AI infrastructure simulator

Size the hardware before you buy it.

Choose an open-weight model and how you will use it. The simulator works out which servers run it, how many people they serve at what speed, and what the rack costs to power and cool in the UAE.

open models
35
accelerators
21
orderable systems
20
specs verified
29 September 2026

Free with a UBM account

Sign in to use the AI simulator

The simulator is free to use with a UBM account. Sign in, and size models, servers and racks for your own workload.

Accounts are by invitation for now. UBM staff can sign in with the address UBM gave them.

We email you about your account only. UBM news is a separate choice.

What will the hardware do

  • Inference Serve a model to your people or your applications.
  • Fine-tuning Adapt a model to your data with full, LoRA or QLoRA fine-tuning.
  • Post-training Reinforcement learning with generated rollouts, GRPO style.
  • Training Train a model from scratch on your own corpus.

How it calculates

Every number, and where it comes from

  1. Read from the model itself

    Each model's architecture is read from the configuration file its lab published: layers, attention type, experts. Nothing is guessed from the name.

  2. Memory is exact

    Weights at the chosen precision, plus the conversation memory for each person at the length you set, plus a runtime reserve. The official release size is used wherever one exists.

  3. Speed is modelled, and says so

    Generating a token is limited by how fast weights and conversation memory stream from GPU memory, by compute, and by traffic between GPUs. The assumptions are listed under every result.

  4. Hardware from the datasheet

    Memory, bandwidth, compute and power come from vendor datasheets. Figures we could not confirm at source are marked, and can be excluded.

  5. Checked against published facts

    The engine is tested against what vendors state: a model that fits one GPU, a model that needs four, the power a server draws. If a check fails, the simulator is not released.

  6. Priced for the UAE

    Energy at the TAQA Distribution commercial tariff, cooling at a typical UAE facility efficiency, costs in dirhams. All of it is editable.