What does AI actually run on?
A stack of 8 layers, from the person asking down to the power socket - pick one to see who runs it.
ComputeThe processing power
The raw computing capacity - rented by the hour, the month or the year - that trains and runs AI models. This is what "the cloud" mostly means in practice.
See the 0 providersUser
The person
Where everything starts: someone typing a question, using a search box, or opening an app that happens to have AI built into it. Nothing below this layer matters unless it eventually reaches a real person doing something with it.
The technical version
The human, or the downstream system, issuing the request. Not itself infrastructure - but the layer every request originates from and every response is optimised to return to, fast enough to feel instant.
This layer sits on top of the infrastructure rather than being part of it, so it is covered in another part of Nakoda.
See the AI products people actually useApplication
The product
The app, website or feature that wraps an AI model into something usable - a chatbot, a copilot inside other software, a search assistant. This is the layer most people actually mean when they say "AI."
The technical version
The product layer: orchestration, prompting, business logic, guardrails and interface sit here, calling one or more models beneath it rather than running any model itself.
This layer sits on top of the infrastructure rather than being part of it, so it is covered in another part of Nakoda.
Browse the companies building AI applicationsModel
0 providers at this layer
The trained system that actually produces the answer - GPT, Gemini, Claude, Llama and others like them - plus the platforms that make models like these available for developers to build on.
The technical version
Foundation models and the model-hosting layer beneath them: managed access to first- and third-party models (Bedrock, Azure OpenAI Service, Vertex AI), or open-weight model deployment (Hugging Face, Together AI).
Inference
0 providers at this layer
The moment a trained model is actually used to produce an answer, in about the time it takes to read this sentence. Training a model happens once; inference happens every single time someone uses it.
The technical version
Serving pre-trained model weights against live requests, optimised for latency and cost per token rather than for training throughput - the domain of purpose-built inference chips, edge inference and serverless model endpoints.
Compute
0 providers at this layer
The raw computing capacity - rented by the hour, the month or the year - that trains and runs AI models. This is what "the cloud" mostly means in practice.
The technical version
Virtualised and bare-metal compute: CPU and GPU instances, managed clusters, HPC-grade interconnect and reserved capacity, provisioned by hyperscale clouds and specialist GPU clouds alike.
GPU
0 providers at this layer
The specialised chips - mostly made by NVIDIA - that do the actual arithmetic behind AI, built to run millions of small calculations in parallel far faster than a general-purpose computer chip can.
The technical version
Graphics processing units and purpose-built AI accelerators (GPUs, LPUs) whose parallel architecture is what makes training and running deep-learning models computationally feasible at scale.
Data Centre
0 providers at this layer
The physical building - racks, servers, cooling - where everything above this actually lives. An AI model has no existence outside a room full of running machines.
The technical version
The physical facility layer: colocation and hyperscale data centres, their power density and cooling architecture, and the interconnection between them - increasingly what limits how much AI compute a company or country can actually deploy.
Power & Network
0 providers at this layer
Everything above this needs two things nobody thinks about until they run out: electricity to run and cool the machines, and network links fast enough to move data between them.
The technical version
Power generation and grid capacity feeding data centres, plus the fibre, interconnection and edge network moving data between facilities and out to users - increasingly the binding constraint on AI build-out, not chip supply.
What is physically in the UAE
0 of the 0 providers on this page have a documented site in the country.
Not in the UAE
The whole stack as a table
| Layer | What it is | Providers |
|---|---|---|
| User | Where everything starts: someone typing a question, using a search box, or opening an app that happens to have AI built into it. Nothing below this layer matters unless it eventually reaches a real person doing something with it. | Covered elsewhere |
| Application | The app, website or feature that wraps an AI model into something usable - a chatbot, a copilot inside other software, a search assistant. This is the layer most people actually mean when they say "AI." | Covered elsewhere |
| Model | The trained system that actually produces the answer - GPT, Gemini, Claude, Llama and others like them - plus the platforms that make models like these available for developers to build on. | 0 providers |
| Inference | The moment a trained model is actually used to produce an answer, in about the time it takes to read this sentence. Training a model happens once; inference happens every single time someone uses it. | 0 providers |
| Compute | The raw computing capacity - rented by the hour, the month or the year - that trains and runs AI models. This is what "the cloud" mostly means in practice. | 0 providers |
| GPU | The specialised chips - mostly made by NVIDIA - that do the actual arithmetic behind AI, built to run millions of small calculations in parallel far faster than a general-purpose computer chip can. | 0 providers |
| Data Centre | The physical building - racks, servers, cooling - where everything above this actually lives. An AI model has no existence outside a room full of running machines. | 0 providers |
| Power & Network | Everything above this needs two things nobody thinks about until they run out: electricity to run and cool the machines, and network links fast enough to move data between them. | 0 providers |
How this page is checked
- Every provider is a real company, checked against its own site or an industry source.
- Pricing links go only to the provider’s own pricing page.
- A UAE site is shown only where a region or facility is documented - otherwise the page says so.
Choosing between a global cloud and a UAE one?
Tell us what you are deploying and we will help you work through it. Infrastructure news and buying guides are there if you would rather read first.