The 10 Best GPU Cloud Providers for AI Training in 2027
By Kory White · Updated Nov 2026 · 7 min read
Direct answer
Lambda is the top GPU cloud provider for AI training in our scoring: H100 SXM from $3.99 and B200 from $6.69 per GPU-hour, plus 1-Click Clusters from 16 to 2,000+ NVIDIA B200 or H100 GPUs. CoreWeave is #2 with the broadest published range, from H100 to GB200, and Together AI is #3 with a full on-demand, preemptible and reserved price table. Prices are vendor-stated from each provider’s own pages read on November 2, 2026.
On-demand H100 SXM from $3.99 per GPU-hour and B200 SXM6 from $6.69 per GPU-hour in 8-GPU instances; GH200 $2.29 and A100 SXM 80 GB $2.79 per GPU-hour (vendor-stated, read November 2, 2026)
Teams that want published per-GPU prices for H100 and B200 and self-serve clusters sized for training
Self-serve H100 $5.50, H200 $6.00 and B200 $9.65 per GPU-hour; from December 1, 2026 these become $6.30, $7.50 and $10.20 (vendor-stated, read November 2, 2026)
Teams building multi-thousand GPU training clusters on InfiniBand with Kubernetes or Slurm managed for them
A single H100 for an hour starts at $1.99 with no contract; every other tier is contact for pricing (vendor-stated, read November 2, 2026)
Teams that want stated InfiniBand H100 clusters and a single entry price
Pick by what you want
Best overall
Lambda GPU Cloud
Best newest hardware
CoreWeave Cloud
Best reserved price ladder
Together AI GPU Clusters
Best InfiniBand scale on H100
Voltage Park AI Cloud
Best lowest listed H100 price
Hyperstack GPU Cloud
The ranking
1
Lambda GPU Cloud
Answrs VerdictThe most complete published price list for training clusters of the providers read.
Price: On-demand H100 SXM from $3.99 per GPU-hour and B200 SXM6 from $6.69 per GPU-hour in 8-GPU instances; GH200 $2.29 and A100 SXM 80 GB $2.79 per GPU-hour (vendor-stated, read November 2, 2026)Training clusters: 1-Click Clusters run from 16 to 2,000+ NVIDIA B200 or H100 GPUsCluster price: HGX B200 clusters cost $9.86 per GPU-hour at 16 GPUs and $8.87 at 256+ GPUs; H100 clusters cost $6.16 at 16 GPUs and $5.54 at 256 GPUs, for terms from 2 weeks to 1 yearReserved: Terms over one year are priced by contacting sales; no figure is publishedNot listed: H200 does not appear on the pricing page, and the page does not mention InfiniBand
Lambda publishes on-demand H100 SXM from $3.99 and B200 SXM6 from $6.69 per GPU-hour in 8-GPU instances, plus GH200 at $2.29 and A100 SXM 80 GB at $2.79. Its 1-Click Clusters run from 16 to 2,000+ NVIDIA B200 or H100 GPUs: B200 clusters cost $9.86 per GPU-hour at 16 GPUs and $8.87 at 256+, and H100 clusters cost $6.16 at 16 and $5.54 at 256, for terms from 2 weeks to 1 year. Longer reservations are contact sales, H200 is not listed, and the page does not mention InfiniBand.
Go ifTeams that want published per-GPU prices for H100 and B200 and self-serve clusters sized for training
Skip ifYou need H200 or a published interconnect specification
Why #1 over #2: It publishes cluster prices at three sizes, where CoreWeave publishes instance prices only.
2
CoreWeave Cloud
Answrs VerdictThe broadest published GPU range, from H100 to GB200, with a stated committed discount.
Price: On-demand 8-GPU HGX H100 instance $49.24 per hour, HGX H200 $50.44 and HGX B200 $68.80 in North America (vendor-stated, read November 2, 2026)Newest hardware: GB200 NVL72 instances (4 GPUs each) are $42.00 per hour; GB300 NVL72 and HGX B300 on-demand are contact salesCommitted use: Up to 60% discounts over on-demand prices for committed usageSpot: HGX H100 spot is $19.71 per hour and HGX H200 spot is $20.93 per hour in North AmericaInterconnect: The pricing page does not mention InfiniBand or NVLink
CoreWeave lists on-demand 8-GPU instances in North America: HGX H100 at $49.24 per hour, HGX H200 at $50.44 and HGX B200 at $68.80, with GB200 NVL72 instances of 4 GPUs at $42.00. GB300 NVL72 and HGX B300 on-demand are contact sales. It states up to 60% discounts over on-demand prices for committed usage, and spot HGX H100 at $19.71 per hour. The pricing page does not mention InfiniBand or NVLink.
Go ifTeams that want the newest Blackwell hardware with published instance prices and committed-use discounts
Skip ifYou need a per-GPU price for a single-GPU H100 or a stated interconnect on the pricing page
Why #2 over #3: It lists GB200 and a stated committed discount, where Together AI leaves GB200 to contact sales.
3
Together AI GPU Clusters
Photo: Pokiiri / CC BY-SA 4.0 via Wikimedia Commons
Answrs VerdictThe clearest published reserved-price ladder, though cluster scale is not stated.
Price: On-demand HGX H100 $3.99, HGX H200 $5.99 and HGX B200 $8.19 per GPU-hour; preemptible H100 $1.99 (vendor-stated, read November 2, 2026)Reserved H100: Reserved H100 falls from $3.69 per GPU-hour for 3 to 30 days to $3.19 for 91 to 180 daysReserved B200: Reserved B200 falls from $7.99 for 3 to 30 days to $6.79 for 91 to 180 daysLonger terms: Reservations of 181+ days and GB200 NVL72 are by contacting salesNot listed: The pricing page does not list InfiniBand or a maximum cluster size
Together AI lists on-demand HGX H100 at $3.99, HGX H200 at $5.99 and HGX B200 at $8.19 per GPU-hour, with preemptible H100 at $1.99. Reserved H100 falls from $3.69 for 3 to 30 days to $3.19 for 91 to 180 days, and reserved B200 from $7.99 to $6.79. Reservations of 181+ days and GB200 NVL72 are contact sales, and the pricing page does not list InfiniBand or a maximum cluster size.
Go ifTeams that want a full table of on-demand, preemptible and reserved prices for H100, H200 and B200
Skip ifYou need the interconnect or maximum cluster size stated before you buy
Why #3 over #4: It publishes reserved prices by term, where Crusoe leaves reservations to contact sales.
4
Crusoe Cloud
Answrs VerdictStrong stated scale and network, but only self-serve prices are published.
Price: Self-serve H100 $5.50, H200 $6.00 and B200 $9.65 per GPU-hour; from December 1, 2026 these become $6.30, $7.50 and $10.20 (vendor-stated, read November 2, 2026)Hardware: Lists NVIDIA GB300, B300, Vera Rubin, H200 and H100 plus AMD MI300X and MI355X; the footer also lists GB200 and B200Scale and network: GPUs connect over InfiniBand and high-performance Ethernet, with clusters from single-node experiments to multi-thousand GPU clusters and zero egress feesCompliance (vendor-stated): ISO 27001, ISO 42001, SOC 2 Type I and II and GDPRReserved: Reserved rates are by contacting sales; no figures are published
Crusoe lists self-serve H100 at $5.50, H200 at $6.00 and B200 at $9.65 per GPU-hour, rising from December 1, 2026 to $6.30, $7.50 and $10.20. Its cloud page says GPUs connect over InfiniBand and high-performance Ethernet, with clusters from single-node experiments to multi-thousand GPU clusters, zero egress fees, and ISO 27001, ISO 42001 and SOC 2 Type I and II, all vendor-stated. Hardware includes GB300, B300, Vera Rubin, H200, H100, MI300X and MI355X; reserved rates are contact sales.
Go ifTeams building multi-thousand GPU training clusters on InfiniBand with Kubernetes or Slurm managed for them
Skip ifYou want a published price for every GPU and for reservations
Why #4 over #5: It states InfiniBand and multi-thousand GPU scale, where RunPod states clusters up to 64 GPUs.
5
RunPod GPU Cloud
Photo: 极客湾Geekerwan / CC BY 3.0 via Wikimedia Commons
Answrs VerdictLow entry prices and a two-tier structure, with cluster scale stated at 64 GPUs.
Price: Secure Cloud pods: H100 SXM $3.99, H200 $5.29, B200 $7.99 and B300 $8.99 per hour (vendor-stated, read November 2, 2026)Community Cloud: Community Cloud H100 SXM is $2.69 per hour and B200 is $5.98 per hourClusters: Instant clusters scale up to 64 GPUs with no commitment; H200 SXM clusters are $6.35 per hourEnterprise: Reserved clusters for enterprises of 10,000+ GPUs are quote-onlyPage caveat: The page metadata shows different figures in places, and the page itself says to verify prices in the console
RunPod lists Secure Cloud pods at H100 SXM $3.99, H200 $5.29, B200 $7.99 and B300 $8.99 per hour, with Community Cloud H100 SXM at $2.69 and B200 at $5.98. Instant clusters scale up to 64 GPUs with no commitment, with H200 SXM clusters at $6.35 per hour, and enterprise reserved clusters of 10,000+ GPUs are quote-only. The page metadata shows different figures in places, so it says to verify prices in the console.
Go ifDevelopers who want pay-as-you-go H100, H200 and B200 pods, from a community tier to secure cloud
Skip ifYou need published multi-hundred GPU cluster prices
Why #5 over #6: It publishes pod prices for B300 and a community tier, where Nebius lists on-demand only.
6
Nebius AI Cloud
Answrs VerdictA clean on-demand price list; reservation discounts and networking are not quantified.
Price: On-demand HGX H100 $4.50, HGX H200 $5.40 and HGX B200 $8.50 per GPU-hour; GB200 NVL72 is contact us (vendor-stated, read November 2, 2026)Commitments: Reserving large-scale clusters for multiple months is priced below on-demand, but no percentage is publishedInference GPUs: L40S is from $1.55 per GPU-hour with an Intel CPU and from $1.82 with an AMD CPUTaxes: Prices exclude applicable taxes such as VAT and may changeInterconnect: The prices page does not mention InfiniBand and points to its documentation for networking
Nebius lists on-demand HGX H100 at $4.50, HGX H200 at $5.40 and HGX B200 at $8.50 per GPU-hour, with GB200 NVL72 as contact us and L40S from $1.55 with an Intel CPU and from $1.82 with an AMD CPU. Reserving large-scale clusters for multiple months is priced below on-demand, but no percentage is published. Prices exclude taxes such as VAT, and the prices page does not mention InfiniBand.
Go ifTeams that want H100, H200 and B200 on a single published price list
Skip ifYou need a published discount percentage or interconnect detail on the prices page
Why #6 over #7: It lists H100, H200 and B200 on one table, where Hyperstack states nothing about scale.
7
Hyperstack GPU Cloud
Photo: Javils00 / CC0 via Wikimedia Commons
Answrs VerdictThe lowest published on-demand and reserved prices on the list, but with little stated about scale.
Price: On-demand H100 SXM $3.20, H200 SXM $3.99, B200 $6.00 and A100 80 GB $1.35 per hour, billed per minute (vendor-stated, read November 2, 2026)Reservation: Reservations start from $2.72 for H100 SXM, $2.79 for H200 SXM, $5.10 for B200 and $0.95 for A100Data transfer: The FAQ says there are no additional charges for bandwidth or data transferInterconnect: The pricing page does not mention InfiniBand or a maximum cluster size
Hyperstack lists on-demand H100 SXM at $3.20, H200 SXM at $3.99, B200 at $6.00 and A100 80 GB at $1.35 per hour, billed per minute, with reservations from $2.72 for H100 SXM, $2.79 for H200 SXM, $5.10 for B200 and $0.95 for A100. Its FAQ says there are no additional charges for bandwidth or data transfer. The pricing page does not mention InfiniBand or a maximum cluster size.
Go ifTeams focused on the lowest published on-demand and reserved prices for H100, H200 and B200
Skip ifYou need stated cluster scale or interconnect for large training runs
Why #7 over #8: It publishes lower prices and reservations, where AWS publishes no price on the page read.
8
AWS EC2 P5 UltraClusters
Photo: Tedder / CC BY-SA 4.0 via Wikimedia Commons
Answrs VerdictThe largest stated cluster scale on the list, with no price on the page read.
Price: The instance page lists no prices and links to the EC2 pricing page; it claims up to 40% savings versus previous-generation GPU instancesGPUs: P5 has 8 NVIDIA H100 GPUs and 640 GB of HBM3 per instance; P5e and P5en have 8 NVIDIA H200 GPUs and 1128 GB of HBM3e per instance (vendor-stated, read November 2, 2026)Network: EFA network bandwidth is 3,200 Gbps on P5, P5e and P5enScale: The page says UltraClusters scale up to 20,000 H100 or H200 GPUs on a petabit-scale nonblocking network, up to 20 exaflops
AWS says P5 instances carry 8 NVIDIA H100 GPUs and 640 GB of HBM3, and P5e and P5en carry 8 NVIDIA H200 GPUs and 1128 GB of HBM3e, all with 3,200 Gbps of EFA network bandwidth. The page says UltraClusters scale up to 20,000 H100 or H200 GPUs on a petabit-scale nonblocking network, up to 20 exaflops. It lists no prices, linking to the EC2 pricing page, and claims up to 40% savings versus previous-generation GPU instances.
Go ifLarge enterprises that need the largest stated H100 and H200 cluster scale inside an existing AWS account
Skip ifYou need a published price on the instance page or B200 availability from this page
Why #8 over #9: It states the largest cluster scale, where DigitalOcean states none.
9
DigitalOcean GPU Droplets
Photo: 极客湾Geekerwan / CC BY 3.0 via Wikimedia Commons
Answrs VerdictStraightforward prices and an AMD option, with no cluster scale stated on the page.
Price: On-demand HGX H100 $4.41, HGX H200 $4.47 and AMD MI300X $2.59 per GPU-hour (vendor-stated, read November 2, 2026)Reserved: 12-month reserved is $3.26 for H100, $3.40 for H200 and $1.91 for MI300X per GPU-hour, by contractual commitmentBlackwell: HGX B300 is $7.94 reserved and $6.17 spot; no B200 price is listedBilling: On-demand is billed per second with a 5-minute minimum, and powered-off Droplets are still billed
DigitalOcean lists on-demand HGX H100 at $4.41, HGX H200 at $4.47 and AMD MI300X at $2.59 per GPU-hour, with 12-month reserved at $3.26, $3.40 and $1.91 by contractual commitment. HGX B300 is $7.94 reserved and $6.17 spot, and no B200 price is listed. On-demand is billed per second with a 5-minute minimum, and powered-off Droplets are still billed.
Go ifDevelopers already on DigitalOcean who want simple per-second GPU Droplets and an AMD option
Skip ifYou need multi-node cluster scale or interconnect details
Why #9 over #10: It publishes reserved and spot prices including B300, where Voltage Park publishes one starting price.
10
Voltage Park AI Cloud
Photo: 极客湾Geekerwan / CC BY 3.0 via Wikimedia Commons
Answrs VerdictStrong stated InfiniBand scale on H100, but little else is published.
Price: A single H100 for an hour starts at $1.99 with no contract; every other tier is contact for pricing (vendor-stated, read November 2, 2026)On-demand clusters: Ethernet clusters of 1 to 1016 HGX H100 and InfiniBand clusters of 8 to 1016 HGX H100 at 3200 Gbps, self-serve and up within 15 minutesLong-term reserve: 32 to 8000+ H100 in a single configuration on 3200 Gbps InfiniBand, with contracts of 6+ monthsHardware: Only H100 is listed; the page gives no H200, B200 or GB200 details
Voltage Park says a single H100 for an hour starts at $1.99 with no contract; every other tier is contact for pricing. On-demand Ethernet clusters run 1 to 1016 HGX H100, and InfiniBand clusters run 8 to 1016 HGX H100 at 3200 Gbps, self-serve and up within 15 minutes. Long-term reserve covers 32 to 8000+ H100 on 3200 Gbps InfiniBand with contracts of 6+ months. Only H100 is listed; the page gives no H200, B200 or GB200 details.
Go ifTeams that want stated InfiniBand H100 clusters and a single entry price
Skip ifYou need H200, B200 or published prices for clusters and reservations
How we ranked these
This list covers GPU clouds for training large models, not inference-only hosts. Each provider was scored out of 100 with the same rubric and evidence standard from its own pricing and product pages, read on November 2, 2026: published GPU generations such as H100, H200 and B200 (25 points), published on-demand and reserved pricing (25), stated training scale and interconnect (25), maturity and operating evidence on its own pages (15) and transparency (10). Scores: Lambda eighty-five, CoreWeave eighty-four, Together AI seventy-nine, Crusoe seventy-eight, RunPod seventy-seven, Nebius seventy-five, Hyperstack sixty-nine, AWS sixty-six, DigitalOcean sixty-four, Voltage Park fifty-four. The weights and scores are our judgment from the pages read, not vendor data. No pick was adjusted for any company, and hyperscalers were weighed on the same rubric as smaller clouds. All prices and specs are vendor-stated, change often and exclude taxes; we did not run any workload, and CoreWeave, AWS and several other pages do not state a per-GPU H100 price or interconnect, which the scores reflect. Google Cloud, Microsoft Azure and Oracle were not scored because their pages were not read for this edition. Facts are taken from each maker's own published information; check the official site before you buy.
Questions people ask
What is the best GPU cloud provider for AI training in 2027?
Lambda scores highest here: H100 SXM from $3.99 and B200 from $6.69 per GPU-hour, and 1-Click Clusters from 16 to 2,000+ NVIDIA B200 or H100 GPUs. CoreWeave is #2 on the broadest published GPU range, from H100 to GB200.
Which GPU cloud has the cheapest H100?
Among listed prices, Voltage Park starts a single H100 at $1.99 per hour, and Hyperstack lists on-demand H100 SXM at $3.20 per hour with reservations from $2.72. Lambda, Together AI and RunPod each list $3.99.
Which provider states the largest training cluster?
AWS says P5, P5e and P5en UltraClusters scale up to 20,000 H100 or H200 GPUs. Voltage Park lists 32 to 8000+ H100 on 3200 Gbps InfiniBand under long-term reserve, and Lambda lists clusters up to 2,000+ B200 or H100 GPUs.
Are B200 GPUs available on demand?
Yes at several providers: Hyperstack lists B200 at $6.00 per hour, Lambda from $6.69, RunPod $7.99, Together AI $8.19, Nebius $8.50 and Crusoe $9.65 self-serve. CoreWeave lists an 8-GPU HGX B200 instance at $68.80 per hour.
Lowest listed on-demand H100 price per GPU-hour, read November 2, 2026 (USD; Voltage Park is a starting price; CoreWeave and AWS are not listed per GPU) at a glance
Voltage Park AI Cloud1.99 USD per GPU-hour
Hyperstack GPU Cloud3.2 USD per GPU-hour
Lambda GPU Cloud3.99 USD per GPU-hour
Together AI GPU Clusters3.99 USD per GPU-hour
RunPod GPU Cloud3.99 USD per GPU-hour
DigitalOcean GPU Droplets4.41 USD per GPU-hour
Nebius AI Cloud4.5 USD per GPU-hour
Crusoe Cloud5.5 USD per GPU-hour
Lowest listed on-demand H100 price per GPU-hour, read November 2, 2026 (USD; Voltage Park is a starting price; CoreWeave and AWS are not listed per GPU), as published by each maker. Picks without a published figure are left off.
