GPU Cloud Comparison Report: Neoclouds for AI Infrastructure

Executive Summary

This report analyzes 17 GPU cloud providers (neoclouds) across the dimensions that determine production readiness for AI workloads: pricing, networking infrastructure, storage performance, orchestration capabilities, and enterprise compliance. It is written for engineers who self-service into these clouds: sign up, add a card, and provision. Where a capability exists only in a reserved, offtake, or bespoke contract, this report says so rather than presenting it as something a self-service user can buy today.

Report Structure:

This report provides comparative analysis tables across all providers, followed by detailed profiles assessing each provider’s infrastructure, strengths, gaps, and optimal use cases. Recommendations by workload type are provided in the “Choosing a Provider” section.

The Neocloud Landscape

The term “neocloud” refers to cloud providers primarily offering GPU-as-a-Service. Unlike hyperscalers with broad service portfolios, neoclouds focus on delivering GPU compute with high-speed interconnects for AI and HPC workloads.

Between 10-15 neoclouds currently operate at meaningful scale in the US, with footprints growing across Europe, the Middle East, and Asia.

Why consider neoclouds over AWS, GCP, or Azure?

The hyperscaler GPU experience involves quota requests, waitlists, and premium pricing:

AWS remains the cheapest hyperscaler per GPU at $6.88/hr, still roughly 4x the cheapest self-service marketplace rate. GCP and Azure are 6-7x more expensive. Beyond price, neoclouds offer self-service provisioning in minutes vs multi-week quota approval processes.

Beyond cost, neoclouds eliminate the friction of quota approvals. On AWS, requesting H100 quota often requires a support ticket explaining your use case, with approval taking days to weeks. GCP and Azure have similar processes. Neoclouds typically offer self-service access: sign up, add payment, deploy GPUs in minutes.

Infrastructure is also optimized differently. Neoclouds that target training treat InfiniBand or a comparable AI-optimized Ethernet fabric as standard: 400Gb/s per GPU for multi-node work. Hyperscalers charge premium tiers for similar networking (AWS EFA, GCP GPUDirect) and availability varies by region. One caveat carried throughout this report: at several neoclouds the high-speed fabric is only exposed in a reserved or “supercloud” tier, not in the public on-demand product.

Market Segmentation

The GPU cloud market has fragmented into distinct tiers with different business models, capital structures, and target customers. Understanding these tiers helps match your needs to the right provider type.

Tier 1: Bespoke Wholesale

Massive, single-tenant infrastructure projects with multi-year, multi-billion dollar contracts. Providers secure 100MW+ of power and build custom data centers to customer specifications.

Known examples: FluidStack (enterprise), Crusoe (Stargate), CoreWeave (enterprise). Most providers will entertain bespoke deals for sufficiently large customers; these arrangements are rarely disclosed publicly.

Characteristics:

• Physical site control and custom network topologies
• Take-or-pay agreements with guaranteed capacity
• Not publicly documented or priced

Tier 2: Sales-Gated Cloud

Standardized multi-tenant infrastructure that requires sales approval to access. Once approved, you use the provider’s standard cloud platform.

Providers: CoreWeave, TensorWave, Nscale (training clusters)

Characteristics:

• Enterprise-grade infrastructure with SLAs
• Requires organizational approval process
• Volume commitments often required

Tier 3: Self-Service Neoclouds

Dedicated GPU clouds with transparent, on-demand pricing via web console. The core of this report.

Providers: Nebius, Lambda, Crusoe, Vultr, Hyperstack, Verda (formerly DataCrunch), RunPod (Secure Cloud), OVHcloud, Voltage Park, GMI Cloud, Hot Aisle, SF Compute

Characteristics:

• Sign up and deploy in minutes
• Published pricing, no sales conversation required
• Bare metal or containerized access

Tier 4: Marketplaces & Aggregators

Software platforms pooling supply from third-party hardware owners. High variability in quality and uptime, but lowest prices.

Providers: Vast.ai, RunPod (Community Cloud), FluidStack (marketplace)

Characteristics:

• Bidding or spot pricing
• Variable hardware quality and reliability
• Ideal for fault-tolerant batch processing

Note: Many providers span multiple tiers. FluidStack now earns roughly 62% of revenue from private cloud (bespoke, multi-year, $100M+ contracts) and 38% from its marketplace, and has shifted decisively toward the former. RunPod offers both Secure Cloud (Tier 3) and Community Cloud (Tier 4). Crusoe runs self-service cloud while building Stargate for OpenAI. The bespoke wholesale tier is likely larger than publicly known, as these deals are rarely disclosed.

This wholesale/offtake reality matters for self-service buyers in a specific way: a provider’s published partnerships (for example with storage vendors like VAST Data and WEKA, or with networking vendors) often reflect what it builds for large offtake customers, not what its public on-demand cloud exposes. Throughout this report, storage and networking entries distinguish what a self-service user can actually provision from what is contract-only.

GPU Hardware & Pricing

Your first decision: which GPUs do you need, and what will they cost? This section covers on-demand pricing for NVIDIA and AMD GPUs. Most providers offer reserved capacity discounts of 30-60%, but those require sales conversations.

On-Demand GPU Pricing

*Vultr lists two product lines at different rates: cloud GPU instances and bare-metal servers. Its H100 is $1.99/GPU/hr on 24- and 48-month contracts (no separate on-demand rate found). B200 is $1.99/GPU/hr as a 48-month cloud instance or $2.99/GPU/hr as bare metal. Several AMD parts are available on-demand as preemptible instances.

Pricing varies significantly based on whether you’re renting individual GPUs, full nodes (typically 8 GPUs), or multi-node clusters with InfiniBand. Some self-service providers raised published on-demand H100 rates in early 2026 (Lambda and Verda each by roughly 30-40%), partly reversing the decline seen through 2025, while others held flat (Nebius at $2.95). Reserved capacity discounts of 25-60% are available from most providers but require sales conversations. “Contact” indicates the GPU exists in the provider’s roadmap or reserved tier but is not available at a published self-service on-demand rate.

AMD GPU Availability

AMD adoption continues to grow. Vultr offers one of the cheapest MI300X options at $1.75-1.85/hr and made MI355X generally available for self-service in early 2026. Hot Aisle and TensorWave are AMD-only providers; Crusoe extended its KVM virtualization work from MI300X to MI355X (with the AMD Pensando Pollara 400 NIC). The next-generation MI400 series was announced at CES 2026 with production shipments slated for H2 2026; it is not yet available on any self-service cloud.

*As of mid-2026 TensorWave no longer publishes hourly rates on its site and directs buyers to sales. Its MI300X/MI325X/MI355X previously listed around $1.71/$1.95/$2.85 per GPU-hour; treat those as historical and confirm current pricing with TensorWave directly.

Training Infrastructure

For multi-node training, your infrastructure determines actual performance regardless of GPU specs. Network bandwidth between GPUs and shared storage throughput are the critical factors. Without proper networking and storage, even the fastest GPUs will sit idle waiting for data.

InfiniBand and High-Speed Networking

For multi-node distributed training, network bandwidth between GPUs is critical. InfiniBand provides lower latency and higher bandwidth than Ethernet, with RDMA enabling GPU-to-GPU communication without CPU involvement.

Note: This table describes publicly available cloud offerings. Bespoke wholesale buildouts (Tier 1) can support arbitrary network configurations. “Not documented” indicates information not publicly available.

Key observations:

• 400Gb/s NDR InfiniBand is the baseline among providers that target training. Crusoe, Verda, GMI Cloud, Nebius, RunPod (clusters), Vultr, and Voltage Park (bare metal) expose it to self-service users. CoreWeave and FluidStack offer it but gate access behind sales or private-cloud contracts.
• Read the “self-service” column carefully. At Hyperstack, InfiniBand exists only in the reserved Supercloud tier; the public on-demand product is plain Ethernet. At SF Compute, InfiniBand is on Kubernetes nodes that are sales-contact-only, while self-service VMs are Ethernet. Voltage Park’s VM instances fall back to Ethernet; only bare metal gets 400Gb/s InfiniBand.
• Ethernet-based AI fabrics are diverging from legacy RoCE. Nscale’s networking is Ultra Ethernet Consortium (UEC) compliant via Nokia IXR-H6 switches, a newer standard than the RoCE used by TensorWave and Hot Aisle. CoreWeave and Nebius pair InfiniBand with NVIDIA Spectrum-X RoCE on newer Blackwell systems.
• Single-GPU instances typically don’t include a high-speed fabric at Lambda, RunPod, and Hyperstack. You need to provision cluster configurations.

Storage Options

Training workloads need three types of storage: block storage for OS and application data, object storage for datasets and checkpoints, and shared filesystems for multi-node data access. This table describes what a self-service user can actually provision. The “Self-service shared FS” column is deliberately strict: many providers announce VAST Data or WEKA partnerships, but at most of them those are reserved or bespoke configurations, not something exposed in the on-demand product. See the note below the table.

Key observations:

• VAST/WEKA partnerships are mostly an offtake story, not a self-service one. Press releases pairing a neocloud with VAST Data or WEKA usually describe what the provider builds for large reserved or single-tenant customers. Verified self-service exposure exists at CoreWeave (VAST is selectable when you provision a cluster) and Crusoe (VAST-backed shared disks are on the on-demand pricing page). At Nebius, Lambda, GMI Cloud, Hyperstack, Voltage Park, TensorWave, and FluidStack, a self-service user gets the provider’s own object/shared storage, and VAST or WEKA requires a sales conversation or only powers the backend. Plan storage around what you can actually provision.
• Crusoe closed its object-storage gap. The December 2025 edition noted Crusoe had no native object storage; it now offers S3-compatible object storage at $0.06/GiB/mo.
• Object storage gaps remain at Verda (still “coming”), Hot Aisle, SF Compute, Vast.ai, and Nscale.
• Shared filesystem is critical for multi-node training. Without it, you copy data to each node’s local storage or stream from object storage.

Orchestration & Platform

How you’ll actually run workloads matters as much as the hardware. But there’s an important distinction: infrastructure orchestration (Kubernetes, Slurm) vs. the platform layer.

Neoclouds provide Kubernetes or Slurm to schedule containers or jobs on GPU nodes. That is infrastructure orchestration: it gets your code running on hardware. But production AI teams need more: hosted dev environments where data scientists can iterate, distributed training orchestration that handles multi-node configurations, parallel job scheduling with automatic retries, and cost allocation by user and project.

Most neoclouds stop at infrastructure. The platform layer (the operational tooling that makes GPU infrastructure actually usable for teams) is what you build on top, or what Saturn Cloud provides out of the box.

Kubernetes and Orchestration

Key observations:

• Nebius and CoreWeave have the most mature Kubernetes offerings, with GPU-optimized features like pre-installed drivers and topology-aware scheduling.
• Slurm remains popular for HPC-style workloads. Nebius’s Soperator (now generally available as a fully managed, one-click option) is notable as the first open-source Kubernetes operator for running Slurm clusters. CoreWeave’s SUNK supports 32,000+ GPU jobs and added a self-service onboarding flow in 2026. Nscale added Managed Slurm and its NKS Kubernetes service in 2026, closing a gap noted in the previous edition, though self-service vs enterprise parity is not documented.
• Watch for “coming soon” autoscaling. Hyperstack’s managed Kubernetes is available on-demand, but HPA and Cluster Autoscaler were still marked coming-soon as of late 2025. OVHcloud added bidirectional autoscaling including scale-to-zero for idle GPU nodes.
• Serverless/container platforms (RunPod, Vast.ai) trade Kubernetes flexibility for simpler deployment models; both expanded their serverless inference offerings in 2026.

Operational Considerations

Hidden costs and networking capabilities often determine whether a provider works for production deployments. Egress fees can add 20-40% to monthly bills at hyperscalers, while load balancers and VPCs are baseline requirements for inference endpoints.

Egress Pricing

Free egress remains standard among GPU neoclouds. This is a significant differentiator from hyperscalers, where egress runs $0.087-0.12/GB and can add 20-40% to monthly bills for data-intensive workloads.

Network Services

Key observations:

• Managed Direct Connect is rare: Only CoreWeave (Equinix, Megaport) and OVHcloud (OVHcloud Connect) offer managed private connectivity to on-prem or other clouds. Most providers expect you to run your own VPN gateway on a VM.
• Load balancers + VPC is the baseline for production inference: Nebius, CoreWeave, Crusoe, Vultr, and OVHcloud meet this bar.
• Marketplace providers (Vast.ai, RunPod) use port mapping instead of dedicated IPs, which complicates production inference deployments.

Developer Experience & Enterprise Readiness

How easy is it to get started, and does the platform meet enterprise requirements? Terraform providers and APIs enable infrastructure-as-code, self-service access determines time-to-first-GPU, and compliance certifications gate enterprise adoption.

Terraform and API Support

Self-Service Access

Compliance and Enterprise Features

Infrastructure Ownership Models

Understanding whether a provider owns their infrastructure or aggregates from others matters for reliability, support, and pricing stability. See the Market Segmentation section for how this maps to business model tiers.

Choosing a Provider

Provider selection should align with workload requirements and organizational constraints. The following recommendations categorize providers by primary use case.

Production Multi-Node Training

Recommended Providers: Nebius, CoreWeave, Crusoe

Selection Criteria:

• InfiniBand networking on all GPU nodes (not cluster-only)
• Managed Kubernetes with GPU-optimized scheduling
• High-performance shared storage (VAST Data or equivalent)
• Enterprise compliance (SOC 2 Type II minimum)
• Terraform provider for infrastructure-as-code

Provider Differentiation:

• CoreWeave: Large scale (~250K GPUs, 43 data centers), GB300 at scale, and one of the few providers where VAST shared storage is genuinely self-service. Sales-gated, with a large debt load to weigh.
• Nebius: Most complete managed service stack (Kubernetes, Slurm via Soperator, MLflow, PostgreSQL, Spark), backed by a well-capitalized public company. Note that VAST/WEKA are not self-service here.
• Crusoe: Offers AMD GPUs alongside NVIDIA with full enterprise features (SOC 2, managed K8s, Slurm) and now native object storage.

One caution for this category: if your design depends on a VAST or WEKA shared filesystem, confirm it is available on the provider’s self-service product before committing. At most providers it is a reserved or bespoke configuration, not something you can provision on-demand.

Cost-Optimized Workloads

Recommended Providers: SF Compute, Vast.ai

Selection Criteria:

• Low per-GPU pricing (H100 from roughly $1.82/hr on SF Compute, varying with marketplace supply)
• Marketplace models enabling spot pricing
• Flexible reservation windows without long-term contracts

Provider Differentiation:

• SF Compute: Marketplace with flexible time-based reservations at guaranteed prices. Self-service is Ethernet VMs; InfiniBand and Kubernetes are sales-contact only.
• Vast.ai: Pure peer-to-peer marketplace with per-minute billing and highly variable pricing/quality, now with a serverless inference layer.

Trade-offs:

• Variable infrastructure quality (aggregated from multiple underlying hosts/providers)
• Limited high-speed networking and storage on the self-service tier

European Data Sovereignty

Recommended Providers: Nebius, Verda, OVHcloud

Selection Criteria:

• EU-based data center operations
• GDPR compliance
• Renewable energy infrastructure (hydro/geothermal)
• European regulatory certifications

Provider Differentiation:

• Nebius: SOC 2 Type II + HIPAA + ISO 27001, managed Kubernetes and Slurm, Finland/France/Iceland locations
• Verda (ex-DataCrunch): SOC 2 Type II + ISO 27001 suite, Finland/Iceland, genuine carbon-neutral operations and early Blackwell Ultra access
• OVHcloud: SecNumCloud (ANSSI) qualification for sensitive French public-sector workloads, full SecNumCloud IaaS GA June 2026

AMD GPU Access

Recommended Providers: Vultr, TensorWave, Hot Aisle

Selection Criteria:

• AMD Instinct GPU availability (MI300X/MI325X/MI355X)
• ROCm software support
• Competitive pricing vs NVIDIA equivalents

Provider Differentiation:

• Vultr: Among the cheapest MI300X ($1.75-1.85/hr), MI355X now self-service, managed Kubernetes, 33 global locations
• TensorWave: Largest AMD cluster in North America, now with managed K8s and Slurm (pricing is now quote-based; confirm rates with sales)
• Hot Aisle: AMD-exclusive specialist with SOC 2 Type II + HIPAA, simple per-minute MI300X access

Provider Profiles

Each profile below covers infrastructure details, strengths, gaps, and best-fit use cases for the provider.

Nebius

Overview

Nebius spun off from Yandex N.V. in 2024 following Russia-related sanctions pressures. The company repositioned from a search conglomerate to a dedicated AI infrastructure provider, led by Yandex co-founder Arkady Volozh, and trades on Nasdaq as NBIS. In December 2024, Nebius raised $700M from NVIDIA and Accel, followed by a string of larger raises through 2025 and 2026.

Growth has been steep. Nebius reported $399M revenue in Q1 2026 (up 684% year-over-year) and $1.9B ARR, with full-year 2026 guidance of $7-9B ARR. In March 2026 it signed a $27B multi-year agreement with Meta ($12B of dedicated capacity from early 2027 plus $15B of optional compute over five years), on top of its earlier Microsoft deal. NVIDIA invested a further $2.0B in March 2026, and Nebius closed $4.34B in convertible notes the same month, leaving roughly $9.3B in cash.

Infrastructure

Nebius owns data centers in Finland (Mäntsälä) and Iceland, with US colocation. Its Iceland region (eu-north2) is operational but currently a limited-access private region. The New Jersey region remains under construction with delivery expected by end of 2026, and the company has secured 1.2GW of power and land for a Pennsylvania data center.

Hardware: H100 ($2.95/hr), H200 ($3.50/hr), B200 ($5.50/hr), L40S, and GB200 NVL72 (contact sales). Preemptible rates are lower (H100 $1.25/hr, H200 $1.45/hr, B200 $2.90/hr). All GPU nodes include 400Gb/s Quantum-2 InfiniBand with rail-optimized fat-tree topology and 3.2 Tbit/s per-host networking; newer Blackwell systems pair InfiniBand with Spectrum-X RoCEv2.

Storage is a strength on Nebius’s own services: an S3-compatible Object Storage and a Shared Filesystem rated at 500+ GB/s aggregate. Note the self-service caveat: Nebius publishes VAST Data (March 2025) and WEKA (June 2025) partnerships, but neither is documented as a self-service on-demand option. A self-service user gets Nebius’s own object and shared storage; VAST/WEKA appear to be reserved or contract configurations.

Strengths

• Most complete managed service stack among neoclouds: Kubernetes ($0 for control plane), Slurm via Soperator, MLflow, Spark, PostgreSQL
• Managed Soperator is now generally available as a one-click, self-service Slurm-on-Kubernetes option; it was the first fully-featured open-source operator of its kind
• Self-service quota of up to 32 GPUs without contacting support
• Competitive pricing held steady (H100 $2.95/hr on-demand, $1.25/hr preemptible) while several peers raised published rates
• Strong sustainability angle: Mäntsälä facility’s heat recovery covers a large share of local municipality heating
• Deep balance sheet and hyperscaler-scale contracts (Meta, Microsoft, NVIDIA) reduce supply risk

Gaps

• VAST/WEKA high-performance storage is not exposed to self-service users
• US footprint still ramping (New Jersey not yet live); no Asia-Pacific data centers
• Most of the announced capacity is committed to large offtake customers, which can constrain on-demand availability

Best for: Teams wanting a fully-managed platform (Kubernetes + Slurm + MLflow) with strong European presence and the backing of a well-capitalized public company.

CoreWeave

Overview

CoreWeave is among the largest neoclouds by GPU count. Founded in 2017 as Atlantic Crypto (Ethereum mining), the company pivoted to GPU cloud in 2019 and went public on Nasdaq (CRWV) in March 2025. The company runs roughly 250,000 GPUs and, as of May 2026, has expanded to 43 data centers.

Major customers include OpenAI (expanded to ~$22.4B), Microsoft, and a March 2026 $21B commitment from Meta. Q1 2026 revenue was $2.08B (up 112% year-over-year), and the revenue backlog reached $99.4B as of March 31, 2026. The stock has been volatile: it traded around $108 at the end of May 2026 within a 52-week range of roughly $64-187, after a November 2025 guidance cut (driven by third-party data center construction delays) triggered a single-day 30% drop.

Infrastructure

CoreWeave has expanded through organic growth, self-builds, and acquisitions, and is targeting 1.7GW of operational capacity by end of 2026. Locations span the US and Europe (UK, Spain, Sweden, Norway).

CoreWeave remains a first-mover on new NVIDIA architectures and is deploying GB300 NVL72 (Blackwell Ultra) at scale. Its networking is a dual strategy: NVIDIA Quantum InfiniBand (400Gb/s NDR, non-blocking fat-tree) for H100/H200/B200 clusters, and NVIDIA Spectrum-X RoCE with BlueField-3 and ConnectX-8 SuperNICs on GB300 systems.

Unusually among neoclouds, CoreWeave exposes VAST Data as a selectable storage option when provisioning clusters (also WEKA, DDN, IBM Spectrum Scale, Pure Storage), backed by a $1.17B commercial agreement with VAST. S3-compatible object storage runs $0.06/$0.03/$0.015 per GB/mo across hot/warm/cold tiers with free ingress and egress.

Strengths

• Large GPU fleet and first-mover on new NVIDIA architectures (GB300 at scale)
• SUNK (Slurm on Kubernetes) supports 32,000+ GPU jobs with GitOps via ArgoCD; a self-service onboarding flow and CoreWeave Sandboxes arrived in 2026
• Non-blocking fat-tree InfiniBand topology with NVIDIA SHARP (2x effective bandwidth)
• One of the few providers where VAST shared storage is genuinely self-service, not offtake-only
• NVIDIA’s top cloud partner; sole holder of SemiAnalysis ClusterMAX Platinum

Gaps

• Sales-gated access: organizational approval still required despite 2026 self-service additions
• Extreme customer concentration in OpenAI and Microsoft
• Debt load has ballooned to ~$25B (roughly tripled in a year), with $536M of quarterly interest expense in Q1 2026
• Stock fell 30% in November 2025 after a guidance cut tied to construction delays
• Custom configurations and large-scale pricing require sales conversations

Best for: Large enterprises needing massive scale, the latest NVIDIA hardware, and willingness to work through a sales process.

Crusoe

Overview

Crusoe was founded in 2018 with a unique angle: converting stranded natural gas (flared at oil wells) into computational power. Their Digital Flare Mitigation technology captures methane with 99.9% combustion efficiency, reducing emissions by ~99% compared to regular flaring.

The company raised $1.375B in Series E (October 2024) at a $10B+ valuation, with investors including NVIDIA, Mubadala, Founders Fund, Fidelity, and Tiger Global. Disclosed funding estimates vary ($2.64B historical, $3.9B+ per Sacra), and Crusoe has since secured $11.6B in debt and equity for Abilene expansion plus a $175M Iceland credit facility.

In March 2025, Crusoe divested its Bitcoin mining operations to NYDIG (which had been 55% of 2024 revenue) to focus purely on AI infrastructure. It is now the lead developer on the Stargate project’s flagship Abilene campus (OpenAI/Oracle/SoftBank’s $500B AI initiative).

Infrastructure

Crusoe operates across multiple regions with 1.6+ GW under operations or construction and 10+ GW in development. The Abilene, Texas Stargate campus reaches 1.2 GW across eight buildings when Phase 2 completes (on track for mid-2026), each building designed for up to 50,000 GB200 NVL72 GPUs. A multi-gigawatt Wyoming campus (Project Jade) is under development.

European presence includes Iceland (57 MW, geothermal/hydro via atNorth) and Norway (12 MW hydro, expandable to 52 MW). Vertical integration through the 2022 Easter-Owens acquisition gives Crusoe in-house data center design and manufacturing.

Hardware: H100 80GB HGX ($3.90/hr), H200 ($4.29/hr), B200 and GB200 (contact sales), L40S, and AMD MI300X ($3.45/hr). Crusoe was the first major cloud to virtualize AMD MI300X on Linux KVM and has since extended that work to the MI355X with the AMD Pensando Pollara 400 NIC. All NVIDIA GPU instances include 400Gb/s InfiniBand with rail-optimized topology.

Strengths

• Energy-first model provides long-term cost predictability and genuine sustainability credentials
• Vertical integration from power generation through hardware to software orchestration
• Full platform: Managed Kubernetes (CMK, Run:ai certified), managed Slurm, Kubeflow, plus the new Command Center observability/orchestration dashboard (February 2026)
• Now offers native S3-compatible object storage ($0.06/GiB/mo), closing a gap from the prior edition; VAST-backed shared disks ($0.07/GiB/mo) are available to on-demand users
• Strong AMD GPU support alongside NVIDIA; 99.98% uptime SLA and SemiAnalysis ClusterMAX “Gold”

Gaps

• B200 and GB200 require a sales conversation; not self-service on-demand
• No documented SAML/SSO for self-service customers
• Energy price volatility exposure: the Texas grid crisis (March 2025) saw costs spike
• Stranded gas supply may decline over the long term as the world transitions away from fossil fuels

Best for: Teams prioritizing sustainability, AMD GPU access, or participation in Stargate-class infrastructure.

Lambda

Overview

Lambda was founded in 2012 by brothers Stephen and Michael Balaban. The company is known for its developer-friendly approach and the Lambda Stack (pre-configured PyTorch/TensorFlow/CUDA environment) used by 100K+ users.

Funding has accelerated: $320M Series C (February 2024), $480M Series D (February 2025), and over $1.5B Series E (November 2025) led by TWG Global and USIT, at a $6B post-money valuation. NVIDIA is a major investor and strategic partner. Lambda raised $350M in pre-IPO convertible notes (reported January 2026); its IPO, originally targeted for H1 2026, has slipped to H2 2026 and remains unconfirmed as of mid-2026.

The NVIDIA relationship is deep: a September 2024 $1.5B GPU leaseback deal has NVIDIA leasing GPUs from Lambda, and a November 2025 multibillion-dollar Microsoft agreement covers tens of thousands of GPUs including GB300 NVL72.

Infrastructure

Lambda operates on a pure colocation model (no owned facilities), across roughly 15 US regions including San Francisco, Dallas/Fort Worth, Plano (TX), Columbus, and Kansas City, with more capacity ramping in 2026.

Hardware: HGX H100 (PCIe $3.29/hr, SXM $3.99-4.29/hr), HGX H200 (clusters), HGX B200 ($6.69-6.99/hr). Published H100 SXM pricing rose roughly 33-43% in early 2026. GB300 NVL72 Bare Metal Instances were announced at GTC 2026 but are reserved/contact-sales only, not self-service on-demand. 1-Click Clusters scale from 16 to 2,000+ GPUs. Newer Blackwell deployments use NVIDIA Quantum-X800 InfiniBand with co-packaged optics; legacy H100 1-Click Clusters still use Quantum-2.

A note on storage: Lambda markets VAST Data and WEKA partnerships, but these are backend infrastructure for Lambda’s own cloud, not self-service customer products. The customer-facing storage is Lambda’s S3-compatible object storage and high-capacity filesystems (with no egress charges).

Strengths

• 1-Click Clusters: multi-node provisioning with a short reservation minimum (no long-term contracts)
• Pre-installed Lambda Stack eliminates environment configuration
• Quantum-X800 InfiniBand in production on its newest Blackwell factory
• No egress/ingress fees; SOC 2 Type II certified

Gaps

• Published on-demand H100 pricing has risen sharply (now $3.99-4.29/hr SXM)
• GB300, Vera CPU, and STX storage are announced for H2 2026 but not yet self-service
• No managed Kubernetes or Slurm; multi-node clusters provision via a reservation/approval flow
• VAST/WEKA are backend-only, not exposed to self-service users
• US-only footprint; no built-in cost allocation/usage tracking by team or project

Best for: Teams wanting fast, simple cluster provisioning without long-term commitments, comfortable with SSH/terminal workflows.

Voltage Park

Overview

Voltage Park was founded in 2023 with an unusual structure: it is backed by a roughly $900M grant from Navigation Fund, a nonprofit founded by Jed McCaleb (Stellar co-founder, Ripple co-founder). The mission is democratizing AI infrastructure access.

Leadership includes Ozan Kaya (CEO, ex-CarLotz President) and Saurabh Giri (Chief Product & Technology Officer, ex-Amazon Bedrock lead). In March 2025 Voltage Park acquired the TensorDock marketplace, and in January 2026 it merged with Lightning AI (which brought 400M+ PyTorch downloads and ~240K developers) to form a company valued at $2.5B with $500M+ in annual recurring revenue. The combined entity is among the larger neoclouds by GPU count.

Infrastructure

After the Lightning AI merger, Voltage Park operates roughly 35,000 NVIDIA GPUs across six US data centers in Washington, Texas, Virginia, and Utah (up from 24,000 H100s). The Quincy, WA facility runs on hydro and wind power.

H100 nodes run on Dell PowerEdge XE9680 servers: 8 H100s per node with NVLink, 1TB RAM, dual Intel Xeon Platinum 8470. Quantum-2 InfiniBand provides 400Gb/s per GPU on bare metal. Next-gen hardware (B200, GB200, B300, GB300) is available via long-term contracts only, not self-service on-demand.

Storage: a June 2025 VAST Data partnership is deployed across most US data centers, but the public pricing page does not advertise VAST to on-demand customers, and the partnership language emphasizes enterprise/regulated use cases. Treat VAST here as a reserved/enterprise capability; self-service instances ship with locally attached NVMe. Managed Kubernetes (launched June 2025) includes the NVIDIA GPU Operator, Prometheus/Grafana, and a VAST CSI driver.

Strengths

• Competitive pricing: $1.99/hr (Ethernet) or $2.49/hr (InfiniBand) for H100s with fast spinup, no contracts required
• Bare-metal access for high training throughput
• SOC 2 Type II, ISO 27001, and HIPAA certified
• Only neocloud partner in the NSF NAIRR pilot; donated 1M H100 GPU hours for research
• Lightning AI merger adds a large developer community and platform tooling

Gaps

• B200/GB200/B300/GB300 are contract-only, not self-service
• VM instances fall back to Ethernet; only bare metal gets 400Gb/s InfiniBand
• VAST storage is not confirmed as self-service; self-service users get local NVMe
• Managed Kubernetes is bare-metal only (VM support still in development)
• No data recovery after instance termination; customers must back up externally

Best for: Researchers, startups, and teams wanting low-cost H100 access, especially those eligible for NAIRR research allocations or already using Lightning AI tooling.

GMI Cloud

Overview

GMI Cloud was founded as an offshoot of Realtek Semiconductors and GMI Technology and is headquartered in San Jose, California. It raised an $82M Series A (October 2024, $15M equity + $67M debt) led by Headline Asia, with strategic backing from Thailand’s Banpu and Taiwan’s Wistron; total capital raised is around $93M.

GMI Cloud is an NVIDIA Cloud Partner, providing access to current GPU architectures including H200 and Blackwell systems.

Infrastructure

Regions span the US, Europe, and Asia-Pacific. GMI has announced large sovereign-AI buildouts: a Taiwan AI Factory and a Kagoshima, Japan AI Factory ($12B, development starting late 2026, with NVIDIA and Wistron). These are real projects but not yet operational.

Hardware: H100 ($2.00/hr, down from $2.10), H200 ($2.60/hr), B200 ($4.00/hr), GB200 ($8.00/hr). B200 and GB200 moved from pre-order to live self-service pricing in early 2026. All training clusters include 400Gb/s InfiniBand via the Cluster Engine.

Storage: GMI markets a VAST Data partnership prominently, but it is gated to enterprise and AI-factory deployments. A self-service user at console.gmicloud.ai gets GMI’s own high-performance filesystem; the VAST-backed S3 “Cold Storage” and GPUDirect tiers target enterprise migrations, not the self-service tier.

Strengths

• GMI Cluster Engine provides managed Kubernetes orchestration (CaaS and BMaaS) with autoscaling
• B200 and GB200 now available at published self-service rates
• H100 pricing competitive at $2.00/hr
• Strong Asia-Pacific presence and sovereign-AI partnerships

Gaps

• VAST Data storage is enterprise/bespoke, not transparently self-service
• Limited public documentation; compliance limited to SOC 2 Type 1 and ISO 27001
• No Slurm offering documented
• Smaller footprint and brand recognition than major neoclouds in North America and Europe

Best for: Teams seeking competitively priced H200/B200 access with managed Kubernetes, especially with Asia-Pacific presence needs.

RunPod

Overview

RunPod was founded in 2022 and is headquartered in New Jersey. Its last disclosed raise was a $20M round (co-led by Intel Capital and Dell Technologies Capital). By January 2026 the company surpassed $120M ARR.

The platform serves 500,000+ developers, from individual researchers to enterprise teams. RunPod’s differentiator is simplicity: GPU instances launch in under a minute with pre-configured ML environments. In April 2026 it launched the open-source Flash SDK for serverless inference without container management.

Infrastructure

RunPod operates 31 data centers globally (14 Secure Cloud, 17 Community Cloud). The platform offers three deployment models:

1. Pods: GPU VMs with persistent storage, available on-demand or spot (up to 80% cheaper)
2. Serverless: Auto-scaling inference endpoints billed per-second
3. Community Cloud: Marketplace of third-party GPU capacity at lower prices

Hardware: H100 (PCIe $2.89/hr, NVL $3.19/hr, SXM $3.29/hr), H200 ($4.39/hr), B200 ($5.89/hr), A100 80GB ($1.49/hr SXM), L40S, RTX 4090/3090. Instant Clusters bring self-service multi-node networking: 1,600-3,200 Gbps east-west via InfiniBand or RoCEv2, deployable at 2 nodes on-demand and up to 8 nodes on request.

Storage: Network volumes ($0.10/GB/mo standard, $0.05-0.07/GB/mo shared), S3-compatible object storage. No VAST or WEKA on the self-service product.

Strengths

• Sub-minute instance launch times with one-click templates
• Serverless inference with pay-per-second billing and automatic scaling; Flash SDK (2026) removes container management
• Self-service InfiniBand/RoCE multi-node clusters (Instant Clusters)
• Spot instances at 50-80% discount for interruptible workloads
• SOC 2 Type II (achieved October 2025), SOC 3, HIPAA with BAAs, GDPR

Gaps

• No managed Kubernetes or native Slurm (container/serverless focus)
• Community Cloud capacity quality varies by host
• Multi-node distributed training beyond Instant Clusters requires manual configuration
• Last disclosed funding round is dated; growth is now self-funded via usage

Best for: Individual developers and small teams wanting fast, simple GPU access for inference and single-node training, with self-service multi-node clusters when needed.

Hyperstack

Overview

Hyperstack is the GPU cloud arm of NexGen Cloud, a UK-based infrastructure company founded in 2020. The platform positions itself as a cost-effective alternative to hyperscalers. NexGen Cloud raised a $45M Series A in April 2025 (about $59M total funding) and reported strong growth in its AI cloud operations.

Infrastructure

Hyperstack operates self-service across three regions (CANADA-1, NORWAY-1, US-1), with a Houston H100 SXM region coming soon, and roughly 13,000 GPUs deployed. The platform offers tiered service levels:

1. Standard (on-demand) tier: GPU VMs with standard Ethernet networking
2. Supercloud tier: Reserved high-performance clusters with 400Gb/s Quantum-2 InfiniBand (SHARP) for distributed training

Hardware: H100 (PCIe $1.90/hr, SXM $2.40/hr), H200 ($3.50/hr on-demand, $2.45/hr reserved), A100, L40S. H100 PCIe at $1.90/hr is among the cheapest in the market. B200 and GB200 are reservation-only via direct sales; they do not appear in the on-demand pricing table.

A key self-service caveat: InfiniBand is exclusive to the reserved Supercloud product. On-demand users get plain Ethernet. The WEKA partnership announced in 2025 also targets the Supercloud line, not the on-demand product, where storage is block, local NVMe, and S3-compatible object (CANADA-1 only).

Strengths

• Aggressive on-demand pricing: H100 at $1.90/hr, H200 at $3.50/hr
• Supercloud tier provides InfiniBand and WEKA for reserved multi-node training
• Simple RESTful API and web console; managed on-demand Kubernetes available
• 100% renewable-powered, GDPR-aligned European data centers

Gaps

• InfiniBand and WEKA are Supercloud-reserved only; the on-demand tier is Ethernet with the provider’s own storage
• B200/GB200 are reservation-only, not self-service
• Kubernetes autoscaling (HPA, Cluster Autoscaler) was still “coming soon” as of late 2025
• Compliance certifications not published publicly
• Smaller GPU fleet and thinner documentation than CoreWeave, Nebius, or Lambda

Best for: Cost-conscious teams wanting affordable on-demand H100/H200 access, comfortable with Ethernet networking unless they reserve a Supercloud cluster.

Verda (formerly DataCrunch)

Overview

DataCrunch was founded in 2019 in Helsinki, Finland, and rebranded to Verda in November 2025 (a rebrand, not an acquisition) to emphasize its sustainability positioning. The company operates in Finland and Iceland.

The core differentiator is renewable energy: Verda’s data centers run on geothermal and hydroelectric power, providing genuine carbon-neutral AI infrastructure rather than offset-based claims. In April 2026 it raised a $117M Series B (about EUR 100M equity plus Nordic debt) led by Lifeline Ventures, and reports being cash-flow positive with an annualized revenue run rate above $60M. It is an NVIDIA Preferred Partner with early access to Blackwell Ultra.

Infrastructure

Primary data centers are in Finland and Iceland, leveraging renewable energy and natural cooling. Expansion to the UK, US, and Asia was announced for 2026.

Hardware (on-demand, current rates): H100 SXM5 ($3.25/hr), H200 ($4.00/hr), B200 ($6.11/hr), B300 ($7.50/hr), GB300 ($8.62/hr), A100. Published pricing rose 30-60% across the board since the prior edition; the previously listed “$1.24/hr B300” was a stale or mislabeled figure. Multi-node clusters use 400Gb/s NDR InfiniBand (16-128 GPUs), while GB300 instances use NVLink v5 for a unified memory domain across Grace CPU and dual B300 per node. Spot pricing offers up to 40% off, with commitment discounts up to 25%.

Storage: Block storage ($0.05-0.20/GB/mo) and NVMe shared filesystem ($0.20/GB/mo). Native S3-compatible object storage was flagged as “coming” in the prior edition and is still not available.

Strengths

• Renewable energy (geothermal/hydro), not offsets
• Early GB300/B300 availability as an NVIDIA Preferred Partner
• 400Gb NDR InfiniBand standard on clusters
• SOC 2 Type II plus ISO 27001/27017/27018/27701; GDPR compliant
• Cash-flow positive, reducing financing risk

Gaps

• Published on-demand pricing rose sharply in early 2026
• No managed Kubernetes (Slurm only); no object storage; no Terraform provider
• No HIPAA despite a broad European compliance stack
• Nordic-only footprint today may add latency for US/Asia users
• Bare-metal focus; less abstraction than serverless platforms

Best for: Organizations with sustainability mandates needing genuine renewable energy infrastructure, or European teams wanting early Blackwell Ultra access with strong compliance.

Vultr

Overview

Vultr was founded in 2014 as a general-purpose cloud provider, making it one of the more established players in this comparison. The company has expanded aggressively into GPU cloud, becoming an NVIDIA Cloud Partner with both NVIDIA and AMD GPU offerings.

Vultr differentiates through global footprint: 33 data center locations worldwide (Milan added in May 2026), more than any neocloud. This enables low-latency inference deployments close to end users. The company raised a $333M Series A at a $3.5B valuation (December 2024, LuminArx and AMD Ventures) and added $329M in debt financing in mid-2025.

Infrastructure

33 locations spanning North America, Europe, Asia-Pacific, South America, and Australia. This geographic diversity is unmatched among neoclouds.

Hardware: NVIDIA H100 ($1.99/GPU/hr on 24- and 48-month contracts), GH200 ($1.99/hr on-demand), A100, L40S. AMD MI300X ($1.85/hr on-demand preemptible, $1.75/hr 48-month), MI325X ($2.00/hr preemptible), MI355X ($2.59/hr on-demand preemptible, $2.65/hr 48-month, now generally available for self-service). B200 is $1.99/GPU/hr as a 48-month cloud instance or $2.99/GPU/hr bare metal. GB300 NVL72 opened for preorder in January 2026 with pricing not yet published. A 50MW MI355X supercluster (24,000 GPUs) in Springfield, Ohio was announced in December 2025 with an early-2026 launch.

Networking: Vultr Clusters provisions Quantum-2 InfiniBand (3200Gb/s, non-blocking) or RoCEv2 Ethernet through self-service automation, no manual reservation required.

Storage: Block storage ($0.10/GB/mo), Object storage ($0.018-0.10/GB/mo with S3 API), NVMe shared filesystem ($0.10/GB/mo). All Vultr-owned; no VAST/WEKA.

Strengths

• Among the cheapest AMD MI300X in market at $1.75-1.85/hr; MI355X now self-service
• 33 global locations enable low-latency edge deployments
• Self-service InfiniBand or RoCE clusters without sales gating
• Full stack: VKE (managed Kubernetes, v1.35 with GPU Operator), bare-metal, block/object/shared storage
• Strong compliance: SOC 2 Type II (HIPAA), ISO 27001, PCI DSS
• Both NVIDIA and AMD GPU availability; established operator

Gaps

• H100 rates are tied to multi-year prepaid contracts; B200 is reserved-only
• GPU fleet smaller than CoreWeave, Lambda, or Nebius
• VKE has no bare-metal worker nodes
• Less AI/ML-specific tooling than specialized neoclouds

Best for: Teams needing global GPU presence for inference, AMD GPU access, or preferring an established provider with self-service clusters and broad compliance.

OVHcloud

Overview

OVHcloud is a French cloud provider founded in 1999, making it the oldest company in this comparison. Publicly traded on Euronext Paris, it reported EUR 555M of revenue in H1 FY2026 (six months to February 2026), with about 6% organic growth and 105% net revenue retention. The company operates 50+ data centers globally with a strong European presence.

OVHcloud’s GPU offerings are part of its broader cloud portfolio. The company emphasizes European data sovereignty, owning and operating all infrastructure without reliance on US hyperscalers. Its SecNumCloud-qualified (ANSSI) Bare Metal Pod hosts H100/H200, and full SecNumCloud IaaS reaches general availability in June 2026, making OVHcloud one of few providers qualified for sensitive French public sector AI workloads.

Infrastructure

50+ data centers across Europe, North America, and Asia-Pacific, with new local zones planned for 2026. OVHcloud uses water-cooled systems that reduce energy consumption.

Hardware: H100 (about $2.99/hr public cloud, EUR 3.33/hr; EUR 3.10/hr managed), H200 (Paris GA, Milan 3-AZ), A100, L4, L40S. Pricing rose effective April 2026. B200 and GB200 are not offered in the public self-service cloud as of mid-2026. Public GPU instances use Ethernet/RoCE with up to 25Gb/s networking; OVHcloud does not document InfiniBand for self-service.

Storage is standard: S3-compatible Object Storage (EUR 0.0119/GB/mo), Block Storage, and a File Storage beta launched in Q1 2026. No VAST or WEKA integration in the public cloud. Managed Kubernetes Standard launched in Q1 2026 with a 99.99% SLA, multi-AZ control plane, 500-node scaling, and bidirectional autoscaling including scale-to-zero for idle GPU nodes.

Strengths

• European data sovereignty with no US hyperscaler dependencies
• SecNumCloud, SOC 2, ISO 27001, PCI DSS, HDS certifications; full SecNumCloud IaaS GA June 2026
• 50+ data centers; water-cooled infrastructure
• Managed Kubernetes with scale-to-zero GPU autoscaling
• 25+ years operational track record; competitive European pricing

Gaps

• No B200/GB200 in the public self-service cloud; slower to adopt the latest architectures
• No InfiniBand for self-service GPU instances
• Storage is undifferentiated (standard S3/block/file), no AI-optimized tier
• AI/ML tooling less developed than specialized providers
• GPU remains a secondary business behind general cloud

Best for: European enterprises with data sovereignty requirements, French public sector organizations needing SecNumCloud certification, or teams preferring established European infrastructure for inference and single-node work.

FluidStack

Overview

FluidStack was founded in 2017 at Oxford University in London. Funding has accelerated: a $200M Series A (February 2025), a $450M Series B (January 2026), and reported Series C talks for around $1B at an $18B valuation (April 2026, Jane Street and Situational Awareness). The valuation jumped from roughly $7.5B to $18B in months. FluidStack has also secured up to $10B in GPU-collateralized debt from Macquarie Group.

The company has transitioned from a GPU marketplace aggregator to primarily building dedicated infrastructure for large enterprises. Revenue mix is roughly 62% Private Cloud (single-tenant, $100M+ average contracts) and 38% Marketplace. Notable customers include Anthropic (a large multi-data-center buildout in NY and TX), Meta, Mistral AI, Character.AI, Poolside, and Black Forest Labs, plus 10-year hosting agreements with TeraWulf.

Infrastructure

FluidStack operates two distinct models:

1. Private Cloud: Purpose-built, single-tenant GPU clusters for enterprise customers with dedicated engineering support. Storage integrates customer-preferred solutions (VAST, WEKA, DDN). The Anthropic buildouts (168 MW Abernathy, TX; 160-360 MW Lake Mariner, NY) fall here, dedicated, not shared self-service capacity.

2. Marketplace: Aggregated GPU capacity from partner data centers with variable specifications.

Hardware: H100, H200, B200, GB200 validated to high performance. Atlas OS provides bare-metal orchestration with managed Kubernetes (Ray, Volcano, Kueue) and Slurm. Self-service hourly pricing is no longer published: the public pricing page routes to a “Request Pricing” sales flow, reflecting the shift toward contracted private cloud.

Strengths

• Proven track record building custom data centers for top AI labs (Anthropic, Meta)
• $10B debt capacity enables rapid scaling without customer pre-funding
• Single-tenant by default with full infrastructure isolation
• Enterprise compliance: GDPR, HIPAA, ISO 27001, SOC 2 Type II
• Flexible storage integration (VAST, WEKA, DDN) for private-cloud deployments

Gaps

• No published self-service on-demand pricing; premium configurations require a sales quote
• VAST and advanced storage are private-cloud only, not exposed to marketplace/self-service users
• Marketplace tier has variable infrastructure quality
• Less public documentation than competitors

Best for: AI labs and enterprises needing custom-built GPU infrastructure with long-term contracts. Self-service buyers will find more transparent pricing elsewhere.

Vast.ai

Overview

Vast.ai was founded in 2018 as a marketplace for GPU compute. Unlike traditional cloud providers, it connects renters with independent GPU hosts, similar to Airbnb for compute. This model enables some of the lowest prices in the market but with significant variability in infrastructure quality. (Note: Vast.ai the GPU marketplace is a different company from VAST Data, the storage vendor referenced elsewhere in this report.)

The platform is popular with researchers, hobbyists, and cost-conscious startups, serving hundreds of thousands of users.

Infrastructure

Vast.ai is a marketplace, not a traditional cloud provider. GPU capacity comes from professional data center hosts and individual hosts, spread across 40+ data centers. The platform shows real-time availability, reliability scores, and host ratings.

Pricing (live, mid-2026): H100 SXM around $2.33/hr, H100 NVL $2.36/hr, H100 PCIe $2.40/hr. H200 lists at $3.88/hr but availability is essentially zero. B200 rentals are appearing with limited supply, expected in the $5.50-7.00/hr range. As a marketplace, these rates float with host supply and demand. Most instances are Ethernet-only; InfiniBand is available only from specific data center hosts. Storage is per-host block storage ($20+/TB/mo) with no native object storage or standardized shared filesystem.

Docker-based deployments with templates for PyTorch, TensorFlow, Stable Diffusion, and other frameworks. A new Serverless product launched in December 2025, adding OpenAI-compatible inference endpoints in April 2026.

Strengths

• Among the lowest H100 prices in market from quality hosts
• Massive selection of GPU types including consumer hardware
• Real-time availability and pricing transparency; host reliability ratings
• Serverless inference layer added in 2026
• No minimum commitments; pay-per-minute billing
• SOC 2 Type II and HIPAA (with BAAs) now available on Secure Cloud

Gaps

• Infrastructure quality varies dramatically by host
• No InfiniBand on most instances (data center hosts only)
• No managed Kubernetes or Slurm; no SSO
• Multi-node training is difficult due to fragmented infrastructure
• No SLA guarantees on marketplace instances; support quality varies by host

Best for: Researchers and hobbyists prioritizing cost over reliability, and inference workloads that fit the new Serverless layer.

TensorWave

Overview

TensorWave was founded in 2023 in Las Vegas by Darrick Horton, Jeff Tatarchuk, and Piotr Tomasik. It closed a $100M Series A in May 2025 led by Magnetar and AMD Ventures (about $147M total funding, including a $43M SAFE), and disclosed ARR exceeding $100M. TensorWave is an AMD-exclusive GPU cloud provider.

While most neoclouds build on NVIDIA infrastructure, TensorWave bet entirely on AMD’s Instinct line. It deployed the largest AMD training cluster in North America (8,192 MI325X GPUs, announced July 2025) and was first to deploy large-scale direct liquid-cooled AMD GPU infrastructure. It holds SOC 2 Type II, ISO 27001, and HIPAA certifications.

Infrastructure

TensorWave operates US data centers (Arizona, plus Pennsylvania capacity added in Q1 2026) purpose-built for AMD GPUs. The infrastructure uses Aviz ONES fabric for RoCE-based Ethernet networking rather than InfiniBand. (Its Dublin presence is an EMEA sales office, not a data center.)

Hardware (bare metal): MI300X, MI325X, and MI355X. TensorWave no longer publishes hourly rates on its site (it directs buyers to sales); earlier listings put these around $1.71, $1.95, and $2.85 per GPU-hour respectively, which should now be confirmed directly. The next-generation MI455X (Helios architecture) is at engineering-sample stage for H2 2026, with mass production expected in Q2 2027; it is not available to self-service users. All systems support AMD ROCm for PyTorch, TensorFlow, and JAX.

Storage: self-service bare metal includes in-node flash only. TensorWave’s WEKA partnership is positioned for large clusters and bespoke deployments with custom pricing; it is not provisionable through self-service.

Strengths

• AMD-first specialization provides access when NVIDIA is constrained
• MI300X at $1.71/hr and MI325X at $1.95/hr are competitive with NVIDIA H100 pricing
• Deep ROCm expertise; AMD Ventures backing suggests priority inventory access
• Managed Kubernetes and Slurm now offered alongside bare metal

Gaps

• AMD ROCm requires workload adaptation (not a drop-in CUDA replacement)
• WEKA storage is enterprise/bespoke only; self-service users get in-node flash
• RoCE has higher tail latency than InfiniBand under congestion
• Kubernetes/Slurm specifics (autoscaling, multi-tenancy) are thinly documented for self-service
• Less ecosystem tooling than NVIDIA-focused providers

Best for: Teams with ROCm expertise seeking AMD GPU access, or those willing to adapt workloads to benefit from AMD’s price/performance.

Hot Aisle

Overview

Hot Aisle was founded in October 2023 by Jon Stevens and Clint Armstrong, with backing from Consensys Mesh (Joseph Lubin’s VC arm). It is headquartered in Cheyenne, WY and remains seed-stage with no announced priced round. The company specializes exclusively in AMD Instinct GPUs. It holds SOC 2 Type II and HIPAA certifications, with ISO 27001 planned.

Infrastructure

Hot Aisle operates from the Switch Pyramid facility in Grand Rapids, Michigan (Tier 5 Platinum, 100% renewable). Infrastructure includes Dell XE9680 servers with Broadcom 57608 200G adapters and Dell PowerSwitch Z9864F spine switches at 400G.

Networking: RoCEv2 delivering 3200 Gbps throughput per node. Per-minute billing with no long-term contracts.

Hardware: MI300X (192GB) at $1.99/hr on-demand for 1x/2x/4x VMs, or $3.39/GPU/hr for reserved 8x bare-metal nodes. The MI355X is reservation-only with no announced launch date or per-GPU price. No managed storage, Kubernetes, or Slurm is offered on the self-service product; orchestration is customer-managed.

Strengths

• AMD MI300X at $1.99/hr is competitive pricing
• SOC 2 Type II and HIPAA certified
• 3200 Gbps RoCEv2 throughput per node on Dell/Broadcom infrastructure
• Per-minute billing, no contracts required

Gaps

• Single data center location (Michigan); no expansion announced
• MI355X has no firm availability date
• No managed storage, Kubernetes, or Slurm; bare-metal/VM only
• Small, early-stage company; AMD ROCm requires workload adaptation from CUDA

Best for: Teams seeking AMD MI300X access at competitive pricing with enterprise compliance certifications, comfortable managing their own orchestration.

Nscale

Overview

Nscale launched from stealth in May 2024 and has raised aggressively: a $155M Series A (December 2024), a $1.1B Series B (September 2025, led by Aker ASA), and a $2B Series C in March 2026 at a $14.6B valuation, backed by NVIDIA, Lenovo, Dell, Citadel, Jane Street, and Nokia. It hired Goldman Sachs and JPMorgan as underwriters and is targeting an H2 2026 IPO. A multi-year Microsoft deal (announced October 2025) carries a roughly $10B commitment.

Nscale focuses on sustainable AI infrastructure, operating data centers in the Nordics powered by renewable energy. It has a joint venture with Aker ASA for “Stargate Norway” (targeting 100,000 NVIDIA GPUs by end of 2026) and a partnership with OpenAI.

Infrastructure

Owned facilities in Glomfjord, Norway plus a colocation deployment with Verne in Iceland, where the first phase completed ahead of schedule in April 2026 (7.5 of 15 MW, including 64 liquid-cooled GB300 racks). A May 2026 Portugal expansion (EUR 695M with Microsoft at Start Campus) targets 66,000+ next-generation Rubin GPUs from late 2027. All facilities use hydroelectric and geothermal power.

Hardware: H100, H200, GB200/GB300 NVL72, A100, and AMD MI300X, all sales-gated (no published self-service hourly pricing). A Serverless Inference API (launched April 2025) is self-service and token-priced, but GPU cluster capacity remains enterprise-only.

Networking is entirely Ethernet-based using Nokia switches. A correction from the prior edition: Nscale’s fabric is Ultra Ethernet Consortium (UEC) compliant via the Nokia IXR-H6 (800GE/1.6TE), a newer AI-fabric standard distinct from legacy RoCE.

Strengths

• Genuine renewable energy (hydro/geothermal), not carbon offsets
• Strong investor backing (NVIDIA, Microsoft, OpenAI, Aker ASA) and a high valuation
• GB300 capacity deployed in Iceland; large Stargate Norway pipeline
• UEC networking via Nokia for next-generation AI fabrics
• Managed Slurm and a Kubernetes service (NKS) added in 2026, closing a prior gap

Gaps

• GPU clusters are sales-gated; only inference endpoints are self-service
• Compliance certifications are not prominently published
• Self-service vs enterprise feature parity is opaque
• Nordic locations may add latency for US/Asia workloads
• Early-stage company; operational track record still developing

Best for: Large enterprises with sustainability mandates seeking renewable-powered GPU infrastructure, especially those interested in OpenAI ecosystem alignment. Not a fit for self-service GPU rental.

SF Compute

Overview

SF Compute (San Francisco Compute) was founded in 2023 by Evan Conrad and raised a $40M Series A in early December 2025, led by DCVC and Wing Venture Capital at a $300M valuation. Other backers include Electric Capital and Alt Capital. The company has around 30 employees and hired Eric Park (former Voltage Park CEO) as CTO in late 2025.

SF Compute operates as a GPU marketplace/broker, deliberately avoiding hardware ownership. The platform enables buyers to access compute and resell unused capacity, creating spot and forward markets for GPU compute. It manages roughly $100M worth of GPU hardware (a dollar figure, not a GPU count, of several thousand GPUs).

The key differentiator is flexible time-based reservations: you can book GPU capacity for arbitrary windows at guaranteed prices, and automate hourly/daily/weekly/monthly purchasing against a maximum price you set.

Infrastructure

As a marketplace, SF Compute does not own infrastructure but provides access to partner capacity. The platform’s self-service product is VM-based:

• VMs (h100v): default self-service option, no InfiniBand, ~5-minute spinup
• Kubernetes clusters (h100i): 3.2Tb/s InfiniBand, but sales-contact only, not self-service
• Bare-metal/Slurm: custom onboarding

Hardware: H100 from roughly $1.82/hr (the marketplace price fluctuated between about $1.20 and $2.20 through May 2026 and varies by zone). H200 shows little to no active market pricing. B300 has slipped from a Q2 2026 target to fall 2026 and is sales-contact only. InfiniBand for self-service VMs has been pushed to Q2-Q3 2026. Storage is local NVMe (1.5TB+ per node); there is no managed object or shared filesystem on self-service, and no VAST/WEKA.

Strengths

• H100 pricing from ~$1.82/hr is among the lowest in market (marketplace rate fluctuates)
• Flexible time-based reservations at guaranteed prices, with programmatic price caps
• Marketplace model allows resale of unused capacity
• Self-service signup and credit-card funding

Gaps

• InfiniBand and Kubernetes are sales-contact only; self-service is Ethernet VMs
• No owned infrastructure; performance varies by underlying provider
• No published SOC 2/ISO/HIPAA certifications, no SSO, no documented uptime SLA
• H200/B300 not yet self-service

Best for: Price-sensitive teams wanting the lowest H100 VM costs and flexible reservation windows, comfortable without InfiniBand or formal compliance on the self-service tier.

Last updated: June 2026. Pricing and features change frequently, and several providers raised published on-demand rates in early 2026. Verify current offerings on provider websites before making decisions, and confirm that any storage or networking capability you rely on is available on the self-service tier rather than only in a reserved or bespoke contract.