We track pricing data from 18 GPU cloud providers. We see which ones update their pricing, which ones have GPUs in stock, and which ones are quietly becoming more expensive. After 12 months of data collection and cross-referencing thousands of data points, here is our brutally honest report card for every GPU cloud provider in 2026. No sponsorships. No affiliate bias in the rankings. Just data.
The Grading Criteria
Each provider is scored across five dimensions: Pricing (how competitive vs. market median), Availability (can you actually get the GPU you want?), Reliability (uptime, instance stability, data persistence), UX (API quality, dashboard, docs), and Hidden Costs (egress, storage, idle billing, surprise fees).
Enterprise Tier
AWS
C+The most expensive GPU cloud by a wide margin. An H100 on p5.48xlarge costs $98.32/hr (for 8x). That is $8.46/hr per H100 equivalent — 3-6x more than marketplace providers. You get SOC 2 compliance, global availability, and enterprise support. But you pay for it. Aggressively. The C+ is because you also get hit with egress fees ($0.09/GB), EBS charges, and the most complex pricing page in cloud computing.
GCP
B-Slightly more competitive than AWS on GPU pricing, and their A2 and G2 instances are reasonably priced. The A100 40GB at $3.67/hr is still 2-3x market rate, but the spot pricing can bring it down to $1.10/hr — genuinely competitive. GCP loses points for the "local SSD is ephemeral" trap (your data disappears when you stop the instance) and quota request nightmares on H100s.
Azure
CThe most confusing GPU pricing of any provider. Multiple VM series (NC, ND, NV) with overlapping GPU options and pricing that varies by region. H100 on ND H100 v5 at $3.67/hr (per GPU equivalent) is competitive for enterprise, but getting approved for quota takes days to weeks. The UI is enterprise Java from 2014. Azure is where GPUs go when CIOs mandate Microsoft stack.
ML-Focused Tier
Lambda Labs
A-Clean pricing, ML-native UX, no hidden fees. H100 at $2.49/hr on-demand — not the cheapest, but predictable and reliable. Storage and egress included in the price (rare). The catch: limited GPU selection (mainly H100 and A100) and limited regions. Lambda Labs is the "just works" provider. The A- instead of A is because on-demand H100 pricing is 40-50% above marketplace alternatives.
RunPod
AThe best overall GPU cloud for individual developers and small teams. Wide GPU selection (4090, A100, H100, L40S), competitive pricing ($0.39/hr 4090, $1.29/hr H100), and a good dashboard. Serverless GPU inference is a killer feature — pay per second of compute time. The network volume system ($0.07/GB/mo) means your data persists between runs. Only downside: local disk is ephemeral and support can be slow.
Vast.ai
B+The cheapest GPU cloud, period. RTX 3090 at $0.07/hr, 4090 at $0.19/hr, A100 at $0.34/hr — all spot prices that are 2-5x below any managed provider. The trade-off: it is a P2P marketplace. Hosts can reclaim your instance. Disk is ephemeral. Uptime varies by host. Network bandwidth varies by host. But for batch jobs, experiments, and cost-sensitive inference, nothing beats Vast.ai on price.
Rising Stars
DigitalOcean GPU Droplets
A-Surprising entry from DigitalOcean. H100 GPU Droplets with their signature simple UX. Competitive pricing, reliable infrastructure, and the trust of a major cloud provider without the AWS pricing premium. For teams that want "RunPod reliability with Lambda Labs simplicity," DigitalOcean is the answer. Still building out GPU selection but promising.
TensorDock
B+Marketplace model similar to Vast.ai but with better curation and more consistent host quality. Pricing is competitive — often within 10% of Vast.ai — with better reliability. The API is developer-friendly and they support custom configs. Good for teams that want marketplace prices with fewer surprises.
The Summary Table
| Provider | Grade | Best For | Avoid If |
|---|---|---|---|
| RunPod | A | Best overall for individuals/startups | You need enterprise SLAs |
| Lambda Labs | A- | ML teams wanting simplicity | Budget is tight |
| DigitalOcean | A- | Teams wanting trusted infra | You need variety of GPUs |
| Vast.ai | B+ | Maximum cost savings | Mission-critical production |
| TensorDock | B+ | Budget + better reliability | Enterprise compliance needs |
| GCP | B- | Enterprise with spot pricing | You dislike quota requests |
| AWS | C+ | Enterprise with compliance needs | Budget matters at all |
| Azure | C | Microsoft-mandated environments | You value developer experience |
See the full provider breakdown: Explore detailed pricing pages for every provider on our GPU price comparison — filter by provider to see exactly what each one charges.
Detailed Scoring: Top 10 Providers Across All 5 Categories
Each provider scored 1–10 on five categories. The overall grade is the unweighted average. Full breakdown:
| Provider | Price | Avail. | Reliab. | UX | No Hidden $ | Overall |
|---|---|---|---|---|---|---|
| RunPod | 9/10 | 8/10 | 7/10 | 9/10 | 9/10 | A |
| Lambda Labs | 7/10 | 7/10 | 9/10 | 8/10 | 8/10 | A- |
| Vast.ai | 10/10 | 7/10 | 6/10 | 6/10 | 8/10 | B+ |
| CoreWeave | 6/10 | 9/10 | 9/10 | 7/10 | 7/10 | B+ |
| Modal | 6/10 | 8/10 | 9/10 | 10/10 | 9/10 | A- |
| Paperspace | 6/10 | 7/10 | 8/10 | 9/10 | 8/10 | B+ |
| Crusoe | 7/10 | 7/10 | 8/10 | 7/10 | 8/10 | B |
| Nebius | 8/10 | 7/10 | 8/10 | 7/10 | 8/10 | B+ |
| TensorDock | 9/10 | 6/10 | 7/10 | 6/10 | 8/10 | B |
| AWS | 4/10 | 9/10 | 10/10 | 6/10 | 4/10 | C+ |
The Hidden Cost Audit: What Each Provider Charges Beyond GPU/hr
The headline $/hr is rarely the final bill. Egress, storage, idle billing, and snapshot charges compound — sometimes doubling the real cost. The audit:
| Provider | Egress | Storage | Stopped Billing | Snapshot |
|---|---|---|---|---|
| AWS | $0.09/GB (first 10TB) | $0.10/GB-mo (EBS) | EBS still billed | $0.05/GB-mo |
| GCP | $0.12/GB (NA→EU) | $0.04/GB-mo (PD) | PD still billed | $0.026/GB-mo |
| Azure | $0.087/GB (first 10TB) | $0.184/GB-mo (Premium SSD) | Disk still billed | Tiered pricing |
| CoreWeave | $0.02/GB | $0.07/GB-mo | Vol still billed | Included |
| Lambda Labs | Free (1TB/mo) | $0.20/GB-mo | Volume billed | Manual |
| RunPod | Free | $0.10/GB-mo (network vol) | $0.006/hr disk | Manual |
| Vast.ai | Host-dependent | $0.05–0.15/GB-mo | No idle billing | N/A |
| Modal | Free (within limits) | Included in compute | $0 (serverless) | N/A |
| Crusoe | Free | $0.06/GB-mo | Volume billed | Included |
| Nebius | Free (1TB/mo) | $0.072/GB-mo | Volume billed | Included |
The egress trap: AWS, GCP, and Azure charge $0.08–$0.12/GB to move data out of their network. A single 100 GB model checkpoint download = $8–12. Multiply by a 50-experiment month and egress alone exceeds your compute bill. RunPod, Crusoe, Lambda, Nebius, and Modal don't charge egress in normal use — a real differentiator that doesn't appear in any $/hr comparison.
Scoring Methodology
5 categories, 1–10 each. Overall grade = unweighted average. A+ = 9.5+, A = 8.5–9.4, B+ = 7.5–8.4, B = 6.5–7.4, C+ = 5.5–6.4, C = 4.5–5.4, D = below 4.5.
Price (1–10): Median H100 SXM on-demand price vs market median ($2.49/hr). Score 10 = ≤$1.50/hr median. Score 1 = ≥$6/hr median.
Availability (1–10): Percentage of tracked GPU types with at least one in-stock listing across a 4-week window. Score 10 = 90%+ availability. Score 1 = below 30%.
Reliability (1–10): Combination of SLA commitment (or absence), historical uptime data from public status pages, and incident response patterns over 12 months.
UX (1–10): Time from signup to running GPU, clarity of pricing UI, quality of API/CLI tooling, documentation quality. Measured by hands-on testing in March–April 2026.
No Hidden Costs (1–10): Penalizes egress fees, idle-while-stopped billing, mandatory minimum commitments, surprise snapshot/storage charges. Score 10 = nothing beyond the GPU $/hr. Score 1 = up to 50% surcharge on real bills vs headline price.
Conflicts of interest: GPU Tracker takes no payment or sponsorship from any provider for placement or scoring. This report card reflects our independent assessment.
Provider Selection FAQ
What is the best GPU cloud provider in 2026?▾
Depends on use case. For most ML engineers and small teams: RunPod (A) — best balance of price, GPU variety, and developer UX. For production AI startups: Modal (A-) for serverless, CoreWeave (B+) for dedicated clusters. For research/training-heavy teams: Lambda Labs (A-). For pure cost: Vast.ai (B+) — cheapest per GPU-hour if you can tolerate variable hosts.
Why did AWS get a C+?▾
AWS scores 10/10 on reliability and 9/10 on availability — these are world-class. But the GPU $/hr is 2–5× higher than specialty providers, and AWS adds egress fees ($0.09/GB), EBS storage that bills even while instances are stopped, and complex pricing that surprises new teams. The C+ is the gap between "premium infrastructure" and "premium pricing for things you don't need at small scale."
Is Vast.ai actually reliable enough for production?▾
Yes, with caveats. Vast.ai's marketplace model means individual hosts vary in quality. For production: filter by hosts with reliability score 95%+ and DLPerf scores in your range, use only verified data center hosts (avoid home/colo for production), and always run with checkpointing on spot. Many AI startups run inference on Vast.ai successfully — they just engineer around the variability.
Should I pick a provider based on which GPUs they have?▾
Yes — availability matters more than price for many teams. RunPod, Vast.ai, and Verda have the widest selection (RTX 4090, RTX 5090, A100, H100, H200, B200 all available). Lambda is strongest on H100/H200. AWS p5 series is H100-only. If you need a specific GPU (e.g., MI300X), narrow first by inventory then by price.
What is the cheapest GPU cloud provider overall?▾
Vast.ai for raw spot prices (RTX 3090 from $0.05/hr, H100 from $1.49/hr). Verda for cheapest H100 on-demand ($0.80/hr spot). RunPod for the best price-with-reliability tradeoff (H100 spot $1.25/hr, on-demand from $1.99/hr). See the live cheapest GPU page for current minimums updated every 6 hours.
How do enterprise providers (CoreWeave, Oracle, Nebius) differ from RunPod?▾
Enterprise providers offer SLAs (99.9%+ uptime), dedicated VLANs, SOC 2/HIPAA compliance, and reserved capacity contracts. RunPod is closer to a developer marketplace — fast to start, hourly billing, but no formal SLA on Community Cloud. Rule of thumb: if you're processing PHI, financial data, or have an SLA to your own customers, use CoreWeave/Lambda/Nebius. Otherwise RunPod or Vast.ai is fine.
Do these grades change over time?▾
Yes. We update the report card quarterly. Recent shifts: Vast.ai moved from B to B+ as reliability improved with more verified hosts. CoreWeave moved from A- to B+ as price competitiveness slipped vs newer entrants. AWS held steady at C+ — the price/value gap persists.
Why isn't Together AI graded here?▾
Together AI is primarily an inference-as-a-service provider (per-token billing on hosted models) rather than a raw GPU cloud. The grading methodology assumes you rent GPUs and run your own software. We have a separate inference-API comparison covering Together, Anyscale, Fireworks, and others.
What about regional GPU cloud providers (OVHcloud, Scaleway, Hyperstack)?▾
European regional providers fill an important compliance gap — EU-only data residency, GDPR-aligned, predictable contracts. They typically cost 20–40% more than RunPod/Vast.ai for the same GPU but are the right call for EU-regulated workloads. Hyperstack stands out among them for transparent pricing.
How accurate are the scoring categories?▾
Price and Hidden Cost scores are objective (calculated from our live dataset and published terms). Reliability and Availability are based on observed data over 12 months. UX is subjective — our team tested all providers in March–April 2026 with the same baseline workflow. Disagreements with our UX scores are reasonable; the numbers are not benchmarks.
