An H100 or B200 GPU is the same silicon wherever it is installed, but what a buyer actually receives varies with the supplier: the surrounding network fabric, storage, support, contract length, the chance of interruption and the legal jurisdiction of the data. Uptime Institute's analysis of the market found large price gaps between provider types for comparable Nvidia instances, and noted that the choice between renting and owning depends heavily on expected utilisation. [1]

Grouping suppliers into types is a simplification, since many companies straddle categories. It is still useful, because each type tends to share a business model, and the business model shapes how capacity is priced, allocated and supported.

AWS, Microsoft Azure, Google Cloud and Oracle Cloud Infrastructure sell GPUs as one part of large general-purpose platforms. AWS, for example, offers instances with eight Nvidia H100 or H200 GPUs, connects them with up to 3,200 Gbps of its EFA networking, and places them in UltraClusters that can scale to 20,000 GPUs. [2]

Their strengths are breadth and integration. GPU instances sit next to managed storage, databases, identity management and compliance tooling that many enterprises already use, and they can be bought through a range of purchasing options, from per-second on-demand billing to multi-year commitments and reservations of GPU clusters for fixed dates. [3]

The trade-off, according to Uptime Institute, is price: its February 2025 comparison found hyperscaler list prices for a comparable H100 instance roughly three times those of neoclouds, a gap it attributed mainly to margin. Hyperscalers suit organisations that value platform integration, existing enterprise agreements and global reach more than the lowest hourly rate. [1]

Neoclouds are cloud providers built mainly around GPU compute for AI. Equinix lists CoreWeave, Crusoe, Denvr Dataworks, GroqCloud, Lambda and Nebius among them and notes that they sell GPU capacity with transparent hourly pricing as well as broader AI services. They typically offer high-speed interconnects and cluster schedulers suited to distributed training. [4]

Neoclouds range from self-service clouds where a developer can launch a single GPU to providers whose revenue comes mostly from multi-year contracts with a few very large customers. That second model means the capacity visible to smaller buyers can be a fraction of a provider's total fleet. Our explainer on neoclouds covers their business models and financing in more depth.

GPU marketplaces do not own most of the hardware they list. On Vast.ai, anyone with Linux experience can register as a host by installing the platform's daemon, while operators that meet minimum server counts and hold third-party certifications such as ISO 27001 can qualify as datacenter hosts. Hosts set their own GPU, storage and reservation prices, and the platform tracks machine reliability. [5]

TensorDock runs a more curated model: hosts must pass vetting and meet data-center-grade requirements such as redundant power and connectivity, static IP addresses and minimum RAM, CPU and NVMe storage per GPU. Customers receive isolated virtual machines on host hardware, and TensorDock keeps a 25 percent revenue share. [7]

Marketplaces often list consumer and workstation GPUs alongside data-center parts and can offer interruptible, bid-based rentals. Vast.ai's interruptible instances, for example, are stopped when another user outbids them. That makes marketplaces well suited to experimentation, batch jobs and inference on smaller models, and less suited to workloads that need uniform hardware, strict security controls or large tightly coupled clusters. [6]

Some buyers need data and operations to stay within a jurisdiction. Global providers have responded with separated offerings: in January 2026 AWS launched its European Sovereign Cloud in Brandenburg, Germany, describing it as physically and logically separate from other AWS regions, run by EU-resident staff and keeping metadata such as billing and identity data in the EU. [8]

National operators are building GPU capacity too. Deutsche Telekom and Nvidia announced an Industrial AI Cloud in Munich with up to 10,000 Blackwell GPUs, aimed at German and European industry and public bodies and scheduled to launch in the first quarter of 2026. In the public sector, the EuroHPC Joint Undertaking selected the first seven AI Factory sites in December 2024, linking AI-focused supercomputers to start-ups and researchers. [9][10]

Regional and sovereign offerings are chosen for compliance, procurement and policy reasons more than for price. Access to publicly funded AI Factories generally runs through application calls rather than a credit card, so they complement rather than replace commercial clouds.

Bare-metal providers rent whole physical servers with no hypervisor, so every CPU core, GPU and disk belongs to one tenant. OVHcloud, for example, describes its bare-metal servers as allocating all hardware resources to the user, deployable across 46 data centers, with the customer keeping full administrative control over configuration and security. [11]

Colocation goes a step further: the buyer owns the servers and rents space, power, cooling and connectivity in someone else's facility. It is typically priced per kilowatt per month, and CBRE's June 2026 report shows that availability in major markets is tight, with vacancy of 0.3 percent in Northern Virginia and 2 percent in Singapore, and power supply cited as the main constraint across regions. Many neoclouds themselves operate from colocation space; Equinix notes that the leading neoclouds it lists all have a presence in its facilities. [12][4]

Uptime Institute's cost comparison illustrates the economics: in its model, owning dedicated infrastructure beat hyperscaler rental at about 22 percent utilisation but needed about 66 percent utilisation to beat neocloud rental. Owning pays off only when hardware stays busy for most of its life. [1]

In our view the useful question is not which type of supplier is best but which risks a workload can tolerate: interruption, hardware variability, jurisdiction, lock-in or capital outlay. Once those are clear, the relevant set of suppliers usually narrows quickly, and price comparison becomes meaningful within that set.

Kovara tracks public GPU prices and availability signals across roughly 100 providers of every type described here. The providers directory profiles the suppliers Kovara tracks, the regions view shows where capacity is listed, and the GPU prices page compares the same accelerator across hyperscalers, neoclouds and marketplaces. Kova can explain the differences for a specific workload, and buyers who need reserved or dedicated capacity can request a quote.