What is the cheapest way to rent an H100?

Peer-to-peer marketplaces like Vast.ai post the lowest on-demand rates, observed around ~$1.87/hr in mid-2026 community pricing digests, though availability and reliability vary by host. Specialist clouds like RunPod's community tier (~$1.99/hr) sit close behind with more standardized infrastructure. These are not independently verified by LocalRig — check current listings before committing a training run to a host.

Why is AWS or Azure H100 pricing so much higher than specialist clouds?

Hyperscaler on-demand rates (AWS ~$6.88/hr, Azure ~$12.29/hr in mid-2026 pricing digests) bundle compliance certifications, enterprise SLAs, dedicated support, and integration with an existing cloud ecosystem (VPC, IAM, egress-free intra-cloud transfer). A hobbyist or indie fine-tuner is paying for guarantees they will never invoke.

Is a cheap H100 rental actually the same chip as an expensive one?

The silicon is the same NVIDIA H100, but the surrounding infrastructure is not — networking (NVLink/InfiniBand vs. shared PCIe), storage persistence, uptime SLA, and support response time all differ by provider tier, and those differences are what the price gap is buying.

Should I rent an H100 or use a smaller GPU for fine-tuning?

Only if your model and batch size actually require H100-class VRAM and bandwidth. Many fine-tuning jobs on 7B-13B models run fine on a single A100 or even a 24GB consumer card; renting an H100 for a job that does not need it is the most common way to overpay in this market. See the cloud fine-tuning vs. inference guide before booking an H100 instance.

H100 Rental Price Comparison 2026: $1.38 to $12+ an Hour for the Same Chip

An H100 rented on a peer-to-peer marketplace in mid-2026 costs roughly $1.38-$1.99 an hour. The same chip rented from Azure costs ~$12.29 an hour. That is not a typo and it is not a scam on either end — it is two different products wearing the same GPU model number, and most of the aggregator tables that rank #1 on Google for “H100 price comparison” will show you the numbers without telling you why they diverge this much.

This page does that second part. It also seeds a live price table LocalRig will maintain going forward, so treat every number below as a dated snapshot, not a permanent quote.

What does an H100 cost per hour in 2026?

On-demand H100 pricing spans roughly $1.38 to $12+ per hour depending on provider tier, with a median cluster around $2.29-$3.12/hr among the GPU-specialist and marketplace providers that most cost-conscious renters actually use. Hyperscalers sit well above that median, not because their hardware is different, but because of what is bundled around it.

The table below compiles figures from community pricing digests (IntuitionLabs, Thunder Compute, Spheron — 2026) alongside provider pricing pages, checked 2026-06-29. Cloud GPU pricing changes without notice — spot/interruptible tiers, reserved commitments, and regional availability all move these numbers, sometimes weekly. Verify current pricing directly with the provider before budgeting a training run against any figure here.

Provider	Tier	~On-demand H100 rate	Category
Vast.ai	marketplace, on-demand	~$1.87/hr	Peer-to-peer marketplace
RunPod	community cloud	~$1.99/hr	GPU specialist
Lambda	on-demand	~$2.99/hr	GPU specialist
AWS	EC2 P5, on-demand	~$6.88/hr	Hyperscaler
Azure	ND H100 v5, on-demand	~$12.29/hr	Hyperscaler

That is roughly a 6.6x spread between Vast.ai’s marketplace floor and Azure’s on-demand rate, and close to 9x if you catch Azure at list price against a discounted marketplace host. Nothing in that spread is fake — it reflects genuinely different products.

Why do hyperscalers charge 4-9x more for the same chip?

Because you are not just renting a GPU from AWS or Azure — you are renting a GPU wrapped in compliance certifications, an existing cloud ecosystem, and enterprise support, and the hyperscaler rate prices all of that in whether you use it or not.

Specifically, the premium buys:

Compliance and certification. SOC 2, HIPAA, FedRAMP, and industry-specific attestations that regulated buyers (healthcare, finance, government contractors) cannot skip. A specialist cloud or marketplace host typically does not carry these.
Egress-free ecosystem integration. If your data already lives in S3 or Azure Blob, training inside the same cloud avoids egress fees and lets you use existing IAM, VPC, and networking policy. That integration has real value — for a team already deep in that ecosystem.
Enterprise SLA and support. A named account team, guaranteed response times, and uptime commitments with financial penalties attached. A specialist cloud’s support is usually a ticket queue or Discord channel — often fine, but not contractually backstopped.
Capacity guarantees. Hyperscalers can promise a reserved instance will be there when you need it. Marketplace GPUs are supply-and-demand priced and can simply be unavailable at peak times.

None of that is padding — it is a real product for a real buyer. The problem is that most people searching “H100 price comparison” are not that buyer.

Who is actually paying for compliance they don’t use?

If you are a solo developer, an indie fine-tuner, or a small team running experiments without a regulatory mandate, you are very likely the buyer hyperscaler pricing was not built for — and paying for HIPAA attestation you will never invoke is the single most common overspend in this market.

This is the constraint-logic split LocalRig uses across the local-vs-cloud cluster: match the provider tier to what you actually need, not to what ranks first in a pricing table.

Hobbyist / indie fine-tuner, no compliance requirement, cost-sensitive. The marketplace and specialist tiers (Vast.ai, RunPod) are the right fit. You are trading some reliability and support depth for a 3-6x lower rate, and for a short training run or an experiment, that trade is usually correct. See RunPod vs. Vast.ai for how those two specifically differ on reliability and interruption risk.
Small team, production inference or fine-tuning, needs predictable uptime but no regulatory mandate. Mid-tier specialist clouds (Lambda, Vultr’s GPU cloud) sit between marketplace pricing and hyperscaler pricing, with more standardized infrastructure and support than a marketplace host. Lambda’s ~$2.99/hr on-demand rate is still a fraction of AWS. LocalRig’s Lambda Cloud review and Vultr GPU cloud review cover the operational differences that matter at this tier.
Regulated industry, existing cloud footprint, or contractual compliance requirement. This is the actual hyperscaler buyer. If your data must stay inside a HIPAA-eligible environment, or your organization already has an AWS/Azure enterprise agreement with negotiated rates, the sticker price above is not the real price you would pay anyway — enterprise discounting changes the math substantially. Get a quote; do not budget off the public on-demand rate.
Uncertain how much compute you actually need. Before renting an H100 at any price, confirm the workload requires H100-class VRAM and bandwidth in the first place. See Cloud GPU: Fine-Tuning vs. Inference — a lot of jobs that get quoted against an H100 would run fine, and far cheaper, on a smaller card.

What do aggregator sites get wrong about H100 pricing?

Sites like gpuperhour and computeprices dominate this search term with tables that are often accurate on the numbers but silent on the “why.” They rank providers by price alone, which implicitly tells a hobbyist and an enterprise compliance team to make the same decision. That is the wrong frame — the cheapest row in the table is not the best answer for every buyer, and the most expensive row is not a rip-off.

LocalRig’s position here is narrower and, we think, more useful: publish the same snapshot pricing, but attach the buyer-constraint logic so the table tells you which row is for you, not just which row is smallest. This page is also the seed for a live, dated H100 price table — check the local-vs-cloud tools hub for the running comparison as providers update rates.

How should I compare H100 rentals beyond the sticker price?

Price per hour is the headline, but it is not the only number that determines what a training run actually costs you. Before booking, check:

Interruption risk. Marketplace and community-tier instances can be preemptible or hosted on less standardized hardware. A run that gets killed mid-training and has to restart from checkpoint can erase the hourly savings fast. RunPod vs. Vast.ai covers this trade-off directly, and how to avoid RunPod data loss is worth reading before any long unattended job.
Networking. Multi-GPU training over NVLink or InfiniBand behaves very differently than GPUs stitched together over commodity networking. If you are training across multiple H100s, the interconnect matters as much as the per-GPU rate.
Storage and egress. Some providers charge separately for persistent storage or data egress; a low compute rate can be offset by storage fees if you are not tracking total cost.
Setup and idle time. Hourly billing that starts at instance boot (not at job start) penalizes slow environment setup. Factor in your own setup time, not just the advertised rate.

Bottom line

The H100 price spread is real and it is not a pricing scam on either end of the range — it reflects genuinely different products bolted onto the same silicon. If you have no compliance mandate and are cost-sensitive, the marketplace and specialist tiers (Vast.ai, RunPod, Lambda) will save you real money and are where most of LocalRig’s audience should be looking first. If you are regulated, embedded in a hyperscaler ecosystem, or need contractual SLAs, the premium buys something specific and is not automatically overpriced — but get an enterprise quote rather than budgeting off the public on-demand rate. Either way, verify current pricing before you commit a run: every number on this page is a snapshot from 2026-06-29, and this market moves.