If you have shopped for GPU rentals lately, you have probably noticed the same H100 costs wildly different amounts depending on where you look. I checked three of the biggest GPU clouds on the same day in August 2026 and the spread was bigger than I expected: the same chip runs from $2.69 an hour to $6.16 an hour depending on the vendor, and that is before spot pricing turns the ranking upside down.
I wrote up the full pricing breakdown, screenshots included, on DevToolLab. Here is the short version.
The headline numbers
As of August 22, 2026: an H100 runs $2.69/hr on RunPod's Community Cloud, $3.29 on its Secure Cloud, $3.99 on Lambda, and $6.16 on CoreWeave on demand. Then spot inverts the table entirely: CoreWeave's 8x H100 node drops from $49.24/hr to $19.71 on spot, which works out to $2.46 per GPU, cheaper than anyone's on-demand rate.
The three companies are as different as their prices. CoreWeave IPO'd in March 2025 and posted $5.13 billion in 2025 revenue. Lambda is reportedly lining up an IPO for the second half of 2026. RunPod raised $100 million at a $1 billion valuation in June after turning down buyout offers.
The trap to avoid: always normalize to price per GPU-hour. CoreWeave's pricing page lists $49.24 for an HGX H100 instance, which is 8 GPUs, not one. Divide it out and you get the same $6.16/GPU their own single-GPU column shows. Lambda and RunPod list per-GPU prices directly, so comparing a CoreWeave node quote against either of them without dividing first is the single most common mistake in this kind of shopping.
CoreWeave: reserved fleets, brutal spot discounts
CoreWeave priced its IPO at $40 a share, raised about $1.5 billion at roughly a $23 billion valuation, and grew revenue 168% year over year in 2025. Its attempted $9 billion acquisition of Core Scientific died in an October 2025 shareholder vote, so it is still expanding by contract rather than merger.
The platform is Kubernetes-native, which is exactly right for a managed fleet and exactly wrong for a weekend fine-tune. On-demand per 8-GPU instance: HGX H100 $49.24/hr, HGX B200 $68.80, A100 $21.60. The number worth remembering is spot: HGX H100 falls to $19.71/hr, or $2.46 per GPU, undercutting every on-demand price in this comparison if your job checkpoints often enough to survive preemption.
Lambda: the simplest pricing model of the three
Lambda sells a plain price-per-GPU-hour model: $3.99 for an H100 SXM, $2.79 for an A100 80GB, $6.69 for a B200 SXM6, and $0.79 for a V100 if you still need one. Above self-serve, 1-Click Clusters offer 16 to 2,000+ GPUs on two-week to one-year terms, starting at $9.86 per GPU-hour at 16 B200s and dropping to $8.87 past 256 GPUs.
Lambda raised a Series E of more than $1.5 billion in November 2025 and has reportedly retained Morgan Stanley, J.P. Morgan and Citi for an IPO targeted at H2 2026. What you give up for the simplicity: no spot market and no serverless per-request billing, so if you want pay-per-inference pricing you are looking at RunPod or a platform like the ones in our inference platforms roundup.
RunPod: the cheapest H100 in the comparison
RunPod is the only one of the three running a two-tier market. Secure Cloud sits in vetted data centers (H100 SXM $3.29, A100 SXM $1.59, B200 $6.79), while Community Cloud runs on vetted third-party hosts and undercuts everyone: we watched the H100 SXM price drop to $2.69 on the pricing page itself. Billing is per second, and RunPod charges no fees for data ingress or egress, a detail the bigger clouds tend to skip.
import runpod
pod = runpod.create_pod(
"h100-box",
"runpod/pytorch:2.4.0-py3.11-cuda12.4.1-devel-ubuntu22.04",
gpu_type_id="NVIDIA H100 80GB HBM3",
gpu_count=1,
)
RunPod raised $100 million led by Summit Partners in June 2026 at a $1 billion valuation, reportedly after turning down buyout offers above $500 million. Community Cloud's tradeoff is host-grade variance in reliability, so keep it for fault-tolerant work rather than anything customer-facing.
What it actually costs for a month
Run an 8x H100 node for a 730-hour month and the totals are RunPod Secure at $19,214, Lambda at $23,302, CoreWeave on-demand at $35,945, and CoreWeave spot at $14,388. The most expensive on-demand provider becomes the cheapest option once you allow spot, which is the single fact worth remembering from this whole market. That math only works for jobs that checkpoint well; a customer-facing endpoint on spot is an outage generator, not a discount.
One tier down, an A100 80GB runs $1.39 to $1.59 on RunPod against $2.79 on Lambda, at roughly a third of H100 pricing and still enough to fine-tune 7B and 13B models comfortably. None of these numbers include storage or egress, and RunPod is the only one of the three explicitly promising zero fees on data transfer. The full DevToolLab comparison has the complete quick-reference table plus L40S and GH200 pricing if you are sizing a memory-bound job instead of a straight H100 run.
If you would rather not choose
SkyPilot (Apache 2.0, 10,500+ stars) treats every cloud as one market: declare the GPUs a job needs, and it launches on whichever configured provider is cheapest with capacity available, recovering automatically when a spot node gets preempted. It is the practical answer to a pricing table that changes every quarter, and it pairs especially well with CoreWeave's spot discounts.
Picking one
Cost-sensitive, bursty jobs, solo developer or small team: RunPod, Community Cloud for fault-tolerant work and Secure Cloud when reliability actually matters. Clean training runs with no Kubernetes involved: Lambda, whose flat per-GPU pricing is the least surprising of the three. Reserved fleets, the newest silicon, or spot arbitrage at scale: CoreWeave. Not picking at all: SkyPilot across two or more providers.
References
- Original article on DevToolLab: Best GPU Cloud Providers for AI Workloads in 2026
- RunPod pricing
- Lambda pricing
- CoreWeave pricing
- SkyPilot
- Unit Price Calculator - normalize node prices and per-GPU prices to the same unit
- Electricity Cost Calculator - price the buy-your-own-GPU alternative
Top comments (0)