H100 cloud pricing per hour and rental comparison
Cheapest on-demand H100 confirmed in stock right now
LimitedThe cheapest on-demand single H100 confirmed in stock is $1.90 per GPU-hour on Lium. Stock checked 19 November 2026 at 22:20 UTC. Limited stock was reported. 8 providers have fresh self-service on-demand stock for this size; the on-demand catalog contains 107 listings across 21 providers.
How this stock answer is checked
Evaluated . Stock checks must be no more than 6 hours old; limited stock counts as in stock. Unknown, stale, out-of-stock and sales-only offers do not qualify. Confirm capacity with the provider before renting.
Selected listing history: fully available for 0% of tracked time in the last 30 days. Limited stock is rentable now but does not count as fully available. A 0% history does not mean there is no stock now.
Coverage: 720 tracked listing-hours. Coverage can be shorter than the 30-day window; hours add up across combined regional listings. This measures reported stock, not machine uptime.
Stock and history methodology. Inspect listings and stock filters.
Compare H100 on-demand and spot rental rates across 26 catalog providers, with 8x H100 node pricing and live stock.
Stock checks run regularly where supported; collection delays can occur and the cheapest in-stock offer can change between checks. Latest stock observation on this page: at page generation; individual listings may be older. Set a price alert to get an email when the H100 price drops.
Single H100 rental summary
stock checked- How have single-GPU H100 catalog prices changed?
- ▲ near the highest monthly listed rate — The November 2026 catalog snapshot lists 1× H100 from $1.29/hr ($1.29/GPU-hr), up 29% versus the previous recorded month. Lowest listed rates ranged from $0.67 to $1.29/hr across 6 recorded monthly snapshots (June 2026–November 2026). These are historical on-demand list prices; they do not establish stock availability or predict current rental prices.
- What is the cheapest single H100 on-demand cloud GPU available right now?
- The cheapest on-demand single H100 confirmed in stock is $1.90 per GPU-hour on Lium. Stock checked 19 November 2026 at 22:20 UTC. Limited stock was reported.
- Which providers have the strongest single H100 stock history?
- For 1× H100 listings across tracked regions, the highest mean fully-available history scores within the trailing 30-day window are Nebius, Akash. Coverage varies by listing. Lower fully-available history scores: Vast, Koyeb.
Current on-demand answers use fresh self-service stock, including limited stock; historical price trends include catalog listings and do not establish current capacity. How we compute this →
H100 rental prices by GPU count
Lowest on-demand H100 price reported in stock at each node size across 11 providers, normalised to per-GPU-hour so 1× and 8× listings compare like for like. Refreshed hourly.
| Node | Per GPU-hour | Node $/hr | Provider | Region | Stock | Spot / GPU-hr |
|---|---|---|---|---|---|---|
| 1× H100 | $1.90 | $1.90 | Lium8 providers at 1× | Kazakhstan | Limited | $0.67 |
| 2× H100 | $2.00 | $3.99 | Lium8 providers at 2× | Russia | Limited | $0.67 |
| 4× H100 | $2.00 | $7.98 | Lium6 providers at 4× | Russia | Limited | $1.33 |
| 8× H100 | $3.00 | $24.00 | PrimeIntellect3 providers at 8× | US | In stock | $1.69 |
H100 in stock now
- Lium $1.90/GPU-hr · 1× · 0% fully available · 30d history
- Shadeform $1.99/GPU-hr · 1× · 10% fully available · 30d history
- Vast $2.44/GPU-hr · 1× · 0% fully available · 30d history
- Koyeb $2.50/GPU-hr · 1× · 0% fully available · 30d history
- Runpod $2.59/GPU-hr · 1× · 0% fully available · 30d history
- Akash $2.73/GPU-hr · 1× · 97% fully available · 30d history
Stock now uses the latest provider report. History measures time reported fully available. Limited stock is rentable now but does not count as fully available. A 0% history can still have stock now.
Cheapest in-stock configuration per provider; “in stock” means the provider reported capacity on our last hourly check. Full list in the pricing table and the availability heatmap.
When to upgrade from the H100 to the H200
The H200 is the same Hopper GH100 die as the H100 — identical 1,979 TFLOPS FP16, NVLink 4.0 and 700 W — with the memory swapped from 80 GB HBM3 to 141 GB HBM3e (76% more capacity, 43% more bandwidth). Upgrade when memory is the constraint: serving a 70B-class model in FP16 that spills to two H100s but fits one H200, long-context inference where the KV cache dominates, or batch sizes capped by VRAM. In those cases one H200 often replaces two H100s and the per-token cost drops even though the hourly rate is 60–70% higher.
Stay on the H100 when the job is compute-bound — pre-training, fine-tuning of models that already fit, and anything that saturates the Tensor Cores at 80 GB — because the H200 runs the same math at the same speed for more money. H100 spot is also far deeper than H200 spot, so price-sensitive, interruptible training is still an H100 workload. The live H100 vs H200 comparison page has today's prices and stock side by side; the blog guide walks through the scenario maths.
Page headline prices reflect the default offers: fresh reported stock, including limited stock, with self-service access.
In-stock price range: 27 1× H100 listings from $1.90 to $7.91 USD per physical GPU-hour. Whole-node prices for other GPU counts are not mixed into this range.
H100 rental prices by provider
Stock now uses the latest provider report. History measures time reported fully available. Limited stock is rentable now but does not count as fully available. A 0% history can still have stock now.
Vast.ai marketplace: 11 more H100 listings were unavailable under our collection filters at the last check, last listed from $2.16/GPU-hr. They are hidden above; browse the provider for current offers and rental conditions.
Browse live Vast.ai offersGet notified when H100 drops in price or reliability changes
One email per change, max once a day. We send a confirmation link first; one-click unsubscribe in every email.
Which H100 rental configuration should you compare?
These listings price cloud rental by the hour. Hardware purchase prices answer a different question. Before choosing a provider, match the GPU allocation and host configuration to your workload.
- Match the H100 variant and memory
- Check the exact instance specification: NVIDIA lists 80 GB per GPU for H100 SXM and 94 GB for H100 NVL. Do not assume every H100 listing has the same memory or bandwidth. Confirm a full GPU allocation rather than a smaller MIG partition.
- Match the billable node size
- A per-GPU rate on an eight-GPU node is a comparison unit. If the provider bills the whole node, all eight GPUs contribute to your bill even when your job uses fewer. Compare the same GPU count, region and on-demand or interruptible terms.
- Verify communication for distributed work
- For a multi-GPU job, ask for the exact GPU-to-GPU topology of the selected instance. For multiple nodes, verify the network between nodes separately. The H100 name or eight-GPU count alone does not establish NVLink, NVSwitch or InfiniBand connectivity.
- Confirm capacity and test the workload
- Use fresh stock observations to shortlist, then confirm capacity and account quota in the provider console. Check CPU, RAM, storage and software requirements. Benchmark a representative run before estimating a larger job; historical stock is not an uptime SLA or a reservation.
Hardware reference: NVIDIA H100 variant and memory specifications. This does not verify any rental host's configuration.
Compare listings and visit a provider →Estimate compute and egress →
H100 pricing by provider
H100 (1× GPU) availability history — last 7 days
2 of 16 providers kept 1× H100 ≥80% fully available all week. Polled hourly · 16 providers tracked · 7-day window.
Percentages show tracked time reported fully available. Limited stock can still be rented, but does not count as fully available here. Missing data is not 0%; this history is not machine uptime.
| Provider | Fri13 | Sat14 | Sun15 | Mon16 | Tue17 | Wed18 | Thu19 | Week |
|---|---|---|---|---|---|---|---|---|
| Akash | 100% | |||||||
| Nebius | 100% | |||||||
| Fluence | 74% | |||||||
| Shadeform | 31% | |||||||
| PrimeIntellect | 28% | |||||||
| OVHcloud | 19% | |||||||
| Digital Ocean | 16% | |||||||
| Lambda | 12% | |||||||
| Lium | 10% | |||||||
| Verda | 4% | |||||||
| EmpirioLabs | 3% | |||||||
| Runpod | 3% | |||||||
| AceCloud | 0% | |||||||
| Hyperstack | 0% | |||||||
| Koyeb | 0% | |||||||
| Vast | 0% |
30-day stock history
For H100, history is the share of tracked time each listing was reported fully available within the last 30 days. Limited stock is rentable now but does not count as fully available. This measures reported stock, not machine uptime. How we compute this →
- 41
- listings with a history percentage
- 77%
- of listings on this page have enough history
- 2,880h
- most tracked listing-hours for one row
Highest fully-available shares: Nebius (100% fully available; 720h tracked), Shadeform (100% fully available; 720h tracked), Akash (97% fully available; 720h tracked).
Coverage can be shorter than the 30-day window. Tracked listing-hours add time across combined regional listings, so they can exceed elapsed calendar time. Percentages stay hidden below 48 tracked listing-hours. Unknown periods are missing data, not time with no stock. Confirm current capacity with the provider.
Price History & Comparison: 8× H100
Full H100 price history (22 months) →Spot pricing decision guide
Treat spot as a risk-adjusted capacity decision, not just a cheaper number.
per GPU-hour
65% below on-demand floor
daily median for the same offers with complete history
median $1.50/GPU-hr
Cheapest trusted spot / interruptible capacity is on Vast; 2 providers currently report spot or interruptible stock (44 units/signals visible).
Nebius has kept spot or interruptible capacity in stock most reliably over the last 30 days (100%). See the availability section for the broader stock history.
The discount is visible, but availability, reliability, or volatility argues for more caution.
Good fit: checkpointed batch jobs, flexible training, and experiments that can restart. Avoid for production serving, deadline-bound runs, or jobs that cannot tolerate eviction.
Source caveat: Runpod exposes explicit spot fields; Vast is labeled interruptible/bid-floor. Lambda is treated as on-demand-only unless a verified spot field is added.
H100 spot, preemptible and interruptible pricing
The cheapest H100 spot price per GPU from each provider that publishes one, ordered by price. Each card uses the provider’s own term and shows how the price is set; the provider can reclaim this capacity. Spot prices for every GPU
VastInterruptible
$0.67/GPU-hr
vs on-demand $3.23/GPU-hr · 79% below on-demand · 2× in Czechia, CZ
If interrupted: Paused; resumes when it regains priority
NebiusPreemptible
$2.15/GPU-hr
vs on-demand $4.50/GPU-hr · 52% below on-demand · 1× in eu-north1
If interrupted: Stopped; restart it yourself (Managed Kubernetes restarts nodes)
H100 spot pricing FAQ
Who has H100 spot GPU pricing?
2 providers list H100 spot, preemptible or interruptible prices on GPU Finder right now, cheapest per GPU first: Vast $0.67/GPU-hr (interruptible), Nebius $2.15/GPU-hr (preemptible). Each provider's own term is shown.
What happens when the provider reclaims H100 spot capacity?
It depends on the provider: Vast: paused; resumes when it regains priority; Nebius: stopped; restart it yourself (Managed Kubernetes restarts nodes). Save checkpoints to storage that survives the instance.
Should I choose the cheapest H100 spot offer?
Only if the workload can tolerate interruption. H100 spot, preemptible and interruptible prices are best for checkpointed batch jobs, flexible training and experiments. Avoid them for production serving, deadline-bound runs or jobs that cannot restart.
About the H100
The NVIDIA H100 is the GPU that defined the generative AI era. Built on the Hopper architecture with a dedicated Transformer Engine, it accelerates large language model training and inference at a scale no previous GPU could match. It remains the default choice for teams training frontier models and any workload where FP8 throughput and NVLink bandwidth are bottlenecks.
Key Specifications
| Architecture | Hopper (GH100) |
| GPU Memory | 80 GB HBM3 (SXM5) |
| Memory Bandwidth | 3.35 TB/s |
| FP16 Tensor Core | 1,979 TFLOPS |
| TDP | 700W (SXM5) / 350W (PCIe) |
| Interconnect | NVLink 4.0 (900 GB/s) |
| Release Year | 2023 |
Cloud Pricing Context
Use the live listings to compare H100 cloud rental rates at the same GPU count and purchase model. Confirm the exact H100 variant, memory allocation and region before comparing providers. A normalized per-GPU rate on a multi-GPU instance does not mean that one GPU can be rented separately. Storage, data transfer and billed setup time can change the final job cost.
Best For
- Pre-training and fine-tuning LLMs at the 7B-70B parameter scale
- High-throughput inference serving with FP8 quantization via the Transformer Engine
- Distributed training after verifying the selected node topology and inter-node network
- Production inference pipelines requiring predictable latency at scale
H100 by node size
Compare across the market
Weighing the H100 against other GPUs? The cheapest cloud GPU prices per hour rank every model we track by its lowest per-GPU-hour rate and the cheapest listing in stock right now.
Dated availability research
Cite the GPU Cloud Availability Index for frozen H100 provider reliability rankings, coverage hours and downloadable data.