What they sell
How customers buy compute from them.
Bare-metal GPUs
A bare-metal GPU server is a whole physical machine rented to one customer, with no hypervisor, the software layer that splits a server into virtual machines. The customer gets every GPU, all the memory and direct access to the network cards
GPU virtual machines
A GPU virtual machine (VM) is a slice of a server, with one or more GPUs attached, that a customer can start in minutes and pay for by the hour or second. It is the standard way clouds sell compute
Training clusters
A training cluster is a block of hundreds to hundreds of thousands of GPUs joined on one fast network, reserved by one customer for months or years. This is how frontier models are trained, and it is the core business of the largest neoclouds
Managed Kubernetes
Kubernetes is open-source software, first released by Google in 2014, that runs programs in containers across many servers and restarts them when they fail. Most AI clouds now offer it as a managed service tuned for GPUs, so customers can run training and inference without managing the machines underneath
Serverless GPUs
Serverless GPUs let customers run code on a GPU without choosing or managing a server. They upload a model or a function, the platform starts a GPU when a request comes in and bills by the second, and when no requests come the cost falls to zero
Model APIs
A model API sells the output of a model rather than the hardware: the customer sends text or images and pays per million tokens. Inference platforms, chip clouds and hyperscalers all serve open models this way, from Llama and Qwen to DeepSeek and Mistral
Colocation and hosting
Colocation means renting space, power and cooling in a data centre for someone else's servers. In AI it usually means a landlord builds or converts a hall and leases it, often for ten years or more, to a neocloud or hyperscaler that brings its own GPUs
Marketplace
A compute marketplace sells GPU time from many suppliers through one place, with prices set by supply and demand or by auction. Buyers choose by price, GPU type, location and reliability
Research allocation
Research allocation is compute granted rather than sold. National supercomputing centres and programmes give time on their machines to scientists, and increasingly to AI start-ups, through calls for proposals judged on merit