What is a neocloud?

A neocloud is a cloud provider built specifically for GPU compute and AI workloads rather than general-purpose enterprise applications. Where a traditional cloud platform offers hundreds of services spanning databases, identity, and analytics, a neocloud concentrates on one thing: getting large numbers of high-end GPUs into a founder's hands quickly, at a predictable price.

The category is recent. SemiAnalysis published its AI Neocloud Playbook and Anatomy in October 2024, which gave the emerging group of GPU-first providers a shared name, and the term moved into common use through 2025 as demand for AI compute outpaced what established platforms could supply.

It is no longer a niche. Synergy Research Group sized the neocloud market at more than $25 billion for 2025 and projects it approaching $400 billion by 2031, a compound annual growth rate near 58%. Providers including CoreWeave, Crusoe, Lambda, and Nebius have built substantial businesses on a proposition that barely existed three years ago.

Neoclouds deliver GPU clusters as bare metal. A customer gets a single-tenant physical server with no hypervisor between the workload and the hardware. High-density NVIDIA GPUs are linked inside a server by NVLink, a direct GPU-to-GPU interconnect, while InfiniBand fabrics carry traffic between nodes with the low latency distributed training requires. Pricing is typically flat per GPU-hour, so a founder can calculate the cost of a job before starting it.

Someone must schedule jobs across the cluster and handle nodes that fail mid-run, which on a long training job means checkpointing strategy rather than a restart. Someone must move training data to storage the cluster can read at speed. Someone must configure the network, keep drivers and CUDA versions aligned across nodes, monitor utilization so expensive hardware is not sitting idle, and apply security patches to machines that hold customer data.

On a managed platform most of this arrives as a service you configure. On bare metal it arrives as work, and the cost of that work is rarely modeled. A founder comparing an hourly GPU rate against a managed alternative is comparing one line item while leaving out the engineering time that fills the gap between them.

Hyperscalers sell breadth. Compute, storage, managed databases, identity, security tooling, compliance certifications, and global regions arrive as one platform under one contract. Neoclouds sell depth in a single dimension. The GPU price per hour is generally lower for raw training work, and access is often faster.

Speed of access has been the decisive factor for many teams rather than price. Capacity at the major providers has been genuinely constrained: on Microsoft's FY26 third-quarter earnings call in April 2026, CFO Amy Hood said Azure was expected to remain capacity constrained at least through 2026. Much of the category's growth reflects that pressure rather than a considered preference.

A B2B startup selling to enterprises inherits requirements that have nothing to do with training throughput: security review, compliance evidence, procurement processes, integration with the identity and data systems the customer already runs. Total cost at scale includes the managed services, security engineering, and compliance work a team would otherwise build and maintain themselves.

Hybrid cloud combines private infrastructure, either on-premises or in a dedicated facility, with public cloud resources, usually to keep data under direct control, satisfy residency requirements, or reduce latency. A neocloud is a public cloud. Running training on a neocloud and production on a hyperscaler is multi-cloud: two public providers, no private infrastructure involved.

Model training runs in bursts on rented GPU capacity, and the resulting model weights move to the platform where the product runs. The application, its database, authentication, monitoring, and customer-facing inference sit on the hyperscaler alongside everything the enterprise sales motion depends on. Training is periodic and compute-heavy. Serving is continuous and integration-heavy.

Moving datasets and model artifacts between providers takes time and incurs egress charges. Two providers means two security reviews, two access control models, and two sets of credentials. Debugging spans environments. None of this is prohibitive, and many teams run this way successfully, but it is a deliberate architecture rather than something to arrive at by accident.

Stage: a pre-product company training a model needs GPUs cheaply and immediately, which favors a neocloud, while a company selling to enterprises needs compliance posture, marketplace presence, and commercial support. Workload: training and inference map onto neocloud strengths, while application hosting, databases, identity, and enterprise integration map onto hyperscaler strengths. Cost structure, lock-in, and whether multi-cloud is a deliberate choice complete the list.

List every cost the alternative absorbs that you would otherwise carry. Start with GPU hours, since that is the number both options quote. Then storage and data transfer at the volume your training actually uses. Then the managed services you would stand up yourself: a database, identity, secrets management, monitoring. Then engineering time, priced at what an infrastructure engineer costs. Then compliance. Finally the commercial infrastructure that reaches enterprise buyers.

Run that list and the comparison usually stops being close. A neocloud generally wins on the first line and is not competing on the rest. That is the right outcome for a company whose only current problem is training throughput, and the wrong one for a company that has a product and needs customers. The correct answer at seed stage is frequently the wrong answer at Series A.

The Microsoft commercial marketplace lets a startup list its product in purchasing channel enterprise IT departments already use and already trust. A buyer can transact through an existing Microsoft relationship, with procurement paths and vendor approvals largely settled. That shortens sales cycles without adding sales headcount.

Co-sell gives qualifying startups access to Microsoft enterprise sellers, who introduce products into accounts a small company would struggle to enter on its own. Compliance works the same way: security certifications, privacy controls, and regulatory attestations are expensive and slow to earn independently, and building on a platform that already holds them moves much of that burden off the engineering team.

None of this exists in a neocloud relationship, and that is not a criticism. Neoclouds sell compute. The point is that a B2B startup needs both compute and distribution, and only one of those categories provides both.

A neocloud fits when the primary workload is model training or large-scale inference, when GPU availability is the bottleneck, when cost per GPU-hour is the variable that decides the budget, and when the team has the engineering capacity to run bare-metal infrastructure competently.

A neocloud does not fit when the company needs compliance certifications to close deals, when reaching buyers requires marketplace presence, when managed database or security services would otherwise have to be built internally, or when a single vendor relationship covering the full production stack is worth more than a lower unit price.

Most neocloud providers are venture-backed with short operating histories in a market Forbes described in November 2025 as holding many risks. The GPU access argument also weakens over time: capacity expands, and startup credit programs put substantial AI infrastructure within reach at low cost. Optimizing purely for training cost can leave a company on infrastructure that cannot support the sales motion the business needs a year later.

Microsoft for Startups provides Azure credits in tiers that increase as a company verifies its business and progresses through the program, with the largest offers available to investor-backed startups. Azure provides NVIDIA GPU infrastructure through its ND-family virtual machines, including H100 and H200 configurations, without requiring a team to operate bare metal.

Microsoft Foundry gives access to enterprise-grade models and the tooling to build on them, so a startup already running on Azure can train and deploy without managing a separate neocloud relationship. The broader program extends past infrastructure into technical advisory, go-to-market support, and developer tooling including GitHub Enterprise.

The useful comparison is total cost: GPU hours plus managed services plus compliance plus the commercial infrastructure a company would otherwise assemble separately. Compared on GPU hours alone, neoclouds usually win. Compared on what a B2B startup needs to reach revenue, the calculation looks different.

What is my primary workload right now? Training a model, running inference at scale, and building a production application for enterprise customers point toward different infrastructure. What engineering capacity do I have for bare-metal operations? Managing infrastructure without managed services costs engineering time, and that time should be measured against the GPU savings.

Does my go-to-market require enterprise trust signals in the next 12 to 24 months? Marketplace presence, compliance certifications, and co-sell relationships take time to establish and are difficult to retrofit. What is the total cost of this choice, including compute, managed services, security, support, and the commercial infrastructure that would otherwise need building? And am I solving for today or for the next funding stage?