Which is cheapest for renting an H100: RunPod, Lambda, or Vast.ai?

Vast.ai, at roughly $1.53-1.87/hr for marketplace H100s, with interruptible instances 50%+ cheaper again. RunPod is next: H100 PCIe from $1.99/hr (Community Cloud) and H100 SXM from $2.69/hr. Lambda charges $3.99-4.29/hr for H100 SXM — a 40-100% premium over the other two — in exchange for a 99.9% uptime SLA and InfiniBand-connected clusters.

How much does a B200 cost per hour in 2026?

RunPod lists B200 (180GB) at $5.89/hr on both Community and Secure Cloud. Lambda's on-demand B200 SXM6 runs $6.69-6.99/GPU/hr depending on instance size. Vast.ai marketplace B200 pricing fluctuates with supply but generally undercuts both when available. For context, B200 rates across more providers are tracked in our GPU price index.

Is Vast.ai reliable enough for production workloads?

Not for uptime-critical ones. Vast.ai is a peer-to-peer marketplace with no platform-level SLA — reliability depends on the individual host, which ranges from data-center-grade operators to consumer hardware. Host reliability scores and datacenter-verified filters mitigate this, and on-demand (non-interruptible) instances are stable in practice, but anything requiring guaranteed uptime belongs on RunPod Secure Cloud (~99.5%) or Lambda (99.9% SLA).

What is the difference between RunPod Community Cloud and Secure Cloud?

Community Cloud runs on vetted third-party hosts with no SLA and uptime around 97-99%, at prices 10-30% below Secure Cloud (e.g., H100 PCIe $1.99 vs $2.89/hr; H200 $3.59 vs $4.39/hr). Secure Cloud runs in certified data centers with ~99.5% uptime, dedicated hardware, and the compliance posture (SOC 2, HIPAA, GDPR) that enterprise and regulated workloads require.

Do these GPU clouds charge egress fees?

Lambda explicitly charges no egress fees, and RunPod likewise does not bill for data transfer — a meaningful saving versus hyperscalers, where moving a few TB of checkpoints out can cost hundreds of dollars. On Vast.ai, bandwidth pricing is set per-host, so check the listing. Storage: RunPod network volumes run $0.05-0.07/GB/month (high-performance $0.14); Lambda bundles substantial NVMe with instances; Vast.ai storage is host-priced.

Which GPU cloud is best for multi-node training?

Lambda. Its 1-Click Clusters provision 16 to 2,000+ InfiniBand-connected HGX H100/B200 GPUs with per-minute billing, which neither RunPod's standard pods nor Vast.ai's marketplace matches for large-scale distributed training. RunPod handles single-node and modest multi-node jobs well; Vast.ai is strongest for single-node, interruption-tolerant workloads.

RunPod vs Lambda vs Vast.ai: Which GPU Cloud to Rent From in 2026

RunPod vs Lambda vs Vast.ai in 2026: H100 from $1.53-3.99/hr, H200 and B200 pricing, marketplace vs managed cloud, SLAs, and which provider fits your workload.

RunPod, Lambda, and Vast.ai are the three GPU clouds most builders shortlist below the hyperscalers, and they occupy three distinct points on the price-reliability curve. Vast.ai is the cheapest — a peer-to-peer marketplace with H100s around $1.53-1.87/hr and no SLA. RunPod is the middle path — a developer-focused platform with H100s from $1.99/hr, per-second billing, and a Secure Cloud tier with SOC 2/HIPAA compliance. Lambda is the premium managed AI cloud — H100 SXM at $3.99-4.29/hr with a 99.9% SLA and InfiniBand clusters to 2,000+ GPUs. Pick by workload tolerance for interruption, not by sticker price alone. Here is the data.

GPU pricing compared (on-demand, per GPU-hour)

Prices below are provider-published on-demand rates as of mid-2026; Vast.ai figures are marketplace averages that move in real time. Live rates across these and other providers are tracked in our GPU price index.

GPU	Vast.ai (marketplace)	RunPod Community	RunPod Secure	Lambda
A100 80GB PCIe	~$0.80-1.20	$1.19	$1.39	$1.99-2.79 (SXM/PCIe)
H100 PCIe	~$1.53-1.87	$1.99	$2.89	—
H100 SXM	~$1.90-2.30	$2.69	$2.99	$3.99-4.29
H200 (141GB)	varies with supply	$3.59	$4.39	~$4.49 when listed
B200 (180GB)	varies with supply	$5.89	$5.89	$6.69-6.99

Sources: RunPod pricing, Lambda pricing, Spheron comparison, ComputePrices.

Three pricing mechanics matter beyond the table:

Vast.ai’s interruptible tier is the true floor. Bidding-based interruptible instances run 50%+ below the marketplace on-demand rates above. For checkpoint-friendly training, that can mean H100 compute under $1/hr — but your job can be preempted whenever a higher bid arrives.
Reserved and spot discounts run 30-50% everywhere. All three offer commitment pricing; Vast.ai reserved instances discount up to 50%, and RunPod savings plans compound with Community Cloud rates.
Billing granularity favors the small. Vast.ai and RunPod bill per second, Lambda per minute. For bursty inference and short experiments, per-second billing plus RunPod’s serverless mode meaningfully cuts waste.

Three different business models

Vast.ai is a marketplace. It owns no GPUs; it matches renters with 40+ data centers and thousands of independent hosts offering 68+ GPU types, and prices are set by supply and demand, not by Vast. That is why it is cheapest — and why quality varies from professional data-center hosts to consumer rigs. Host reliability scores, datacenter-verified badges, and secure-cloud filters are your quality controls, per Lyceum’s comparison.

RunPod is a hybrid platform. Community Cloud aggregates vetted third-party capacity (marketplace economics, ~97-99% uptime, no SLA); Secure Cloud runs on certified data center infrastructure at ~99.5% uptime with dedicated hardware. On top sit developer conveniences: 18+ GPU classes, templates, a serverless autoscaling tier for inference, per-second billing, and compliance certifications (SOC 2, HIPAA, GDPR) that make it the most enterprise-usable of the budget options, per GPUCloudList.

Lambda is a managed AI cloud. A deliberately narrow catalog — H100, H200, B200, A100, mostly SXM with NVLink — sold as single instances or 1-Click Clusters of 16 to 2,000+ InfiniBand-connected GPUs. It is the only one of the three built for serious multi-node distributed training out of the box, and the only one with a contractual 99.9% uptime SLA plus human support over Slack, email, and phone. You pay 40-100% over RunPod/Vast rates for it.

Reliability, support, and hidden costs

Factor	Vast.ai	RunPod	Lambda
Uptime guarantee	None (host-dependent)	~99.5% Secure; 97-99% Community (no SLA)	99.9% SLA
Interruption risk	High on interruptible; low on on-demand	Low on Secure Cloud	Minimal
Multi-node training	Limited; single-node focus	Modest multi-node	1-Click Clusters, 16-2,000+ GPUs, InfiniBand
Egress fees	Host-priced bandwidth	None	None
Storage	Host-priced	$0.05-0.14/GB/month	NVMe bundled with instances
Compliance	Varies by host; verified-DC filter	SOC 2, HIPAA, GDPR	SOC 2; enterprise agreements
Support	Community + docs	Tickets, Discord; solid docs	Human support (Slack/email/phone)
Billing	Per second	Per second	Per minute

The “hidden cost” picture is unusually clean in this trio compared with hyperscalers: no egress charges at Lambda or RunPod means moving datasets and checkpoints is free, whereas the same traffic on a big-three cloud bills at $0.08-0.12/GB. The real hidden cost is engineering time: babysitting preempted Vast.ai jobs, or re-queuing for capacity when a cheap Community Cloud region sells out. Price that time honestly before declaring the marketplace cheaper. If your utilization is high and sustained for 12+ months, also run the rent-vs-colo math — owning GPUs in colocation space can undercut all three; see our data center catalog and colocation price index for the inputs.

A worked example: one month of fine-tuning

Numbers make the trade-offs concrete. Take a typical mid-size job: fine-tuning a 70B-parameter model on 8x H100 SXM for 200 GPU-hours of compute, plus 2 TB of dataset/checkpoint storage and 3 TB of data moved in and out over the month.

Cost component	Vast.ai (interruptible)	RunPod Secure	Lambda
Compute (200 GPU-h x 8)	~$1.10/h x 1,600 = ~$1,760	$2.99/h x 1,600 = $4,784	$3.99/h x 1,600 = $6,384
Storage (2 TB-month)	host-priced, ~$100-200	$0.05-0.07/GB = $100-140	bundled NVMe = $0
Data transfer (3 TB)	host-priced, often ~$0	$0	$0
Preemption overhead	+10-25% reruns = $175-440	~$0	~$0
Approx. total	~$2,050-2,400	~$4,900	~$6,400

The marketplace saves roughly 55-60% versus Lambda even after paying a preemption tax — if the job checkpoints cleanly and nobody is waiting on the result. Flip the assumptions (deadline-driven, multi-node with heavy inter-GPU traffic, compliance requirements) and the Lambda premium buys wall-clock time and certainty that is worth more than the difference. The honest comparison is never hourly rate versus hourly rate; it is total cost per completed job.

Availability: the constraint pricing tables hide

Published prices assume you can actually get the GPU, and in 2026 that is the weakest assumption in the stack. B200 capacity remains allocation-constrained everywhere: Lambda routes serious volume through sales conversations and reserved contracts, RunPod’s B200 pools sell out regionally, and Vast.ai B200 listings appear and vanish with individual hosts. H100s, by contrast, have loosened markedly as A100 fleets retire and Blackwell ramps — which is exactly why marketplace H100 rates have drifted below $2/hr.

Three practical consequences. First, if your training schedule is fixed, reserve — a 30-50% discount for commitment is also a capacity guarantee. Second, multi-region flexibility is worth real money on RunPod and Vast.ai; teams that can run in any region catch cheap capacity that single-region teams miss. Third, watch the curve: GPU rental is the fastest-repricing corner of the data center market, which is why we track it continuously in the GPU price index alongside the slower-moving colocation index.

Which to choose if…

You are a hobbyist, researcher, or running interruption-tolerant batch jobs → Vast.ai. Cheapest H100s on the market (~$1.53-1.87/hr, less on interruptible), per-second billing, no commitment. Checkpoint often, filter for high-reliability hosts, and treat any single instance as disposable.
You are a startup shipping an AI product → RunPod. The best balance: Community Cloud for experiments ($1.99/hr H100 PCIe), Secure Cloud for production (~99.5% uptime, SOC 2/HIPAA), serverless for spiky inference, and no egress fees. This is the default recommendation for most small teams.
You are an enterprise or regulated business → RunPod Secure Cloud or Lambda. RunPod if cost matters and single-node/modest multi-node suffices; Lambda if you need a contractual 99.9% SLA, named support humans, and procurement-friendly paper.
You are training large models across many nodes → Lambda. 1-Click Clusters with InfiniBand from 16 to 2,000+ H100/B200 GPUs is the differentiator; neither competitor offers comparable interconnect at scale. The 40-100% hourly premium is usually cheaper than the wall-clock time lost to slow inter-node networking.
You are cost-optimizing an existing pipeline → split by stage. A common pattern: dataset preprocessing and ablations on Vast.ai interruptible, main training runs on Lambda clusters, production inference on RunPod serverless. Nothing locks you in — all three bill by the minute or second.
You are weighing renting vs owning → do the math at your utilization. At $2/hr an H100 costs ~$17,500/year at full utilization; at 50% utilization rented, ownership in colo starts to win within 18-24 months. Request a quote and we will benchmark colo options against current rental rates.

One final note on lock-in: there is essentially none, and that is the strategic point of this tier of the market. All three providers support standard Docker containers, SSH access, and API-driven provisioning, so a pipeline built on any one of them moves to the others in days, not months. Run your first serious workload on two of the three in parallel, measure real throughput per dollar rather than list price, and let the data pick the winner.

Prices in this market move monthly — B200 rates in particular have been falling as supply ramps. Check the live GPU tracker before committing to reserved terms.