Skip to main content
BSS/OSS Academy

Who uses AI infrastructure, and why?

The organisation that uses the AI compute is not always the organisation that owns the infrastructure. An AI company may consume GPU capacity from a hyperscaler or specialist provider, while a telecom operator or sovereign programme may build dedicated infrastructure for its own workloads and make capacity available to customers. This distinction between infrastructure owner, service provider and compute consumer is an important part of understanding how the AI infrastructure ecosystem works.

Hyperscalers & Big Tech

They train foundation models, run inference for their own products and sell AI capacity to cloud customers, so they build at the largest scale.

Not one recipe: alongside NVIDIA GPUs, several design their own accelerators, and they still rent extra capacity when their own build cannot keep up.

Builds & operatesMicrosoft Google (TPUs) AWS (Trainium) Meta (GPUs and MTIA)

Through partnersMicrosoft renting GPU clouds

AI-native companies

Model developers need very large amounts of compute for training, fine-tuning and serving their models at scale.

Most rent rather than own: compute comes through cloud partners and purpose-built joint ventures, with some now building sites of their own.

Builds & operatesxAI (Colossus)

Through partnersOpenAI OpenAI at Stargate Norway

Builds and rentsAnthropic

Tier-1 telecom operators

Sovereign AI, enterprise AI services, AI compute they can sell, AI for their own networks and edge inference close to customers, built on the sites, power, fibre and customer relationships they already have.

Not one model and not one scale: some build and sell their own AI factories, others sell through partners, and some mainly use AI to run their networks.

Two terms you’ll hear: AI factory · AI grid
AI factory
Infrastructure that runs the whole AI life cycle, from data to training to inference. Some operators package it as a platform selling compute and AI services to enterprises, governments and other organisations. A vendor term, not a standard. Source
AI grid
AI capacity spread across many interconnected sites, so inference runs close to users. Operators already run geographically distributed networks, data centres and edge sites, which is their potential advantage. A new vendor term; not the same as AI-RAN. Source

Builds & operatesDeutsche Telekom SoftBank KDDI Singtel TELUS Telenor Telefónica (edge GPUs)

Through partnersAT&T (with Cisco) Vodafone Egypt (with Cassava)

Builds and rentsOrange

Sovereign & national AI

Governments fund AI compute for data and AI sovereignty, research capability, national security and industrial competitiveness.

Cloud infrastructure is a service you rent wherever the provider runs it. Sovereign AI infrastructure adds where the data and compute sit, whose laws govern them, and who can operate them. Funding compute is not the same as owning a data centre.

Publicly fundedEuroHPC AI Factories EU AI gigafactories (announced) UK AI Research Resource Japan’s ABCI

Specialist GPU / AI clouds

Their business is selling GPU capacity: they sit between the infrastructure and the AI workload, renting compute to AI labs, hyperscalers and enterprises.

Unlike a hyperscaler, GPU capacity is the whole offer, not one service among hundreds. Some build their own sites; others operate in leased or partner-run facilities.

Builds & operatesCoreWeave Crusoe

Builds and rentsNscale Lambda

Enterprises & research

Where the requirements justify it: sensitive data, regulation, performance, predictable workloads, research, or not wanting to depend on external capacity.

A minority, and rarely at hyperscale. Most enterprises rent AI capacity rather than build it.

Builds & operatesEli Lilly

Publicly fundedUniversity of Bristol (Isambard-AI)

Each name links to the organisation’s own announcement (or its government’s). Tags say only what that source supports; they are on record, not yet independently verified. Names, not scale: no figures are repeated here.

Why they use AI infra

The reasons vary by organisation, but they come down to three drivers: building AI, running it, and scaling and controlling it.

Build AI
Train foundation and specialised models. Fine-tune existing models for an industry, an organisation or a use case. Power research, simulation and other computationally intensive work.
Run AI
Run inference: answer prompts and generate responses, predictions, images and video. Run enterprise workloads on proprietary, sensitive or regulated data. Provide AI services to customers: GPU compute, model hosting and AI platforms.
Scale and control AI
Keep data and workloads within a country or jurisdiction (sovereign AI). Bring AI closer to users and devices with edge and distributed inference. Apply AI to running networks, including AI-RAN. Combine compute, networking, storage, power and cooling to grow efficiently.

What drives each group

  • Hyperscalers
    Primary driver

    AI services and cloud

    Typical workload
    • Training
    • Inference
  • AI-native companies
    Primary driver

    Model development

    Typical workload
    • Training
    • Inference
  • Telecom operators
    Primary driver

    Sovereign AI, AI services, network AI

    Typical workload
    • Inference
    • Enterprise AI
    • Network AI
  • Sovereign programmes
    Primary driver

    National capability

    Typical workload
    • Government
    • Research
    • Industry
  • GPU / AI clouds
    Primary driver

    Selling compute

    Typical workload
    • GPU cloud for others’ AI workloads
  • Enterprises
    Primary driver

    Data, control, performance

    Typical workload
    • Enterprise AI
  • Research
    Primary driver

    Scientific computing

    Typical workload
    • Training
    • HPC
    • Simulation

High level, and not universal: individual organisations in each group differ.