding and video analytics count NVENC and NVDEC engines, generation is sized by memory, as inference is.

    Estimate Rendering and video

  • Scientific computing (FP64)

    Scientific and engineering computing in double precision (FP64).

    Sized by FP64 throughput, not by memory.

    Estimate Scientific computing (FP64)

  • I know which GPU I need

    Pick a GPU and a count, and we’ll work out the GPU-hours.

    Choose a GPU

Billing, console, support and data

How billing works, what the console has, how support answers and where the data is kept.

  • Billing

    GPU time billed per second

    • Models via API billed per token
    • No minimum rental term
    • Time counts from the moment the server is ready to work
    • Creating a server is not billed
    • Downtime caused by us is not billed
    • A stopped server: you pay only for its disks, not for the GPU or the IP address
    • Traffic is not billed
    • Invoiced for the time used, with no prepayment

    How much time goes on the bill

    Per second

    7 min

    With an hourly step

    1 h

    Not billed: 53 min

    Per second

    1 h 5 min

    With an hourly step

    2 h

    Not billed: 55 min

    Per second

    8 h 10 min

    With an hourly step

    9 h

    Not billed: 50 min

    An example in server time. We send prices in our reply.

  • Console

    • SSH access with root rights
    • Ready images with the NVIDIA driver, CUDA and Docker
    • Your own server images and snapshots
    • Persistent disks
    • Two-factor sign-in
    • GPU monitoring
    • Invoices

  • Support

    10 min

    Reply to a critical issue, for every customer

    • Available around the clock, every day
    • Channels: email, Telegram and WhatsApp
    • Languages: Russian, Uzbek and English

  • Data and data centre

    Tashkent

    Partner data centre

    • Your servers’ disks, snapshots, backups and logs, and the console and billing data, are held in a partner data centre in Tashkent
    • Redundant power and cooling, diesel generators

Early-access terms

We’ll reply with terms for your workload, as we don’t publish prices yet.

Confirmed

  • We are placing our GPU capacity in Uzbekistan
  • Contract with a legal entity registered in Uzbekistan
  • We work with companies and individuals
  • Billing in Uzbek soum (UZS)
  • Nothing to pay before the contract
  • Access: UZ Dynamics Console, SSH
  • Storage: persistent disks, snapshots and backups

In preparation

  • Four formats in preparation: cloud GPUs, serverless inference, models via API, dedicated servers and clusters; terms for each are not yet confirmed
  • Access methods and provisioning

Discussed per request

  • Prices and what is included
  • Availability and form factor per model
  • SLA guarantees
  • Trial access
  • Lead time
What we know and what is not yet confirmed

What we know

GPU capacity and contract in Uzbekistan

  • We are placing our GPU capacity in Uzbekistan
  • A partner data centre in Tashkent
  • Contract with a legal entity registered in Uzbekistan
  • Billing in Uzbek soum (UZS)
  • Seller: LLC UZ Dynamics

Catalogue and formats

  • 11 GPU models in the data-centre catalogue, L4 to B300; availability not yet confirmed
  • Four formats in preparation: cloud GPUs, serverless inference, models via API, dedicated servers and clusters; terms for each are not yet confirmed

Billing, storage and support

  • GPUs are billed per second
  • Models via API are billed per token
  • Minimum term in every format: none
  • Storage: persistent disks, snapshots and backups
  • Backups and monitoring: disk snapshots and backups, monitoring in the console
  • Support: email, Telegram and WhatsApp, around the clock, every day

Not yet confirmed

  • Prices and what is included
  • Availability and form factor per model
  • Security and isolation
  • Access methods and provisioning
  • Bandwidth and cluster networking
  • SLA guarantees
  • Trial access
  • Lead time

From request to access

Tell us about your workload first. You don’t need a GPU configuration to enquire.

  1. Describe your workload

    Leave your contact details and what you want to run. Add a configuration if you have one.

  2. Discuss the terms

    Send terms for your configuration

  3. Agree and get access

    Confirm the configuration and contract. Access details are agreed with you.

All steps and responsibilities

  1. Request

    What you do
    Pick a configuration in the sizing tool or describe the job
    What we do
    Receive the request with its configuration

  2. Reply

    What you do
    Wait for our reply; nothing to pay
    What we do
    Send terms for your configuration

  3. Details

    What you do
    Answer questions about the job and data
    What we do
    Confirm the GPU model, count and format

  4. Contract

    What you do
    Sign the contract
    What we do
    Sign the contract as an Uzbek legal entity and bill in UZS

  5. Access

    What you do
    Run your workloads
    What we do
    Provide access; access methods come with the terms
Rent or buy GPU servers

Owning servers pays off with steady high utilisation and an in-house ops team. Renting suits uneven load or changing workloads.

  • Upfront cash

    Buy
    A large upfront payment before any work (or leasing/financing)
    Rent
    Payments under a contract, no hardware purchase

  • Time to first job

    Buy
    Delivery, installation and setup
    Rent
    Lead time being confirmed

  • Power and cooling

    Buy
    Your monthly cost
    Rent
    Not your cost: covered by whoever provides the GPUs

  • Operations

    Buy
    Your own engineers for drivers and repairs
    Rent
    Hardware upkeep isn’t on you; who maintains it will be set out in the terms

  • Utilisation

    Buy
    An idle GPU is still paid for
    Rent
    You pay for the time you rent, not years of ownership. GPU count per format comes with the terms

  • New GPU generations

    Buy
    A server stays the generation you bought
    Rent
    The catalogue spans several generations, from Ampere to Blackwell

  • Networking between servers

    Buy
    Your own switches and links
    Rent
    We’ll confirm in our reply

Flagship GPUs by generation

  1. 2020: A100 (Ampere)
  2. 2022: H100 (Hopper)
  3. 2023: H200 (Hopper)
  4. 2024: B200 (Blackwell)
  5. 2025: B300 (Blackwell)

Announcement years per NVIDIA.

How many GPUs does your workload need?

Pick a workload: the sizing tool shows the memory it needs, the configurations that fit and the GPU-hours a month. An estimate from open formulas, not an invoice.

TrainingFine-tuningInferenceRendering and videoScientific computing (FP64)

Example: LLM inference, 70B parameters in FP8. It needs about 87 GB of GPU memory, which fits on 1 × RTX PRO 6000 or 1 × H200.

Open the sizing toolI know which GPU I need

Explore GPU options

Four examples from the catalogue. Compare memory and workload fit, or browse every model and its specifications.

Specs from NVIDIA’s official materials. Availability and form factor of each model are being confirmed.

Browse all 11 GPU models

Specs from NVIDIA’s official materials, checked 26 September 2026.

Formats: from cloud GPUs to dedicated servers

These formats are in preparation: we confirm each one’s availability in our reply. If you’re not sure which fits your workload, choose “Not sure yet” in your request.

  • Cloud GPUs on request

    When you need GPUs for an experiment or a fine-tune for a limited time

    Minimum term
    none
    Billing increment
    per second
    Access
    UZ Dynamics Console, SSH
    Request “Cloud GPUs on request”

  • Serverless inference

    When your model serves requests and traffic is uneven

    Minimum term
    none
    Billing increment
    per second
    Access
    Access methods are being confirmed; we’ll send them with the terms.
    Request “Serverless inference”

  • Models via API

    When you want a ready model over an API, not your own server

    Minimum term
    none
    Billing increment
    per token
    Access
    Access methods are being confirmed; we’ll send them with the terms.
    Request “Models via API”

  • Dedicated servers and clusters

    When the load is steady or you need whole GPU servers

    Minimum term
    none
    Billing increment
    per second
    Access
    UZ Dynamics Console, SSH
    Request “Dedicated servers and clusters”

From test to production

Cloud GPUs will suit an experiment, and steady load will be able to move to dedicated servers. We’ll discuss how to switch formats in our reply.

For companies: clusters and contracts

For clusters over 8 GPUs and dedicated servers. Describe the workload and we’ll reply with the configurations we can offer and our terms.

Discuss a cluster

Our reply will include

  • A configuration for your workload: GPU models and server count
  • Contract terms with a legal entity in Uzbekistan, billing in UZS
  • Storage of your workload data in Tashkent, set out in the terms and contract
  • Answers on security and isolation: ask your questions in the comment

About the company

UZ Dynamics: GPU rental with compute capacity in Uzbekistan, a contract with a legal entity in Uzbekistan and billing in UZS.

What is confirmed

  • We are placing our GPU capacity in Uzbekistan
  • A partner data centre in Tashkent
  • Contract with a legal entity registered in Uzbekistan
  • Billing in Uzbek soum (UZS)
  • 11 GPU models in the data-centre catalogue, L4 to B300; availability not yet confirmed

Contacts

Seller: LLC UZ Dynamics

GPU rental FAQ

How much does GPU rental cost?

We don’t publish prices yet. We’ll reply with terms for your configuration, billed in UZS. The “Sizing” section already shows how many GPUs and GPU-hours a month your job needs.

Which GPU should I choose?

GPU memory (VRAM) comes first: the model or scene has to fit on the GPU. Pick your workload in the sizing tool to see which GPUs fit and how many. Not sure? Choose “Not sure yet” in your request and describe the job.

When can I get access?

We’re collecting early-access requests now. We’re not giving dates yet; our reply will say what we know at that point.

Can I test before paying?

Trial access terms are still being confirmed. If a trial matters to you, say so in the comment on your request.

Is SSH available?

Yes, cloud GPUs and dedicated servers are accessed over SSH with root rights and managed in the console. Access methods for serverless inference and models via API come with the terms.

Where is my data stored?

Your servers’ disks, snapshots, backups and logs, and the console and billing data, are held in a partner data centre in Tashkent. Isolation between customers’ workloads and access to the servers are still being confirmed and will be set out in the terms and contract.

Where is the GPU compute located?

In Tashkent, in a partner data centre. Its power and cooling are redundant, and it has diesel generators.

What happens after I send a request?

We read your request and reply with terms for your configuration. Then we confirm details, sign a contract with a legal entity in Uzbekistan and provide access. Nothing to pay before the contract.

13 more questions
What is a GPU server?

A server with graphics accelerators (GPUs). Our catalogue lists NVIDIA cards. They run matrix operations in parallel, which suits neural networks and rendering.

Can I pay hourly?

Yes, and more precisely: cloud GPUs, serverless inference and dedicated servers are billed per second, models via API per token, all in UZS. Time counts from the moment the server is ready to work; creating it is not billed. A stopped server bills for its disks only, at the usual storage rate; its GPU and IP address are not billed. A stopped server does not hold its GPU: you can turn it on again when a GPU of the same model is free.

Can I run LLM inference?

We’re preparing serverless inference, models via API and cloud GPUs for early access. The sizing tool shows how much memory your model needs.

Can I train AI models?

We’re preparing cloud GPUs, dedicated servers and clusters for training. We’re currently accepting early-access requests. The sizing tool shows the minimum GPUs to hold the weights, gradients and optimiser state.

Can I fine-tune LLMs?

We’re preparing cloud GPUs for fine-tuning and accepting early-access requests. The sizing tool estimates memory for LoRA, QLoRA and full fine-tuning.

Can I install Docker?

Yes. Docker, the NVIDIA driver and CUDA come in the ready images, and SSH access has root rights: run your own Docker images on cloud GPUs and dedicated servers.

Can I get several GPUs?

You can request several GPUs for early access. The sizing tool shows how many you need. The sizing estimate assumes up to 8 GPUs per server. We’ll confirm the actual configuration and multi-GPU availability in our reply.

Can I scale?

We’re preparing formats from cloud GPUs to dedicated servers and clusters. How quickly you can add GPUs depends on availability; we’ll tell you in our reply.

Is persistent storage available?

Yes: persistent disks, snapshots and backups.

What happens if a GPU is unavailable?

We check availability for every request. If a GPU isn’t available, we’ll say so and suggest GPUs with similar memory from the sizing tool.

Do you provide support?

Support: email, Telegram and WhatsApp, around the clock, every day. Languages: Russian, Uzbek and English. We reply to a critical issue within 10 minutes.

Can I run production workloads?

We’ll discuss production terms with you, including SLA and redundancy. There’s no public SLA yet. The “Formats” section outlines the path from test to production; details come in our reply.

Can I sign with an Uzbek legal entity and pay in UZS?

Yes. The contract is with a legal entity registered in Uzbekistan, and billing is in UZS.

GPU planning guides

All guides

Early access

Need GPUs?

Pick a configuration in the sizing tool or describe the job. We’ll reply with terms.

Request early access

More than 8 GPUs or dedicated servers?

Describe the workload and we’ll reply with the configurations we can offer and our terms.

Discuss a cluster