NeoClouds

NeoClouds are specialist cloud providers focused on AI compute and the services needed to use it. Buying capacity involves more than selecting a GPU: storage, networking, scheduling, support and the contract determine whether a workload can run as intended.

This topic focuses on AI infrastructure and ways to obtain it. It does not treat chip manufacturers, ordinary application software or a model-only API as interchangeable infrastructure providers. A marketplace can give access to several clouds without operating every underlying cluster itself.

Use the branches below to understand the operating model. Use the companion assessment to compare its evaluated providers. Neither a category label nor advertised GPU availability is a guarantee of capacity for your workload.

AI cloud platforms

Compute offered with a broader operating environment: storage, networking, cluster orchestration and management services. The buyer must still determine which responsibilities remain with its own team.

Self-service GPU specialists

Developer-oriented access to GPU instances, containers or serverless workers. These are useful models to evaluate for experiments and variable workloads; check isolation, persistence, limits and production support rather than assuming all configurations are equivalent.

Managed or dedicated capacity

A defined cluster or capacity commitment with an agreed delivery and operating arrangement. The important questions are who operates each layer, when capacity becomes usable, what is reserved and what the contract guarantees.

Marketplaces and aggregators

A way to find, procure or manage capacity across multiple underlying suppliers. Distinguish the marketplace’s commercial and management role from the infrastructure operator’s service responsibilities.

These models overlap. One company may offer on-demand access and dedicated commitments. Sovereignty, location and private deployment are evaluation dimensions across the models—not proof that a provider belongs to a separate market.

Cloud platform examples

CoreWeave describes compute, storage, networking and managed Kubernetes services. Nebius documents GPU-enabled virtual machines and cluster connectivity. Compare the offered environment and contract for the intended workload, not only the accelerator label.

Self-service examples

Runpod Pods provide configurable compute environments. Runpod Serverless uses workers and endpoints that can scale with requests. These are distinct operating choices within one provider; serverless also introduces startup and lifecycle considerations.

Capacity aggregation

Shadeform provides a common interface for comparing, launching and managing GPU instances from multiple providers. An aggregator and an underlying cloud can therefore be parts of the same purchase rather than direct substitutes.

Dedicated infrastructure

Fluidstack describes building and operating large compute infrastructure. A dedicated-capacity discussion requires explicit delivery milestones and division of operational responsibilities; the website is not evidence of capacity available to a specific buyer.

Examples illustrate operating roles, not a new ranking. Software schedulers, monitoring and data services can complement infrastructure. Two compute offers are meaningful alternatives only when both satisfy the same workload, location and service requirements.

These are planning scenarios, not claims of measured savings or performance.

Train or fine-tune a model

Run a defined training job with the required accelerators, interconnect and dataset throughput. Evaluate completed-job time, failure recovery and the total reserved capacity commitment.

Serve an application

Host inference for a product or internal service. Test latency, concurrency, availability and the cost of idle capacity as well as the cost of active requests.

Run development and batch work

Give a team access to GPUs for experiments, evaluation or scheduled processing. Check environment reproducibility, storage persistence and whether interrupted work can resume.

Add capacity beyond an existing estate

Place a suitable workload with another provider when existing infrastructure cannot meet the required schedule. Account for data movement, identity integration and the operational burden of a second environment.

• Capacity: Which hardware, quantity and location are contractually available, on what date, and for how long?
• Workload fit: What memory, interconnect, storage throughput and software environment does the actual job require?
• Service boundary: Who operates the scheduler, containers, drivers, backups and incident response?
• Data and control: Where are data, logs and backups stored? Who can access them, and what residency or isolation commitments are documented?
• Reliability: What happens when a node fails, a job is interrupted or on-demand capacity is unavailable? What do the service terms cover?
• Economics and exit: Include storage, networking, idle reservations, support and migration. Compare a completed workload, not an isolated hourly GPU price.

• NeoClouds Institute Quadrant — provisional comparative assessment; provider briefings, customer evidence and specified verification remain pending.
• NeoClouds 2026: Buyer’s Guide — October 2026 companion guide covering operating models, procurement questions and provider profiles. Read its confidence and evidence limitations.
• Measure AI Economics at the Workflow Level — draft explainer for review; a method for evaluating economics, not observed savings.

Topic primer · Updated October 23, 2026. The linked primary documentation supports descriptions of provider offerings and operating models. This page does not independently verify delivery capacity, benchmark performance, compliance claims or savings.

The branches are a practical guide to buying models, not mutually exclusive company classifications. The existing NeoClouds assessment remains provisional; this introductory page does not increase the confidence of its scores or replace its evidence notes.