Your own AI cloud: GPUs, storage and software in a data centre you control
Run AI assistants, agents, search and model training on GPUs dedicated to you, next to your data, at a cost you can predict. We design, build and run GPU private clouds from a single server to full racks, on your premises or in a UK data centre.
A GPU private cloud gives your teams the self-service experience of a public cloud (pick a model, start a notebook, deploy an assistant) on GPU servers that only you use. The hardware sits on your own premises or in a colocation data centre, and the platform software, security policies and data all stay under your control.
Public cloud is excellent for trying ideas and covering short peaks. But AI assistants and agents run all day, every interaction is metered, and the data they need is often the data you can least afford to move. Once usage is steady, a private cloud usually makes costs predictable, keeps sensitive data in one place and cuts the latency between your models and your systems.
It is not one product. A private AI cloud is a stack: power and cooling, GPU servers, a fast network, storage that can keep GPUs busy, a container and scheduling platform, model serving and data tools, and the governance around them. The rest of this page goes through each layer, then lets you try the options for your own workloads.
The reasons we hear most often, and what each one means in practice.
Assistants, agents and APIs that run 24/7 cost the same each month on hardware you own or lease, instead of a bill that grows with every token and every user.
Keep customer, patient, legal or financial data in the UK, on your own site, or completely offline, and prove it to auditors and regulators.
Models run beside your databases, documents and applications, so retrieval and agent tool calls don't cross the internet on every step.
Run open-weight models, your fine-tuned models and commercial ones side by side, and change them without re-platforming. No lock-in to one provider's catalogue.
Single sign-on, role-based access, quotas per team, audit logs of every model call and policies on which data each assistant can reach.
Data scientists start notebooks and fine-tuning jobs, developers deploy model endpoints, and IT keeps one console for capacity, health and cost.
Your GPUs are there when a project starts, not sold out in the region you need. Scheduling decides which team gets them, not a queue at a provider.
Older GPUs move to inference, development and teaching when new ones arrive, so each generation keeps earning for years.
Played from each manufacturer's own YouTube channel. Each video loads only when you press play.
GPU memory (VRAM) decides which models fit and how many users each GPU can serve; memory bandwidth decides how fast it answers. These are the data-centre GPUs we build private clouds with most often.
| GPU | Maker | Memory per GPU | মেমরি ব্যান্ডউইথ | Form | Good for |
|---|---|---|---|---|---|
| RTX PRO 6000 Blackwell Server Edition | NVIDIA | 96 GB GDDR7 | 1.6 TB/s | PCIe card, air cooled | Inference, RAG, fine-tuning small and medium models, graphics and simulation |
| L40S | NVIDIA | 48 GB GDDR6 | 864 GB/s | PCIe card, air cooled | Inference of small models, vision, rendering |
| H200 | NVIDIA | 141 GB HBM3e | 4.8 TB/s | 8-GPU HGX board | Large-model inference and fine-tuning |
| B200 | NVIDIA | 180 GB HBM3e | 8 TB/s | 8-GPU HGX board | Training and high-throughput inference |
| B300 (Blackwell Ultra) | NVIDIA | 288 GB HBM3e | 8 TB/s | 8-GPU HGX board, air or liquid | Reasoning models, very large models, training |
| GB300 NVL72 | NVIDIA | 288 GB HBM3e per GPU, 72 GPUs per rack | 8 TB/s per GPU | Liquid-cooled rack | Frontier training and the largest inference services |
| Instinct MI325X | AMD | 256 GB HBM3E | 6 TB/s | 8-GPU OAM platform | Large-model inference with lots of memory per GPU |
| Instinct MI355X | AMD | 288 GB HBM3E | 8 TB/s | 8-GPU OAM platform, liquid | Training and inference of very large models |
| DGX Spark (GB10) | NVIDIA | 128 GB unified | 273 GB/s | Desk-side system | Development, prototyping and pilots before the cluster |
Figures are the manufacturers' published specifications. Partner servers may run GPUs at different power limits.
Two to eight PCIe GPUs such as RTX PRO 6000 or L40S in a standard air-cooled server, for example Dell PowerEdge XE7745 or R760xa. The easiest way into a private cloud: fits most racks and power budgets.
Eight GPUs joined by NVLink or AMD Infinity Fabric, for example Dell PowerEdge XE9680 (H200 or AMD Instinct MI300X), XE9780 (B300) and XE9785 (AMD Instinct MI355X), or NVIDIA DGX B200 and DGX B300.
Whole liquid-cooled racks that act as one giant GPU, such as NVIDIA GB300 NVL72 or AMD Helios. For organisations training or serving the largest models, in a data centre built for them.
DGX Spark and AI workstations for developers and pilots. The same software stack as the cluster, so work moves across unchanged.
A quick way to size GPU memory before you talk to anyone. Use the sizer further down for your own numbers.
| Model size | Precision | Memory for weights | Fits on (with room for users) |
|---|---|---|---|
| 8 billion parameters | 16-bit | about 16 GB | One L40S or RTX PRO 6000; DGX Spark |
| 32 billion parameters | 8-bit | about 32 GB | One RTX PRO 6000 |
| 70 billion parameters | 8-bit | about 70 GB | One H200, or two RTX PRO 6000 |
| 120 billion parameters (MoE) | 4-bit | about 60 to 65 GB | One RTX PRO 6000 or H200 |
| 405 billion parameters | 8-bit | about 405 GB | One 8-GPU H200, B200 or MI325X server |
| 670 billion parameters and up (MoE) | 8-bit | about 670 GB or more | One 8-GPU B300 or MI355X server, or several servers |
GPUs that wait for data waste money. AI storage is usually three tiers, sized for the read speeds training needs and the capacity your data needs.
Fast scratch space for datasets, checkpoints and model caches, close to the GPUs. Typically several NVMe drives per GPU server.
NVMe SSDs in each GPU server
One namespace all servers read in parallel, fast enough to feed training and to load models in seconds. Look for support for GPUDirect Storage, which moves data straight into GPU memory.
Dell PowerScale, VAST Data, WEKA, DDN, IBM Storage Scale, Lustre
Large, low-cost capacity for raw data, datasets, model versions and backups, using the S3 interface most AI tools already speak.
Dell ECS or ObjectScale, Ceph, S3-compatible appliances, tape or cloud for off-site copies
Assistants that answer from your documents need a vector database and the original files, with the same access controls as the source systems. For example pgvector, Milvus, Qdrant or Elasticsearch.
Inside a server, GPUs talk over NVLink or Infinity Fabric. Between servers, the network decides whether many GPUs can work as one.
For training or serving across servers: NVIDIA Quantum InfiniBand (400 to 800 Gb/s per port), NVIDIA Spectrum-X Ethernet, or Ethernet with AMD Pensando Pollara AI NICs. Usually one 400 or 800 Gb/s port per GPU.
A separate Ethernet network for storage traffic, users and applications, typically 100 to 400 Gb/s per server, so it never competes with the compute fabric.
Out-of-band management for every server, DPUs such as NVIDIA BlueField to isolate tenants and offload security, and firewalls between the AI cloud and the rest of the business.
Our checklist for a private AI cloud that teams actually use. We tick these off at design, build and handover.
The usual trade-offs. Most organisations end up hybrid: a private base for steady work and the public cloud or leased capacity for peaks.
| Public cloud AI services | GPU private cloud in a colocation data centre | GPU private cloud on your premises | Hybrid | |
|---|---|---|---|---|
| How you pay | Per token, per hour or per user, monthly | Hardware bought, financed or leased, plus rack space and power | Hardware bought, financed or leased; your own power and cooling | A private base plus metered cloud for peaks |
| Best when | Trying ideas, spiky or short workloads | Steady AI use and you lack a suitable server room | Steady use, strict data rules and a site with power and cooling | Steady base load with occasional big peaks |
| Where data lives | The provider's regions | A UK data centre you choose | Your own building, can be air-gapped | Sensitive data private, the rest where it's cheapest |
| Time to start | Minutes | Weeks to months: hardware lead times and installation | Weeks to months, longer if the site needs power or cooling work | Start in the cloud, move steady work in as hardware arrives |
| Cost as usage grows | Rises with every user and every call | Flat until you need more GPUs | Flat until you need more GPUs | Flat base, metered peaks |
| Latency to your systems | Over the internet or a private link | Low, with a private connection to your sites | Lowest, on your own network | Low for private workloads |
| Model and tool choice | The provider's catalogue and limits | Anything you can run | Anything you can run | Both |
| Skills needed | Cloud and application skills | Platform and GPU operations, or a managed service | Platform, GPU and facilities operations, or a managed service | Both, plus workload placement |
| GPU availability | Newest GPUs often sold out or rationed | Guaranteed once installed | Guaranteed once installed | Guaranteed base, best-effort peaks |
| Hardware refresh | The provider's choice | Your choice, every 3 to 5 years; older GPUs move to lighter work | Your choice, every 3 to 5 years | Your choice for the base |
The eight stages we take every private cloud through, from the first workshop to a platform your teams use every day.
List the assistants, agents, searches and training jobs you want to run, the models, the users and the data. We turn that into GPUs, VRAM, storage and network.
You decide: Which workloads come first, and what success looks like
Your own server room, your data centre or a UK colocation hall. We survey power, cooling, floor loading and connectivity, or find colocation space built for GPU racks.
You decide: On-premises or colocation
GPU servers (NVIDIA or AMD, Dell or NVIDIA DGX), CPU nodes, storage tiers, compute fabric and management network, with spares and room to grow.
You decide: GPU family, vendor and how much growth to build in
Buy outright, spread the cost with a finance lease or lease to own, or lease some capacity while your own hardware arrives.
You decide: Capital, finance lease or leased capacity
Delivery, racking, cabling and power-up, then firmware baselines, fabric tests, storage benchmarks and a multi-day burn-in of every GPU.
You decide: Acceptance tests and dates
Kubernetes and the GPU operators, schedulers, GPU sharing, image registry, identity, secrets and monitoring, all as code so it can be rebuilt.
You decide: Kubernetes, Slurm or both
Model serving, the model catalogue, a staff assistant, RAG over your documents, notebooks and fine-tuning, connected to your systems.
You decide: Approved models and which data each service can use
Runbooks, training for your team, cost and usage reporting, and either your own operations or our managed service.
You decide: Who runs it day to day
Four ways to explore the options for your own AI workloads. Nothing you enter leaves your browser unless you choose to send it to us.
Six questions, one per stop. At the end you get a starting design: hardware, VRAM, storage, network, power and the platform, plus where to go next.
1 of 6
এআই ডেটা সেন্টার · GPU ক্লাস্টার · GPU লিজিং · প্রাইভেট AI · AI Factory · MLOps সেটআপ
Tell us the workloads, users and data rules, or send your build journey answers, and we'll come back with a design, a dated plan and the cost of buying, financing or leasing it.