GPU Cloud Comparison 2026: 15 Providers Ranked

Fifteen GPU cloud providers span a 6x price range for the same NVIDIA H100: roughly $1.49-2.10/hr on marketplaces and boutique neoclouds versus $10.98-12.29/hr on Azure and Google Cloud, according to GetDeploying and Spheron pricing trackers. Price alone does not decide the winner — scale, interconnect, SLA, and compliance matter as much for production workloads. This guide ranks 15 providers across those dimensions and gives the pricing, feature, and fit data behind each rank.

Key takeaways

• Neoclouds beat hyperscalers on price by 2-4x for the same GPU, per TensorWave/Spheron market comparisons — CoreWeave, Lambda, Nebius, RunPod, and FluidStack all undercut AWS, Azure, GCP, and Oracle on published on-demand rates.
• CoreWeave is the scale leader: ~45,000 GPUs and a $99.4 billion contracted backlog as of Q1 2026, but it sells only 8-GPU nodes — no path to a single self-serve GPU.
• Vast.ai and FluidStack post the lowest H100 rates (~$1.49-2.10/hr), with no or minimal platform SLA; RunPod’s Secure Cloud (~99.5% uptime) is the cheapest option with real reliability guarantees.
• Hyperscalers remain the compliance and ecosystem default: broadest certification portfolios (SOC 2, HIPAA, FedRAMP, ISO), but Azure’s $12.29/GPU-hr H100 rate is the highest tracked among major providers.
• AMD is a distinct, thinner market: TensorWave (AMD specialist) and Vultr both list MI300X under $2/GPU-hr, roughly 20% below comparable-memory H100 pricing, but with far fewer providers to compare.
• GB200/Blackwell-generation capacity is still largely reserved-contract only — CoreWeave and Together AI sell it, but most GB200 access runs through quote-only agreements, not published on-demand rates.

For live rates across GPU models, see our GPU price index; for the colocation side of AI infrastructure, see the colocation price index and data center catalog.

How this ranking works

Providers are ranked on four weighted factors, in order of importance for most AI buyers in 2026:

1. Price competitiveness — published on-demand and reserved rates for H100/H200/B200 against the tracked market range.
2. Scale and reliability — GPU fleet size, uptime SLA (if any), and whether capacity is actually available at the advertised price.
3. Training-grade networking — availability of InfiniBand/NVLink multi-node clusters versus single-GPU-only provisioning.
4. Developer and enterprise fit — self-serve provisioning, billing granularity, compliance certifications, and support model.

This produces a “best overall value for a typical AI buyer” order, not a pure price ranking — a workload with strict compliance needs or 2,000-GPU training runs should weight scale and networking above headline price. The “Which to choose” section at the end maps ranking to specific buyer profiles.

The ranking: 15 GPU clouds

Rank	Provider	Category	H100 on-demand	Standout strength
1	CoreWeave	Enterprise neocloud	~$6.16/hr (8-GPU nodes only)	Largest independent scale, $99.4B backlog
2	Lambda	Managed AI cloud	$3.99-4.29/hr	99.9% SLA, 1-Click Clusters to 2,000+ GPUs
3	Nebius	Vertically integrated neocloud	$3.85/hr on-demand, $2.15 spot	Self-serve 1-GPU, up to 35% commit discounts
4	RunPod	Hybrid platform	$1.99 (Community) / $2.89 (Secure)	Best price-to-reliability ratio, no egress fees
5	Crusoe	Sustainable-power neocloud	~$3.90/hr single-GPU on-demand	262% revenue growth 2024-25, stranded-energy model
6	Together AI	Inference + training hybrid	$5.49/hr on-demand, $3.49 reserved	Clusters to 4,000+ GPUs, per-token API option
7	FluidStack	Boutique neocloud	~$2.10/hr	Among the lowest enterprise-grade rates tracked
8	TensorWave	AMD specialist	N/A (AMD-only)	MI355X from $2.85/hr, MI300X sold out on demand
9	Vast.ai	Marketplace	~$1.53-1.87/hr	Cheapest widely available rate; no platform SLA
10	Vultr	Budget global cloud	$0.118-19.18/hr range	Cheapest MI300X/MI355X (~$1.85-2.59/hr)
11	DigitalOcean	Simplified GPU droplets	$3.39/hr ($3.26 reserved)	Simplest billing, no long-term commitment needed
12	Oracle Cloud (OCI)	Bare-metal hyperscaler	$10/hr flat (8-GPU node = $80/hr)	Only major hyperscaler offering true bare metal
13	Google Cloud	Hyperscaler	~$10.98-11.06/hr (A3, on-demand)	Deepest integration with Vertex AI/TPU ecosystem
14	AWS	Hyperscaler	~$5.19-6.87/hr (Capacity Blocks)	Broadest compliance and enterprise procurement fit
15	Azure	Hyperscaler	~$12.29/hr (ND H100 v5)	Deepest Microsoft/OpenAI-adjacent tooling, highest price

Rates are published on-demand figures as of mid-2026 from provider pricing pages and GetDeploying/Spheron/Thunder Compute aggregation; several (CoreWeave, Together AI, Oracle) also sell in fixed multi-GPU node sizes rather than per-GPU. AWS’s Capacity Block rates rose roughly 20% on July 1, 2026. Note AWS ranks above Azure despite a narrower spread because Capacity Blocks list below Azure’s ND H100 v5 rate and AWS carries broader compliance coverage.

Tier 1: Enterprise neoclouds (ranks 1-3)

CoreWeave is the scale leader — the largest independent GPU provider in North America at roughly 45,000 GPUs, with a $99.4 billion contracted revenue backlog reported on its Q1 2026 earnings call, up nearly 4x year-over-year. It sells exclusively in 8-GPU HGX nodes with NVLink and InfiniBand, aimed at teams training 70B+ parameter models who need guaranteed multi-node availability — but there is no path to a single self-serve GPU, and H100 pricing (~$6.16/hr) sits well above RunPod or Nebius.

Lambda is the managed-AI-cloud benchmark: a narrow catalog (H100, H200, B200, A100) backed by a contractual 99.9% uptime SLA and human support over Slack, email, and phone. 1-Click Clusters scale from 16 to 2,000+ InfiniBand-connected GPUs — the differentiator against RunPod and Vast.ai for serious distributed training, at a 40-100% price premium over budget neoclouds.

Nebius undercuts both on price while staying self-serve: $3.85/hr on-demand H100 (vs. CoreWeave’s $6.16/hr) with spot at $2.15/hr, plus commitment discounts to 35%. Unlike CoreWeave, Nebius offers single-GPU presets, making it viable for teams that don’t need 8-GPU minimums.

Tier 2: Developer-friendly and specialist neoclouds (ranks 4-8)

RunPod rank 4 on the strength of its price-to-reliability ratio: Community Cloud at $1.99/hr H100 (no SLA, ~97-99% uptime) or Secure Cloud at $2.89/hr (~99.5% uptime, SOC 2/HIPAA/GDPR), both with no egress fees and per-second billing. It is the default recommendation for small-to-mid AI teams that don’t need CoreWeave/Lambda-scale clusters. Full comparison against Lambda and Vast.ai in our RunPod vs. Lambda vs. Vast.ai guide.

Crusoe pairs a distinctive infrastructure story — stranded natural gas and flared energy powering Tier III mobile data centers across 86 sites — with genuine growth: revenue rose from $276 million (2024) to nearly $1 billion (2025), a 262% jump, with cloud ARR up 150% year over year. On-demand single-GPU pricing (~$3.90/hr) is competitive though not the cheapest; Microsoft, Databricks, and Together AI are customers.

Together AI occupies a hybrid niche: per-token serverless inference, dedicated endpoints, and Instant GPU Clusters (8 to 4,000+ GPUs, InfiniBand) at $5.49/hr on-demand or $3.49/hr reserved. Best fit for teams that want to move between renting raw GPUs and consuming a managed inference API without changing providers.

FluidStack posts some of the lowest enterprise-grade rates tracked — ~$2.10/hr H100, ~$2.30/hr H200 — positioning it below CoreWeave and Lambda on price while still targeting production, not marketplace, customers.

TensorWave is the AMD specialist: MI355X from $2.85/GPU-hr, MI325X from $1.95/GPU-hr, with MI300X frequently sold out given AMD’s thinner rental market. The right choice specifically for AMD Instinct workloads rather than a general-purpose pick.

Tier 3: Marketplaces and budget neoclouds (ranks 9-11)

Vast.ai is the price floor for widely available H100 capacity (~$1.53-1.87/hr on-demand, 50%+ lower on interruptible instances) via a peer-to-peer marketplace of 40+ data centers and thousands of independent hosts. There is no platform-level SLA — reliability depends entirely on the individual host — making it best suited to interruption-tolerant, checkpoint-friendly workloads.

Vultr spans an unusually wide $0.118-19.18/hr range across its GPU catalog (A16 through B200) and is the cheapest tracked source for AMD MI300X ($1.85/hr) and MI355X ($2.59/hr) pods, undercutting even TensorWave on those SKUs.

DigitalOcean trades price for simplicity: GPU Droplets (H100, H200, L40S, MI300X) at $3.39/hr with per-hour billing and no long-term commitment, a 12-month reserve trimming only to $3.26/hr — a minor discount that signals DigitalOcean is optimizing for ease of use, not lowest cost.

Tier 4: Hyperscalers (ranks 12-15)

The big four price 2-6x above neoclouds for identical silicon but remain the default for regulated industries, existing enterprise agreements, and workloads already integrated into a specific cloud’s ecosystem.

Hyperscaler	H100 on-demand	Notable model	Why it ranks where it does
Oracle Cloud (OCI)	$10/GPU-hr flat	BM.GPU.H100.8 bare metal	Only hyperscaler with true bare metal (no virtualization overhead); simplest flat-rate pricing among the four
Google Cloud	~$10.98-11.06/GPU-hr	A3 (H100), A3 Ultra (H200)	Best fit if already standardized on Vertex AI, TPUs, or BigQuery; A4 (B200) pricing not yet on the public list
AWS	~$5.19-6.87/GPU-hr (Capacity Blocks)	P5, P5e, P5en	Broadest compliance/procurement ecosystem; Capacity Block rates rose ~20% July 1, 2026
Azure	~$12.29/GPU-hr	ND H100 v5	Highest tracked on-demand rate among major providers; strongest for Microsoft-stack and OpenAI-adjacent tooling

Spot/preemptible pricing cuts hyperscaler rates 70-82% with short eviction notice (as low as 30 seconds on Azure), and 1-year reserved terms discount roughly 35% — closing much of the gap to neocloud on-demand rates for teams that can tolerate interruption or commit in advance.

Pricing by GPU model across the market

GPU	Tracked on-demand range	Typical/median	Cheapest tier	Most expensive tier
L40S (48GB)	$0.48-3.50/hr	~$1.55/hr	Verda (~$0.48)	Oracle Cloud (~$3.50)
A100 (80GB)	$1.09-5.07/hr	~$1.85/hr	Thunder Compute (~$1.09)	GCP a2-ultragpu (~$5.07)
H100 (80GB)	$1.49-11.06/hr	~$3.72/hr	Vast.ai/FluidStack	Google Cloud A3
MI300X (192GB, AMD)	$1.85-7.86/hr	~$3.00/hr	Vultr	Azure
H200 (141GB)	$2.30-13.78/hr	~$4.15/hr	FluidStack	Azure ND H200 v5
MI355X (288GB, AMD)	$2.59-8.60/hr	~$4.95/hr	Vultr	Oracle bare metal
B200 (180GB)	$3.50-16.11/hr	~$6.20/hr	Vultr	Azure/AWS premium SKUs
GB200 (per GPU, NVL72)	$10.50-27.04/hr	~$17-18/hr	CoreWeave	Azure bare metal

Data compiled from GetDeploying, AIMultiple, Thunder Compute, and provider pricing pages, current as of the GPU price index update on 2026-08-05. GB200/B200 figures understate real availability: most GB200 capacity and much B200 volume moves through quote-only reserved contracts rather than published on-demand listings.

Reliability and enterprise readiness compared

Factor	Hyperscalers (AWS/Azure/GCP/OCI)	Enterprise neoclouds (CoreWeave/Lambda/Nebius)	Budget neoclouds (RunPod/Vultr/DigitalOcean)	Marketplace (Vast.ai)
Uptime SLA	99.9%+ contractual, platform-wide	99.5-99.9% on managed tiers	None to ~99.5% depending on tier	None (host-dependent)
Compliance	Broadest: FedRAMP, HIPAA, SOC 2, ISO, PCI	SOC 2, HIPAA on managed tiers	SOC 2/HIPAA on some (RunPod Secure)	Varies by host
Multi-node training	Yes, at premium pricing	Yes, core product (CoreWeave, Lambda)	Limited to modest scale	Limited, single-node focus
Self-serve single GPU	Yes	No (CoreWeave), yes (Nebius, Lambda)	Yes	Yes
Billing granularity	Per hour, some per-second	Per hour/minute	Per second common	Per second
Egress fees	Yes, $0.08-0.12/GB typical	Varies; some waive	Often waived (RunPod, Lambda)	Host-priced

What this means for buyers

The right provider depends on workload shape more than headline price:

• Large-scale, deadline-driven training (100+ GPUs, multi-week runs): CoreWeave or Lambda. Guaranteed multi-node capacity and InfiniBand interconnect are worth the premium over marketplace rates when a stalled training run costs more than the rate difference.
• Startup or mid-size team shipping an AI product: RunPod or Nebius. Both post competitive rates with real reliability tiers and no egress fees — the best balance of cost and operational safety for teams without dedicated infrastructure staff.
• Regulated industry or existing enterprise cloud agreement: AWS, Azure, Google Cloud, or Oracle. Compliance certification breadth and procurement relationships outweigh the 2-4x price premium for finance, healthcare, and government buyers.
• AMD-specific workloads: TensorWave for guaranteed MI300X/MI355X access, Vultr if MI300X/MI355X price is the only variable that matters.
• Batch, interruption-tolerant, or exploratory work: Vast.ai. Interruptible pricing runs 50%+ below its own on-demand rate, and on-demand instances themselves already undercut every managed provider — checkpoint often and treat any single host as disposable.
• Sustainability-weighted procurement: Crusoe’s stranded-energy model is the only one in this ranking built around that constraint specifically, at competitive (not lowest) pricing.

Run the same job on two shortlisted providers in parallel before committing to a reserved term — published rates are a starting point, not the total cost once egress, storage, and preemption overhead are added. Track live rates across all these GPU models in the GPU price index, and see H100 rental prices for how this market has moved since 2023. For colocation-side infrastructure decisions that pair with GPU rental — housing your own hardware instead of renting — compare against the colocation price index or request a quote.