Choose the Cortex AI for local inference and everyday assistant use, the AI Pro for RAG pipelines, diffusion work and light fine-tuning, the AI Ultra for single-GPU 30B-class fine-tuning, and the AI Ultimate X — this build — when you need multi-GPU training or 70B+ model serving. Not sure? Call us on 0207 096 5008 and we'll match a build to your framework and model size.

As a rough guide, 64GB across two cards serves quantised 70B-class models with tensor parallelism, runs 30B-class models comfortably on a single card, and handles QLoRA fine-tuning at 70B. Bear in mind a model has to be sharded to use both cards — a single 32GB card is the ceiling for anything that cannot split. Context length and quantisation both change the maths, so tell us your target and we’ll size it precisely.

Yes. Every Cortex AI ships with NVIDIA's data-science driver stack, with CUDA and cuDNN version-matched to PyTorch and validated against a live test job before dispatch, so you're training on arrival rather than untangling version mismatches.

Yes. We supply Cortex AI workstations with Windows 11, Ubuntu LTS, or a dual-boot configuration, with the CUDA stack installed and validated on whichever you choose.

Both cards sit on the X870E AORUS Master running at PCIe 5.0 x8 each, in a Corsair 9000D super-tower sized for the airflow two flagship cards need, powered by a 1600W Seasonic PRIME PX. Frameworks that shard across GPUs — vLLM, DeepSpeed, FSDP, Accelerate — see both cards; work that cannot split runs on one. We validate a multi-GPU test job before dispatch.

Yes — every Cortex AI workstation is fully customisable. We can adjust GPU, memory, storage and cooling to suit your exact model size and budget before we build.

A 3-Year Standard Warranty: 1 year of Collect & Return service, 1 year of parts coverage and 3 full years of technical labour support.

Both. Free UK delivery is included, or you can collect for free from our Central London store at 48 Tottenham Court Road.