Skip to main content
BSS/OSS Academy

4.1 · X-as-a-Service Landscape

From data centre infrastructure to commercial AI services

A building full of GPUs earns nothing until it is sold as something: space and power, GPU-hours, a cluster, a model behind an API. This chapter shows the service models, what each one stands on, how each is charged and what a provider must run to sell it, and the gap between running a data centre and operating an AI cloud.

A · Service Landscape

Sixteen ways to sell AI infrastructure

Each card is a service model: what it is, who buys it, what it needs underneath and how it is charged. The names overlap and several are not standard terms; the badge on each card says how settled the term is. Token-based inference in particular is a way of charging for hosted models, not a separate technology.

16 of 16 services · Compare up to 3

  • Data Centre as a Service

    DCaaS
    Established category

    Space, power, cooling, physical security and connectivity for the customer's own servers.

    Who buys it
    GPU clouds, hyperscalers, AI labs and enterprises that own their hardware.
    Typical use
    A GPU cloud signs for 5 megawatts in a new hall and installs its own racks.
    Needs underneath
    Physical data centre · Billing and metering
    Charged
    Per kilowatt per month · Per rack per month · Reserved capacity / committed usage
    Provider runs
    The building, power and cooling plant, physical security and the meet-me room. Not the servers.
    Offered by (verified)
    None of the eight companies in Real-World Examples lists it on its own pages.
    Related
    Bare Metal as a Service · GPU as a Service · Sovereign AI as a Service

  • GPU as a Service

    GPUaaS
    Established category

    GPU compute rented on demand or reserved, usually as virtual machines with one to eight GPUs.

    Who buys it
    AI start-ups, researchers and enterprise teams that need GPUs without buying them.
    Typical use
    A team rents eight GPUs for a week to test a model, then releases them.
    Needs underneath
    Physical data centre · GPU servers · Storage · Customer portal · Billing and metering
    Charged
    Per GPU-hour · Reserved capacity / committed usage
    Provider runs
    Hardware, hypervisor and isolation, images and drivers, capacity, and the portal and billing around them.
    Offered by (verified)
    NVIDIA DGX Cloud, CoreWeave, Lambda, Nebius, Crusoe, AWS, Microsoft Azure, Google Cloud
    Related
    Bare Metal as a Service · Cluster as a Service · Kubernetes as a Service

  • Bare Metal as a Service

    BMaaS
    Established category

    Whole physical servers, dedicated to one customer, with no virtualisation layer.

    Who buys it
    Teams that want every bit of performance, their own operating system or full control of drivers.
    Typical use
    An AI lab takes thirty-two dedicated eight-GPU servers and installs its own software stack.
    Needs underneath
    Physical data centre · GPU servers · Customer portal · Billing and metering
    Charged
    Per server-hour · Reserved capacity / committed usage
    Provider runs
    The hardware, its firmware, the network to it and its replacement when it fails. The customer runs the software.
    Offered by (verified)
    CoreWeave
    Related
    GPU as a Service · Cluster as a Service · Data Centre as a Service

  • Cluster as a Service

    Label in use, not standardised

    Many GPU servers joined by a high-speed fabric and shared storage, handed over as one managed cluster.

    Who buys it
    AI labs and large enterprises training or serving models across many servers.
    Typical use
    A lab reserves 1,024 GPUs on one InfiniBand fabric for six months to train a model.
    Needs underneath
    Physical data centre · GPU servers · High-performance networking · Storage · Customer portal · Billing and metering
    Charged
    Reserved capacity / committed usage · Per cluster-hour
    Provider runs
    Servers, fabric, storage, cluster manager and scheduler, and the health checks that pull failing nodes out.
    Offered by (verified)
    NVIDIA DGX Cloud, CoreWeave, Lambda, Nebius, Crusoe, AWS, Microsoft Azure, Google Cloud
    Related
    Bare Metal as a Service · Kubernetes as a Service · Training as a Service

  • Kubernetes as a Service

    KaaS
    Established category

    A managed Kubernetes environment with GPU nodes, where the provider runs the control plane.

    Who buys it
    Platform teams that package their AI workloads as containers.
    Typical use
    An enterprise runs its inference services and batch jobs on a managed GPU Kubernetes cluster.
    Needs underneath
    Physical data centre · GPU servers · Storage · Kubernetes platform · Customer portal · Billing and metering
    Charged
    Per GPU-hour · Subscription · Reserved capacity / committed usage
    Provider runs
    The Kubernetes control plane, GPU drivers and device plugins, node pools and upgrades.
    Offered by (verified)
    CoreWeave, Lambda, Nebius, Crusoe, AWS, Microsoft Azure, Google Cloud
    Related
    GPU as a Service · Cluster as a Service · MLOps as a Service · AI Platform as a Service

  • AI Platform as a Service

    Established category

    A managed environment to build, train, deploy and run models: notebooks, pipelines, registries and endpoints in one place.

    Who buys it
    Data science and engineering teams that want tools, not infrastructure.
    Typical use
    A bank's data scientists build and deploy a fraud model without touching servers.
    Needs underneath
    Physical data centre · GPU servers · Storage · Kubernetes platform · Model serving platform · Customer portal · Billing and metering
    Charged
    Per GPU-hour · Subscription · Per API request
    Provider runs
    The whole toolchain and the compute under it, its upgrades and its integration.
    Offered by (verified)
    NVIDIA DGX Cloud, AWS, Microsoft Azure, Google Cloud
    Related
    MLOps as a Service · Training as a Service · Inference as a Service · Kubernetes as a Service

  • Model as a Service

    MaaS
    Label in use, not standardised

    Hosted AI models the customer calls through an API, without managing any GPUs.

    Who buys it
    Developers and businesses building applications on existing models.
    Typical use
    A support team adds a hosted language model to its help desk through an API.
    Needs underneath
    Physical data centre · GPU servers · Storage · Model serving platform · Customer portal · Billing and metering
    Charged
    Per million input tokens · Per million output tokens · Per API request · Reserved capacity / committed usage
    Provider runs
    The models, the serving fleet, capacity, the API and its rate limits.
    Offered by (verified)
    NVIDIA DGX Cloud, CoreWeave, Nebius, Crusoe, AWS, Microsoft Azure, Google Cloud
    Related
    Token-based Inference Service · Inference as a Service · Fine-Tuning as a Service · RAG as a Service

  • Token-based Inference Service

    A charging model, not a technology

    Model usage billed by the token: the customer pays for the text sent in and generated, not for GPUs or time.

    Who buys it
    Anyone calling a hosted model whose usage rises and falls.
    Typical use
    An app pays for 500 million input and 100 million output tokens in a month, and nothing for idle time.
    Needs underneath
    Physical data centre · GPU servers · Storage · Model serving platform · Customer portal · Billing and metering
    Charged
    Per million input tokens · Per million output tokens · Reserved capacity / committed usage
    Provider runs
    Everything Model as a Service needs, plus accurate token counting per customer and per model.
    Offered by (verified)
    CoreWeave, Nebius, AWS, Microsoft Azure, Google Cloud
    Related
    Model as a Service · Inference as a Service

  • Inference as a Service

    Established category

    Managed endpoints that run a model, the provider's or the customer's own, and answer requests at an agreed capacity.

    Who buys it
    Companies putting their own or fine-tuned models into production.
    Typical use
    A company deploys its fine-tuned model on two dedicated GPU endpoints that scale with traffic.
    Needs underneath
    Physical data centre · GPU servers · Storage · Model serving platform · Customer portal · Billing and metering
    Charged
    Per GPU-hour · Per million input tokens · Per million output tokens · Reserved capacity / committed usage
    Provider runs
    Endpoint deployment, autoscaling, serving engines and the GPUs under them.
    Offered by (verified)
    NVIDIA DGX Cloud, CoreWeave, Nebius, Crusoe, AWS, Microsoft Azure, Google Cloud
    Related
    Model as a Service · Token-based Inference Service · Fine-Tuning as a Service · MLOps as a Service

  • Training as a Service

    Established category

    Managed training runs: the customer supplies code and data, the provider supplies and runs the cluster.

    Who buys it
    Teams training models from scratch or at large scale.
    Typical use
    A research group submits a training job that runs on 256 GPUs for ten days.
    Needs underneath
    Physical data centre · GPU servers · High-performance networking · Storage · Customer portal · Billing and metering
    Charged
    Per GPU-hour · Per training job · Reserved capacity / committed usage
    Provider runs
    The training cluster, job scheduling, checkpoint storage and restarts after failures.
    Offered by (verified)
    NVIDIA DGX Cloud, Nebius, AWS, Microsoft Azure, Google Cloud
    Related
    Cluster as a Service · Fine-Tuning as a Service · AI Platform as a Service · Storage as a Service

  • Fine-Tuning as a Service

    Established category

    Adapting an existing model to the customer's data, tone or task, without training one from scratch.

    Who buys it
    Enterprises that need a model to follow their terminology, formats or domain.
    Typical use
    An insurer fine-tunes an open model on its claims documents, then serves it as a private endpoint.
    Needs underneath
    Physical data centre · GPU servers · Storage · Model serving platform · Customer portal · Billing and metering
    Charged
    Per training job · Per million input tokens · Per GPU-hour
    Provider runs
    The tuning pipeline, the GPUs it runs on and, usually, hosting of the tuned model.
    Offered by (verified)
    NVIDIA DGX Cloud, CoreWeave, Nebius, Crusoe, AWS, Microsoft Azure, Google Cloud
    Related
    Training as a Service · Inference as a Service · Model as a Service · MLOps as a Service

  • RAG as a Service

    Label in use, not standardised

    A managed service that searches the customer's documents and passes the relevant passages to a model with each question.

    Who buys it
    Enterprises that want answers grounded in their own documents.
    Typical use
    A service desk asks questions of 50,000 internal procedures and gets answers with references.
    Needs underneath
    Physical data centre · GPU servers · Storage · Model serving platform · Customer portal · Billing and metering
    Charged
    Per million input tokens · Per million output tokens · Per terabyte per month · Per API request
    Provider runs
    Document ingestion, embedding, the vector index, retrieval and the model call.
    Offered by (verified)
    AWS, Microsoft Azure, Google Cloud
    Related
    Model as a Service · AI Agents as a Service · Storage as a Service

  • AI Agents as a Service

    Label in use, not standardised

    Hosted agents that plan steps, call tools and systems, and run multi-step workflows on the customer's behalf.

    Who buys it
    Enterprises automating workflows that span several systems.
    Typical use
    An agent reads a customer email, checks the order system, drafts a reply and opens a ticket.
    Needs underneath
    Physical data centre · GPU servers · Storage · Kubernetes platform · Model serving platform · Customer portal · Billing and metering
    Charged
    Per million input tokens · Per million output tokens · Per API request · Subscription
    Provider runs
    The agent runtime, tool connectors, memory, the model calls and logs of every action.
    Offered by (verified)
    AWS, Microsoft Azure, Google Cloud
    Related
    RAG as a Service · Model as a Service · AI Platform as a Service

  • Storage as a Service

    STaaS
    Established category

    High-performance file and object storage for datasets, checkpoints and model weights, close to the GPUs.

    Who buys it
    Anyone training or serving at scale; GPU clouds sell it alongside compute.
    Typical use
    A training cluster reads 2 petabytes of data and writes a checkpoint every hour.
    Needs underneath
    Physical data centre · Storage · Customer portal · Billing and metering
    Charged
    Per terabyte per month · Reserved capacity / committed usage
    Provider runs
    The storage systems, their performance, durability and capacity.
    Offered by (verified)
    CoreWeave, Lambda, Nebius, Crusoe, AWS, Microsoft Azure, Google Cloud
    Related
    Training as a Service · Cluster as a Service · RAG as a Service

  • MLOps as a Service

    Established category

    Managed tooling for a model's life: versioning, pipelines, deployment, monitoring and retirement.

    Who buys it
    Organisations running many models in production.
    Typical use
    A company tracks 40 production models, retrains them monthly and rolls back a bad version in minutes.
    Needs underneath
    Physical data centre · GPU servers · Storage · Kubernetes platform · Customer portal · Billing and metering
    Charged
    Subscription · Per GPU-hour
    Provider runs
    The registry, pipelines, deployment tooling and model monitoring.
    Offered by (verified)
    CoreWeave, Nebius, AWS, Microsoft Azure, Google Cloud
    Related
    AI Platform as a Service · Inference as a Service · Fine-Tuning as a Service

  • Sovereign AI as a Service

    Label in use, not standardised

    AI infrastructure and services run inside one jurisdiction, with defined controls over data, operations and ownership.

    Who buys it
    Governments, regulated industries and national champions that must keep AI and data under local control.
    Typical use
    A public agency uses models and GPUs hosted in-country, operated by locally cleared staff.
    Needs underneath
    Physical data centre · GPU servers · Storage · Model serving platform · Customer portal · Billing and metering
    Charged
    Reserved capacity / committed usage · Per GPU-hour · Per million input tokens · Subscription
    Provider runs
    The underlying services, plus the sovereignty controls: residency, local staff, audit and legal structure.
    Offered by (verified)
    AWS, Microsoft Azure, Google Cloud
    Related
    Data Centre as a Service · GPU as a Service · Model as a Service · Cluster as a Service